publish v1
Browse files- .gitignore +3 -0
- app.py +107 -0
- core/chessboard_detector.py +137 -0
- core/helper_34.py +316 -0
- core/helper_cls.py +23 -0
- core/kpt_34_with_xanything.py +197 -0
- core/runonnx/base_onnx.py +88 -0
- core/runonnx/full_classifier.py +203 -0
- core/runonnx/rtmdet.py +117 -0
- core/runonnx/rtmpose.py +356 -0
- requirements.txt +1 -0
.gitignore
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
__pycache__/
|
| 2 |
+
*.pyc
|
| 3 |
+
*.DS_Store
|
app.py
ADDED
|
@@ -0,0 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
from core.chessboard_detector import ChessboardDetector
|
| 3 |
+
|
| 4 |
+
detector = ChessboardDetector(
|
| 5 |
+
det_model_path="onnx/det/v1.onnx",
|
| 6 |
+
pose_model_path="onnx/pose/v1.onnx",
|
| 7 |
+
full_classifier_model_path="onnx/layout_recognition/v1.onnx"
|
| 8 |
+
)
|
| 9 |
+
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
# 数据集路径
|
| 13 |
+
dict_cate_names = {
|
| 14 |
+
'.': '.',
|
| 15 |
+
'x': 'x',
|
| 16 |
+
'红帅': 'K',
|
| 17 |
+
'红士': 'A',
|
| 18 |
+
'红相': 'B',
|
| 19 |
+
'红马': 'N',
|
| 20 |
+
'红车': 'R',
|
| 21 |
+
'红炮': 'C',
|
| 22 |
+
'红兵': 'P',
|
| 23 |
+
|
| 24 |
+
'黑将': 'k',
|
| 25 |
+
'黑仕': 'a',
|
| 26 |
+
'黑象': 'b',
|
| 27 |
+
'黑傌': 'n',
|
| 28 |
+
'黑車': 'r',
|
| 29 |
+
'黑砲': 'c',
|
| 30 |
+
'黑卒': 'p',
|
| 31 |
+
}
|
| 32 |
+
|
| 33 |
+
dict_cate_names_reverse = {v: k for k, v in dict_cate_names.items()}
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
with gr.Blocks(
|
| 37 |
+
css="""
|
| 38 |
+
.image {
|
| 39 |
+
max-height: 512px;
|
| 40 |
+
}
|
| 41 |
+
"""
|
| 42 |
+
) as demo:
|
| 43 |
+
gr.Markdown("""
|
| 44 |
+
## 棋盘检测, 棋子识别
|
| 45 |
+
|
| 46 |
+
步骤:
|
| 47 |
+
1. 流程分成两步,第一步检测边缘
|
| 48 |
+
2. 对整个棋盘画面进行棋子分类预测
|
| 49 |
+
"""
|
| 50 |
+
)
|
| 51 |
+
|
| 52 |
+
with gr.Row():
|
| 53 |
+
with gr.Column():
|
| 54 |
+
image_input = gr.Image(label="上传棋盘图片", type="numpy", elem_classes="image")
|
| 55 |
+
|
| 56 |
+
with gr.Column():
|
| 57 |
+
original_image_with_keypoints = gr.Image(
|
| 58 |
+
label="step1: 原图带关键点",
|
| 59 |
+
interactive=False,
|
| 60 |
+
visible=True,
|
| 61 |
+
elem_classes="image"
|
| 62 |
+
)
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
with gr.Row():
|
| 66 |
+
with gr.Column():
|
| 67 |
+
transformed_image = gr.Image(
|
| 68 |
+
label="step2: 拉伸棋盘",
|
| 69 |
+
interactive=False,
|
| 70 |
+
visible=True,
|
| 71 |
+
elem_classes="image"
|
| 72 |
+
)
|
| 73 |
+
|
| 74 |
+
with gr.Column():
|
| 75 |
+
use_time = gr.Textbox(
|
| 76 |
+
label="用时",
|
| 77 |
+
interactive=False,
|
| 78 |
+
visible=True,
|
| 79 |
+
)
|
| 80 |
+
layout_pred_info = gr.Dataframe(
|
| 81 |
+
label="棋子识别",
|
| 82 |
+
interactive=False,
|
| 83 |
+
visible=True,
|
| 84 |
+
)
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
def detect_chessboard(image):
|
| 88 |
+
original_image_with_keypoints, transformed_image, cells_labels_str, scores, time_info = detector.pred_detect_board_and_classifier(image)
|
| 89 |
+
|
| 90 |
+
# 将 cells_labels 转换为 DataFrame
|
| 91 |
+
# cells_labels 通过 \n 分割
|
| 92 |
+
annotation_10_rows = [item for item in cells_labels_str.split("\n")]
|
| 93 |
+
# 将 annotation_10_rows 转换成为 10 行 9 列的二维数组
|
| 94 |
+
annotation_arr_10_9 = [list(item) for item in annotation_10_rows]
|
| 95 |
+
|
| 96 |
+
# 将 棋子类别 转换为 中文
|
| 97 |
+
annotation_arr_10_9 = [[dict_cate_names_reverse[item] for item in row] for row in annotation_arr_10_9]
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
return original_image_with_keypoints, transformed_image, annotation_arr_10_9, time_info
|
| 101 |
+
|
| 102 |
+
image_input.change(fn=detect_chessboard,
|
| 103 |
+
inputs=[image_input],
|
| 104 |
+
outputs=[original_image_with_keypoints, transformed_image, layout_pred_info, use_time])
|
| 105 |
+
|
| 106 |
+
if __name__ == "__main__":
|
| 107 |
+
demo.launch()
|
core/chessboard_detector.py
ADDED
|
@@ -0,0 +1,137 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
import time
|
| 3 |
+
import numpy as np
|
| 4 |
+
import cv2
|
| 5 |
+
from typing import List, Tuple, Union
|
| 6 |
+
from pandas import DataFrame
|
| 7 |
+
from .runonnx.rtmdet import RTMDET_ONNX
|
| 8 |
+
from .runonnx.rtmpose import RTMPOSE_ONNX
|
| 9 |
+
from .runonnx.full_classifier import FULL_CLASSIFIER_ONNX
|
| 10 |
+
|
| 11 |
+
from core.helper_34 import extract_chessboard
|
| 12 |
+
|
| 13 |
+
class ChessboardDetector:
|
| 14 |
+
def __init__(self,
|
| 15 |
+
det_model_path: str,
|
| 16 |
+
pose_model_path: str,
|
| 17 |
+
full_classifier_model_path: str = None
|
| 18 |
+
):
|
| 19 |
+
|
| 20 |
+
self.det = RTMDET_ONNX(
|
| 21 |
+
model_path=det_model_path,
|
| 22 |
+
)
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
self.pose = RTMPOSE_ONNX(
|
| 26 |
+
model_path=pose_model_path,
|
| 27 |
+
)
|
| 28 |
+
|
| 29 |
+
if full_classifier_model_path is not None:
|
| 30 |
+
self.full_classifier = FULL_CLASSIFIER_ONNX(
|
| 31 |
+
model_path=full_classifier_model_path,
|
| 32 |
+
)
|
| 33 |
+
|
| 34 |
+
self.board_positions = [] # 存储棋盘位置坐标
|
| 35 |
+
self.current_image = None
|
| 36 |
+
self.current_filename = None
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
# 检测中国象棋棋盘
|
| 40 |
+
def pred_detect_and_keypoints(self, image_bgr: Union[np.ndarray, None] = None) -> Tuple[List[int], float, List[List[int]], List[float]]:
|
| 41 |
+
|
| 42 |
+
xyxy, conf = self.det.pred(image_bgr)
|
| 43 |
+
|
| 44 |
+
# 预测关键点, 绘制关键点
|
| 45 |
+
keypoints, scores = self.pose.pred(image=image_bgr, bbox=xyxy)
|
| 46 |
+
|
| 47 |
+
return xyxy, conf, keypoints, scores
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
def draw_pred_with_keypoints(self, image_rgb: Union[np.ndarray, None] = None) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
|
| 51 |
+
if image_rgb is None:
|
| 52 |
+
return None, None, None, None
|
| 53 |
+
|
| 54 |
+
image_rgb = image_rgb.copy()
|
| 55 |
+
|
| 56 |
+
original_image = image_rgb.copy()
|
| 57 |
+
|
| 58 |
+
image_bgr = cv2.cvtColor(image_rgb, cv2.COLOR_RGB2BGR)
|
| 59 |
+
|
| 60 |
+
xyxy, conf, keypoints, scores = self.pred_detect_and_keypoints(image_bgr)
|
| 61 |
+
|
| 62 |
+
# 绘制棋盘框架
|
| 63 |
+
draw_image = self.det.draw_pred(image_rgb, xyxy, conf)
|
| 64 |
+
|
| 65 |
+
# 绘制关键点
|
| 66 |
+
draw_image = self.pose.draw_pred(img=draw_image, keypoints=keypoints, scores=scores)
|
| 67 |
+
|
| 68 |
+
# 融合 self.pose.bone_names 与 keypoints, 再转换成 DataFrame
|
| 69 |
+
keypoint_list = []
|
| 70 |
+
for bone_name, keypoint in zip(self.pose.bone_names, keypoints):
|
| 71 |
+
keypoint_list.append({"name": bone_name, "x": keypoint[0], "y": keypoint[1]})
|
| 72 |
+
|
| 73 |
+
keypoint_df = DataFrame(keypoint_list)
|
| 74 |
+
|
| 75 |
+
return draw_image, original_image, [xyxy], keypoint_df
|
| 76 |
+
|
| 77 |
+
# 拉伸棋盘 detect board, 然后预测
|
| 78 |
+
def extract_chessboard_and_classifier_layout(self,
|
| 79 |
+
image_rgb: Union[np.ndarray, None] = None,
|
| 80 |
+
keypoints: Union[np.ndarray, None] = None
|
| 81 |
+
) -> Tuple[np.ndarray, List[List[str]], List[List[float]]]:
|
| 82 |
+
|
| 83 |
+
# 提取棋盘, 绘制 每个位置的 范围信息
|
| 84 |
+
transformed_image, _transformed_keypoints, _corner_points = extract_chessboard(img=image_rgb, keypoints=keypoints)
|
| 85 |
+
|
| 86 |
+
transformed_image_copy = transformed_image.copy()
|
| 87 |
+
|
| 88 |
+
# 预测每个位置的 棋子类别
|
| 89 |
+
_, _, scores, pred_result = self.full_classifier.pred(transformed_image_copy, is_rgb=True)
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
return transformed_image, pred_result, scores
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
# 检测棋盘 detect board
|
| 96 |
+
def pred_detect_board_and_classifier(self,
|
| 97 |
+
image_rgb: Union[np.ndarray, None] = None,
|
| 98 |
+
) -> Tuple[np.ndarray, np.ndarray, str, List[List[float]], str]:
|
| 99 |
+
|
| 100 |
+
"""
|
| 101 |
+
@param image_rgb: 输入的 RGB 图像
|
| 102 |
+
@return:
|
| 103 |
+
- transformed_image_layout # 拉伸棋盘
|
| 104 |
+
- original_image_with_keypoints # 原图关键点
|
| 105 |
+
- layout_pred_info # 每个位置的 棋子类别
|
| 106 |
+
- scores # 每个位置的 置信度
|
| 107 |
+
- time_info # 推理用时
|
| 108 |
+
"""
|
| 109 |
+
|
| 110 |
+
if image_rgb is None:
|
| 111 |
+
return None, None, [], [], ""
|
| 112 |
+
|
| 113 |
+
image_rgb_for_extract = image_rgb.copy()
|
| 114 |
+
|
| 115 |
+
start_time = time.time()
|
| 116 |
+
|
| 117 |
+
image_bgr = cv2.cvtColor(image_rgb, cv2.COLOR_RGB2BGR)
|
| 118 |
+
|
| 119 |
+
xyxy, conf, keypoints, scores = self.pred_detect_and_keypoints(image_bgr)
|
| 120 |
+
|
| 121 |
+
# 绘制棋盘框架
|
| 122 |
+
draw_image = self.det.draw_pred(image_rgb, xyxy, conf)
|
| 123 |
+
|
| 124 |
+
"""
|
| 125 |
+
绘制 原图关键点
|
| 126 |
+
"""
|
| 127 |
+
original_image_with_keypoints = self.pose.draw_pred(img=draw_image, keypoints=keypoints, scores=scores)
|
| 128 |
+
|
| 129 |
+
transformed_image, cells_labels, scores = self.extract_chessboard_and_classifier_layout(image_rgb=image_rgb_for_extract, keypoints=keypoints)
|
| 130 |
+
|
| 131 |
+
|
| 132 |
+
use_time = time.time() - start_time
|
| 133 |
+
|
| 134 |
+
time_info = f"推理用时: {use_time:.2f}s"
|
| 135 |
+
|
| 136 |
+
return original_image_with_keypoints, transformed_image, cells_labels, scores, time_info
|
| 137 |
+
|
core/helper_34.py
ADDED
|
@@ -0,0 +1,316 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import numpy as np
|
| 2 |
+
import cv2
|
| 3 |
+
from typing import Tuple, List
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
BONE_NAMES = [
|
| 7 |
+
"A0", "A1", "A2", "A3", "A4", "A5", "A6", "A7", "A8",
|
| 8 |
+
"J0", "J1", "J2", "J3", "J4", "J5", "J6", "J7", "J8",
|
| 9 |
+
"B0", "C0", "D0", "E0", "F0", "G0", "H0", "I0",
|
| 10 |
+
"B8", "C8", "D8", "E8", "F8", "G8", "H8", "I8",
|
| 11 |
+
]
|
| 12 |
+
|
| 13 |
+
def check_keypoints(keypoints: np.ndarray):
|
| 14 |
+
"""
|
| 15 |
+
检查关键点坐标是否正确
|
| 16 |
+
@param keypoints: 关键点坐标, shape 为 (34, 2)
|
| 17 |
+
"""
|
| 18 |
+
if keypoints.shape != (34, 2):
|
| 19 |
+
raise Exception(f"keypoints shape error: {keypoints.shape}")
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
def build_cells_xywh_by_cronners(corner_points: np.ndarray, padding: int = 3) -> np.ndarray:
|
| 23 |
+
"""
|
| 24 |
+
根据 棋盘的 corner 点坐标 计算 每个位置的 xywh
|
| 25 |
+
@param corner_points: 棋盘的 corner 点坐标, shape 为 (4, 2)
|
| 26 |
+
@param padding: 棋盘边框 padding
|
| 27 |
+
|
| 28 |
+
@return: 棋盘的 xywh, shape 为 (10, 9, 4), 4 为 center_x, center_y, w, h
|
| 29 |
+
"""
|
| 30 |
+
|
| 31 |
+
if corner_points.shape != (4, 2):
|
| 32 |
+
raise Exception(f"corner_points shape error: {corner_points.shape}")
|
| 33 |
+
|
| 34 |
+
top_left_xy = corner_points[0]
|
| 35 |
+
top_right_xy = corner_points[1]
|
| 36 |
+
bottom_left_xy = corner_points[2]
|
| 37 |
+
bottom_right_xy = corner_points[3]
|
| 38 |
+
|
| 39 |
+
# 计算 每个框的 w 和 h
|
| 40 |
+
item_w = (top_right_xy[0] - top_left_xy[0]) / (9 - 1)
|
| 41 |
+
item_h = (bottom_left_xy[1] - top_left_xy[1]) / (10 - 1)
|
| 42 |
+
|
| 43 |
+
item_w = item_w
|
| 44 |
+
item_h = item_h
|
| 45 |
+
|
| 46 |
+
item_w_with_padding = item_w - padding * 2
|
| 47 |
+
item_h_with_padding = item_h - padding * 2
|
| 48 |
+
|
| 49 |
+
# 计算 每个框的 center 坐标
|
| 50 |
+
cells_xywh = np.zeros((10, 9, 4))
|
| 51 |
+
|
| 52 |
+
for i in range(10):
|
| 53 |
+
for j in range(9):
|
| 54 |
+
center_x = top_left_xy[0] + item_w * j
|
| 55 |
+
center_y = top_left_xy[1] + item_h * i
|
| 56 |
+
|
| 57 |
+
cells_xywh[i, j] = [center_x, center_y, item_w_with_padding, item_h_with_padding]
|
| 58 |
+
|
| 59 |
+
return cells_xywh
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
# todo: 需要优化
|
| 64 |
+
def build_cells_xywh(keypoints: np.ndarray, width: int = 450, height: int = 500, padding: int = 3) -> np.ndarray:
|
| 65 |
+
"""
|
| 66 |
+
@param keypoints: 关键点坐标, shape 为 (34, 2)
|
| 67 |
+
@param width: 棋盘宽度
|
| 68 |
+
@param height: 棋盘高度
|
| 69 |
+
@param padding: 棋盘边框 padding
|
| 70 |
+
@return: 棋盘的 xywh, shape 为 (10, 9, 4), 4 为 center_x, center_y, w, h
|
| 71 |
+
"""
|
| 72 |
+
check_keypoints(keypoints)
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
# 生成 A0 到 J8 的坐标, 如 B1 坐标 为 A1-J1 与 B0-B8 的交集点
|
| 76 |
+
cells_xywh = np.zeros((10, 9, 4), dtype=np.int16)
|
| 77 |
+
|
| 78 |
+
# 遍历 full_points 的每个点,计算其坐标
|
| 79 |
+
for i in range(10):
|
| 80 |
+
for j in range(9):
|
| 81 |
+
# 计算 第 i 行 第 j 列 的坐标
|
| 82 |
+
row_name = chr(ord('A') + i)
|
| 83 |
+
col_name = str(j)
|
| 84 |
+
flag_name = f"{row_name}{col_name}"
|
| 85 |
+
if flag_name in BONE_NAMES:
|
| 86 |
+
# 计算 第 i 行 第 j 列 的坐标
|
| 87 |
+
cur_xy = keypoints[BONE_NAMES.index(flag_name)]
|
| 88 |
+
cells_xywh[i, j] = [cur_xy[0], cur_xy[1], 0, 0]
|
| 89 |
+
else:
|
| 90 |
+
# 计算 第 i 行 第 j 列 的坐标
|
| 91 |
+
row_start_name = f"{row_name}0"
|
| 92 |
+
row_end_name = f"{row_name}8"
|
| 93 |
+
|
| 94 |
+
col_start_name = f"A{col_name}"
|
| 95 |
+
col_end_name = f"J{col_name}"
|
| 96 |
+
|
| 97 |
+
row_start_xy = keypoints[BONE_NAMES.index(row_start_name)]
|
| 98 |
+
row_end_xy = keypoints[BONE_NAMES.index(row_end_name)]
|
| 99 |
+
|
| 100 |
+
col_start_xy = keypoints[BONE_NAMES.index(col_start_name)]
|
| 101 |
+
col_end_xy = keypoints[BONE_NAMES.index(col_end_name)]
|
| 102 |
+
|
| 103 |
+
# 计算 row_start_xy 到 row_end_xy 的直线 与 col_start_xy 到 col_end_xy 的直线 的交点
|
| 104 |
+
# 使用参数方程法计算交点
|
| 105 |
+
x1, y1 = row_start_xy # 横向直线起点
|
| 106 |
+
x2, y2 = row_end_xy # 横向直线终点
|
| 107 |
+
x3, y3 = col_start_xy # 纵向直线起点
|
| 108 |
+
x4, y4 = col_end_xy # 纵向直线终点
|
| 109 |
+
|
| 110 |
+
# 计算交点坐标
|
| 111 |
+
# 使用克莱姆法则求解
|
| 112 |
+
denominator = (x1 - x2) * (y3 - y4) - (y1 - y2) * (x3 - x4)
|
| 113 |
+
|
| 114 |
+
# 计算交点的 x 坐标
|
| 115 |
+
x = ((x1 * y2 - y1 * x2) * (x3 - x4) - (x1 - x2) * (x3 * y4 - y3 * x4)) / denominator
|
| 116 |
+
# 计算交点的 y 坐标
|
| 117 |
+
y = ((x1 * y2 - y1 * x2) * (y3 - y4) - (y1 - y2) * (x3 * y4 - y3 * x4)) / denominator
|
| 118 |
+
|
| 119 |
+
cells_xywh[i, j] = [int(x), int(y), 0, 0]
|
| 120 |
+
|
| 121 |
+
# 计算每个点位的 wh
|
| 122 |
+
for i in range(10):
|
| 123 |
+
for j in range(9):
|
| 124 |
+
cur_xy = cells_xywh[i, j]
|
| 125 |
+
# 获取上下左右 4 个点, 根据 4 个点计算 wh, 宽高为 4 个点 计算出来的 x1y1x2y2 的距离 的 1/2
|
| 126 |
+
if i == 0:
|
| 127 |
+
# [i+1, j] 的 反向点
|
| 128 |
+
up_xy = 2 * cur_xy - cells_xywh[i+1, j]
|
| 129 |
+
else:
|
| 130 |
+
up_xy = cells_xywh[i - 1, j]
|
| 131 |
+
|
| 132 |
+
if i == 9:
|
| 133 |
+
# [i-1, j] 的 反向点
|
| 134 |
+
down_xy = 2 * cur_xy - cells_xywh[i-1, j]
|
| 135 |
+
else:
|
| 136 |
+
down_xy = cells_xywh[i+1, j]
|
| 137 |
+
|
| 138 |
+
if j == 0:
|
| 139 |
+
left_xy = 2 * cur_xy - cells_xywh[i, j+1]
|
| 140 |
+
else:
|
| 141 |
+
left_xy = cells_xywh[i, j-1]
|
| 142 |
+
|
| 143 |
+
if j == 8:
|
| 144 |
+
right_xy = 2 * cur_xy - cells_xywh[i, j-1]
|
| 145 |
+
else:
|
| 146 |
+
right_xy = cells_xywh[i, j+1]
|
| 147 |
+
|
| 148 |
+
min_x = min(up_xy[0].tolist(), down_xy[0].tolist(), left_xy[0].tolist(), right_xy[0].tolist())
|
| 149 |
+
min_y = min(up_xy[1].tolist(), down_xy[1].tolist(), left_xy[1].tolist(), right_xy[1].tolist())
|
| 150 |
+
|
| 151 |
+
min_x += padding
|
| 152 |
+
min_y += padding
|
| 153 |
+
|
| 154 |
+
# 防止 min_x 和 min_y 为 0
|
| 155 |
+
min_x = max(min_x, 1)
|
| 156 |
+
min_y = max(min_y, 1)
|
| 157 |
+
|
| 158 |
+
max_x = max(up_xy[0].tolist(), down_xy[0].tolist(), left_xy[0].tolist(), right_xy[0].tolist())
|
| 159 |
+
max_y = max(up_xy[1].tolist(), down_xy[1].tolist(), left_xy[1].tolist(), right_xy[1].tolist())
|
| 160 |
+
|
| 161 |
+
max_x -= padding
|
| 162 |
+
max_y -= padding
|
| 163 |
+
|
| 164 |
+
# 防止 max_x 和 max_y 超出边界
|
| 165 |
+
max_x = min(max_x, width - 1)
|
| 166 |
+
max_y = min(max_y, height - 1)
|
| 167 |
+
|
| 168 |
+
w = (max_x - min_x) / 2
|
| 169 |
+
h = (max_y - min_y) / 2
|
| 170 |
+
|
| 171 |
+
cells_xywh[i, j] = [int(cur_xy[0]), int(cur_xy[1]), int(w), int(h)]
|
| 172 |
+
|
| 173 |
+
return cells_xywh
|
| 174 |
+
|
| 175 |
+
|
| 176 |
+
def perspective_transform(
|
| 177 |
+
image: cv2.UMat,
|
| 178 |
+
src_points: np.ndarray,
|
| 179 |
+
keypoints: np.ndarray,
|
| 180 |
+
dst_size=(450, 500)) -> Tuple[cv2.UMat, np.ndarray, np.ndarray]:
|
| 181 |
+
"""
|
| 182 |
+
透视变换
|
| 183 |
+
@param image: 图片
|
| 184 |
+
@param src_points: 源点坐标
|
| 185 |
+
@param keypoints: 关键点坐标
|
| 186 |
+
@param dst_size: 目标尺寸 (width, height) 10 行 9 列
|
| 187 |
+
|
| 188 |
+
@return:
|
| 189 |
+
result: 透视变换后的图片
|
| 190 |
+
transformed_keypoints: 透视变换后的关键点坐标
|
| 191 |
+
corner_points: 棋盘的 corner 点坐标, shape 为 (4, 2) A0, A8, J0, J8
|
| 192 |
+
"""
|
| 193 |
+
|
| 194 |
+
check_keypoints(keypoints)
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
# 源点和目标点
|
| 198 |
+
src = np.float32(src_points)
|
| 199 |
+
padding = 50
|
| 200 |
+
corner_points = np.float32([
|
| 201 |
+
# 左上角
|
| 202 |
+
[padding, padding],
|
| 203 |
+
# 右上角
|
| 204 |
+
[dst_size[0]-padding, padding],
|
| 205 |
+
# 左下角
|
| 206 |
+
[padding, dst_size[1]-padding],
|
| 207 |
+
# 右下角
|
| 208 |
+
[dst_size[0]-padding, dst_size[1]-padding]])
|
| 209 |
+
|
| 210 |
+
# 计算透视变换矩阵
|
| 211 |
+
matrix = cv2.getPerspectiveTransform(src, corner_points)
|
| 212 |
+
|
| 213 |
+
# 执行透视变换
|
| 214 |
+
result = cv2.warpPerspective(image, matrix, dst_size)
|
| 215 |
+
|
| 216 |
+
# 重塑数组为要求的格式 (N,1,2)
|
| 217 |
+
keypoints_reshaped = keypoints.reshape(-1, 1, 2).astype(np.float32)
|
| 218 |
+
transformed_keypoints = cv2.perspectiveTransform(keypoints_reshaped, matrix)
|
| 219 |
+
# 转回原来的形状
|
| 220 |
+
transformed_keypoints = transformed_keypoints.reshape(-1, 2)
|
| 221 |
+
|
| 222 |
+
return result, transformed_keypoints, corner_points
|
| 223 |
+
|
| 224 |
+
|
| 225 |
+
|
| 226 |
+
def get_board_corner_points(keypoints: np.ndarray) -> np.ndarray:
|
| 227 |
+
"""
|
| 228 |
+
计算棋局四个边角的 points
|
| 229 |
+
@param keypoints: 关键点坐标, shape 为 (34, 2)
|
| 230 |
+
@return: 边角的坐标, shape 为 (4, 2)
|
| 231 |
+
"""
|
| 232 |
+
check_keypoints(keypoints)
|
| 233 |
+
|
| 234 |
+
# 找到 A0 A8 J0 J8 的坐标 以及 A4 和 J4 的坐标
|
| 235 |
+
a0_index = BONE_NAMES.index("A0")
|
| 236 |
+
a8_index = BONE_NAMES.index("A8")
|
| 237 |
+
j0_index = BONE_NAMES.index("J0")
|
| 238 |
+
j8_index = BONE_NAMES.index("J8")
|
| 239 |
+
|
| 240 |
+
a0_xy = keypoints[a0_index]
|
| 241 |
+
a8_xy = keypoints[a8_index]
|
| 242 |
+
j0_xy = keypoints[j0_index]
|
| 243 |
+
j8_xy = keypoints[j8_index]
|
| 244 |
+
|
| 245 |
+
# 计算新的四个角点坐标
|
| 246 |
+
dst_points = np.array([
|
| 247 |
+
a0_xy,
|
| 248 |
+
a8_xy,
|
| 249 |
+
j0_xy,
|
| 250 |
+
j8_xy
|
| 251 |
+
], dtype=np.float32)
|
| 252 |
+
|
| 253 |
+
return dst_points
|
| 254 |
+
|
| 255 |
+
def extract_chessboard(img: cv2.UMat, keypoints: np.ndarray) -> Tuple[cv2.UMat, np.ndarray, np.ndarray]:
|
| 256 |
+
"""
|
| 257 |
+
提取棋盘信息
|
| 258 |
+
@param img: 图片
|
| 259 |
+
@param keypoints: 关键点坐标, shape 为 (34, 2)
|
| 260 |
+
@return:
|
| 261 |
+
transformed_image: 透视变换后的图片
|
| 262 |
+
transformed_keypoints: 透视变换后的关键点坐标
|
| 263 |
+
transformed_corner_points: 棋盘的 corner 点坐标, shape 为 (4, 2) A0, A8, J0, J8
|
| 264 |
+
"""
|
| 265 |
+
|
| 266 |
+
check_keypoints(keypoints)
|
| 267 |
+
|
| 268 |
+
source_corner_points = get_board_corner_points(keypoints)
|
| 269 |
+
|
| 270 |
+
transformed_image, transformed_keypoints, transformed_corner_points = perspective_transform(img, source_corner_points, keypoints)
|
| 271 |
+
|
| 272 |
+
return transformed_image, transformed_keypoints, transformed_corner_points
|
| 273 |
+
|
| 274 |
+
|
| 275 |
+
def collect_cells_images(image: cv2.UMat, cells_xywh: np.ndarray) -> List[List[np.ndarray]]:
|
| 276 |
+
"""
|
| 277 |
+
收集 棋盘的 cells_xywh 对应的图片集合
|
| 278 |
+
"""
|
| 279 |
+
width = image.shape[1]
|
| 280 |
+
height = image.shape[0]
|
| 281 |
+
crop_cells: List[List[np.ndarray]] = []
|
| 282 |
+
|
| 283 |
+
for i in range(10):
|
| 284 |
+
row_cells = []
|
| 285 |
+
for j in range(9):
|
| 286 |
+
x, y, w, h = cells_xywh[i, j]
|
| 287 |
+
|
| 288 |
+
x_0 = max(int(x-w/2), 0)
|
| 289 |
+
y_0 = max(int(y-h/2), 0)
|
| 290 |
+
x_1 = min(int(x+w/2), width-1)
|
| 291 |
+
y_1 = min(int(y+h/2), height-1)
|
| 292 |
+
|
| 293 |
+
crop_img = image[y_0:y_1, x_0:x_1]
|
| 294 |
+
row_cells.append(crop_img)
|
| 295 |
+
crop_cells.append(row_cells)
|
| 296 |
+
|
| 297 |
+
return crop_cells
|
| 298 |
+
|
| 299 |
+
def draw_cells_box(image: cv2.UMat, cells_xywh: np.ndarray) -> cv2.UMat:
|
| 300 |
+
"""
|
| 301 |
+
绘制 棋盘的 cells_xywh 对应的 矩形框
|
| 302 |
+
"""
|
| 303 |
+
width = image.shape[1]
|
| 304 |
+
height = image.shape[0]
|
| 305 |
+
for i in range(10):
|
| 306 |
+
for j in range(9):
|
| 307 |
+
x, y, w, h = cells_xywh[i, j]
|
| 308 |
+
|
| 309 |
+
x_0 = max(int(x-w/2), 0)
|
| 310 |
+
y_0 = max(int(y-h/2), 0)
|
| 311 |
+
x_1 = min(int(x+w/2), width-1)
|
| 312 |
+
y_1 = min(int(y+h/2), height-1)
|
| 313 |
+
|
| 314 |
+
cv2.rectangle(image,(x_0, y_0), (x_1, y_1), (0, 0, 255), 1)
|
| 315 |
+
|
| 316 |
+
return image
|
core/helper_cls.py
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
|
| 3 |
+
dict_cate_names = {
|
| 4 |
+
'point': '.',
|
| 5 |
+
'other': 'x',
|
| 6 |
+
'red_king': 'K',
|
| 7 |
+
'red_advisor': 'A',
|
| 8 |
+
'red_bishop': 'B',
|
| 9 |
+
'red_knight': 'N',
|
| 10 |
+
'red_rook': 'R',
|
| 11 |
+
'red_cannon': 'C',
|
| 12 |
+
'red_pawn': 'P',
|
| 13 |
+
'black_king': 'k',
|
| 14 |
+
'black_advisor': 'a',
|
| 15 |
+
'black_bishop': 'b',
|
| 16 |
+
'black_knight': 'n',
|
| 17 |
+
'black_rook': 'r',
|
| 18 |
+
'black_cannon': 'c',
|
| 19 |
+
'black_pawn': 'p',
|
| 20 |
+
}
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
full_cate_names = list(dict_cate_names.keys())
|
core/kpt_34_with_xanything.py
ADDED
|
@@ -0,0 +1,197 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import cv2
|
| 2 |
+
import os
|
| 3 |
+
import json
|
| 4 |
+
import numpy as np
|
| 5 |
+
|
| 6 |
+
from .helper_34 import BONE_NAMES
|
| 7 |
+
|
| 8 |
+
class Shape:
|
| 9 |
+
|
| 10 |
+
@staticmethod
|
| 11 |
+
def init_from_dict(data: dict):
|
| 12 |
+
shape_ins = Shape(data["label"], data["points"], data["group_id"], data["shape_type"])
|
| 13 |
+
|
| 14 |
+
return shape_ins
|
| 15 |
+
|
| 16 |
+
def __init__(self, label="", points=None, group_id=1, shape_type=""):
|
| 17 |
+
self.label = label
|
| 18 |
+
self.score = None
|
| 19 |
+
self.points = points
|
| 20 |
+
self.group_id = group_id
|
| 21 |
+
self.description = ""
|
| 22 |
+
self.difficult = False
|
| 23 |
+
self.shape_type = shape_type
|
| 24 |
+
self.flags = {}
|
| 25 |
+
self.attributes = {}
|
| 26 |
+
|
| 27 |
+
def to_dict(self):
|
| 28 |
+
return {
|
| 29 |
+
"label": self.label,
|
| 30 |
+
"score": self.score,
|
| 31 |
+
"points": self.points,
|
| 32 |
+
"group_id": self.group_id,
|
| 33 |
+
"description": self.description,
|
| 34 |
+
"difficult": self.difficult,
|
| 35 |
+
"shape_type": self.shape_type,
|
| 36 |
+
"flags": self.flags,
|
| 37 |
+
"attributes": self.attributes
|
| 38 |
+
}
|
| 39 |
+
|
| 40 |
+
class KeyPoint(Shape):
|
| 41 |
+
def __init__(self, label="", point_xy=list[float, float], group_id=1):
|
| 42 |
+
# 校验 point_xy 是否为 2 个元素的列表
|
| 43 |
+
if len(point_xy) != 2:
|
| 44 |
+
raise ValueError("point_xy 必须是一个包含 2 个元素的列表")
|
| 45 |
+
super().__init__(label, [point_xy], group_id, "point")
|
| 46 |
+
|
| 47 |
+
class Rectangle(Shape):
|
| 48 |
+
def __init__(self, label="A1", xyxy=list[float, float, float, float], group_id=1):
|
| 49 |
+
|
| 50 |
+
if len(xyxy) != 4:
|
| 51 |
+
raise ValueError("xyxy 必须是一个包含 4 个元素的列表")
|
| 52 |
+
|
| 53 |
+
"""
|
| 54 |
+
bbox [左上角坐标, 右上角坐标, 右下角坐标, 左下角坐标] [[x1,y1],[x2,y2],[x3,y3],[x4,y4]]
|
| 55 |
+
"""
|
| 56 |
+
x1, y1, x2, y2 = xyxy
|
| 57 |
+
|
| 58 |
+
bbox = [
|
| 59 |
+
[x1, y1],
|
| 60 |
+
[x2, y1],
|
| 61 |
+
[x2, y2],
|
| 62 |
+
[x1, y2]
|
| 63 |
+
]
|
| 64 |
+
|
| 65 |
+
super().__init__(label, bbox, group_id, "rectangle")
|
| 66 |
+
|
| 67 |
+
class Annotation:
|
| 68 |
+
@staticmethod
|
| 69 |
+
def init_from_dict(data: dict):
|
| 70 |
+
"""
|
| 71 |
+
从 dict 初始化 Annotation 类
|
| 72 |
+
"""
|
| 73 |
+
|
| 74 |
+
image_height = data["imageHeight"]
|
| 75 |
+
image_width = data["imageWidth"]
|
| 76 |
+
|
| 77 |
+
ann = Annotation(image_path=data["imagePath"], image_width=image_width, image_height=image_height)
|
| 78 |
+
|
| 79 |
+
for shape in data["shapes"]:
|
| 80 |
+
if shape["shape_type"] == "rectangle":
|
| 81 |
+
ann.add_shape(Rectangle.init_from_dict(shape))
|
| 82 |
+
elif shape["shape_type"] == "point":
|
| 83 |
+
ann.add_shape(KeyPoint.init_from_dict(shape))
|
| 84 |
+
|
| 85 |
+
return ann
|
| 86 |
+
|
| 87 |
+
def __init__(self, image_path="", image_width=-1, image_height=-1):
|
| 88 |
+
self.version = "2.4.4"
|
| 89 |
+
self.flags = {}
|
| 90 |
+
self.shapes = []
|
| 91 |
+
self.image_data = None
|
| 92 |
+
|
| 93 |
+
self.image_path = image_path
|
| 94 |
+
self.image_height = image_height
|
| 95 |
+
self.image_width = image_width
|
| 96 |
+
|
| 97 |
+
def add_shape(self, shape: Rectangle | KeyPoint):
|
| 98 |
+
self.shapes.append(shape.to_dict())
|
| 99 |
+
|
| 100 |
+
def to_dict(self):
|
| 101 |
+
if self.image_path == "":
|
| 102 |
+
raise ValueError("image_path 不能为空")
|
| 103 |
+
if self.image_height == -1 or self.image_width == -1:
|
| 104 |
+
raise ValueError("image_height 和 image_width 不能为 -1")
|
| 105 |
+
|
| 106 |
+
return {
|
| 107 |
+
"version": self.version,
|
| 108 |
+
"flags": self.flags,
|
| 109 |
+
"shapes": self.shapes,
|
| 110 |
+
"imagePath": self.image_path,
|
| 111 |
+
"imageData": self.image_data,
|
| 112 |
+
"imageHeight": self.image_height,
|
| 113 |
+
"imageWidth": self.image_width
|
| 114 |
+
}
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
def save_kpt_34_with_xanything(image_input: np.ndarray, image_ann_path, bbox: list[float, float, float, float], kpt_34: list[tuple[str, float, float]], save_dir: str):
|
| 118 |
+
"""
|
| 119 |
+
保存 34 个关键点 和 一个 bbox 到 xanything 的 json 文件
|
| 120 |
+
"""
|
| 121 |
+
x1, y1, x2, y2 = bbox
|
| 122 |
+
x1, y1, x2, y2 = float(x1), float(y1), float(x2), float(y2)
|
| 123 |
+
|
| 124 |
+
|
| 125 |
+
if image_input is None:
|
| 126 |
+
raise ValueError("image_input 不能为 None")
|
| 127 |
+
|
| 128 |
+
image_height, image_width = image_input.shape[:2]
|
| 129 |
+
|
| 130 |
+
|
| 131 |
+
# image_ann_path 缺省 .json
|
| 132 |
+
if not image_ann_path.endswith(".json"):
|
| 133 |
+
image_ann_path = image_ann_path + ".json"
|
| 134 |
+
|
| 135 |
+
# 读取 image_ann_path 的 文件名
|
| 136 |
+
file_name = os.path.basename(image_ann_path)
|
| 137 |
+
|
| 138 |
+
annotation = Annotation(file_name, image_width, image_height)
|
| 139 |
+
|
| 140 |
+
kpt_34_dict = {}
|
| 141 |
+
for bone_name, x, y in kpt_34:
|
| 142 |
+
kpt_34_dict[bone_name] = [float(x), float(y)]
|
| 143 |
+
|
| 144 |
+
for bone_name in BONE_NAMES:
|
| 145 |
+
x, y = kpt_34_dict[bone_name]
|
| 146 |
+
annotation.add_shape(KeyPoint(bone_name, [x, y]))
|
| 147 |
+
|
| 148 |
+
# 添加 bbox
|
| 149 |
+
annotation.add_shape(Rectangle("bbox", [x1, y1, x2, y2]))
|
| 150 |
+
|
| 151 |
+
ann_file_path = os.path.join(save_dir, file_name)
|
| 152 |
+
# 保存
|
| 153 |
+
with open(ann_file_path, "w") as f:
|
| 154 |
+
json.dump(annotation.to_dict(), f)
|
| 155 |
+
|
| 156 |
+
# 保存图片
|
| 157 |
+
image_input_rgb = image_input.copy()[:, :, ::-1]
|
| 158 |
+
|
| 159 |
+
# print('ann_file_path:', ann_file_path.replace(".json", ".jpg"))
|
| 160 |
+
|
| 161 |
+
cv2.imwrite(ann_file_path.replace(".json", ".jpg"), image_input_rgb)
|
| 162 |
+
|
| 163 |
+
def read_xanything_to_json(json_path) -> tuple[list[tuple[str, float, float]], list[float, float, float, float]]:
|
| 164 |
+
"""
|
| 165 |
+
读取 xanything 的 json 文件
|
| 166 |
+
"""
|
| 167 |
+
data = {}
|
| 168 |
+
with open(json_path, "r") as f:
|
| 169 |
+
data = json.load(f)
|
| 170 |
+
|
| 171 |
+
# data
|
| 172 |
+
annotation = Annotation.init_from_dict(data)
|
| 173 |
+
|
| 174 |
+
keypoints_34_dict: dict[str, list[float, float]] = {}
|
| 175 |
+
# x1, y1, x2, y2
|
| 176 |
+
bbox: list[float, float, float, float] = []
|
| 177 |
+
|
| 178 |
+
for shape in annotation.shapes:
|
| 179 |
+
if shape["shape_type"] == "point":
|
| 180 |
+
keypoints_34_dict[shape["label"]] = [shape["points"][0][0], shape["points"][0][1]]
|
| 181 |
+
elif shape["shape_type"] == "rectangle":
|
| 182 |
+
bbox = [shape["points"][0][0], shape["points"][0][1], shape["points"][2][0], shape["points"][2][1]]
|
| 183 |
+
|
| 184 |
+
keypoints_34: list[tuple[str, float, float]] = []
|
| 185 |
+
|
| 186 |
+
for item in BONE_NAMES:
|
| 187 |
+
keypoints_34.append((item, keypoints_34_dict[item][0], keypoints_34_dict[item][1]))
|
| 188 |
+
|
| 189 |
+
return keypoints_34, bbox
|
| 190 |
+
|
| 191 |
+
|
| 192 |
+
|
| 193 |
+
|
| 194 |
+
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
|
core/runonnx/base_onnx.py
ADDED
|
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import onnxruntime
|
| 2 |
+
import numpy as np
|
| 3 |
+
import cv2
|
| 4 |
+
from abc import ABC, abstractmethod
|
| 5 |
+
from typing import Any, Tuple, Union, List
|
| 6 |
+
|
| 7 |
+
class BaseONNX(ABC):
|
| 8 |
+
def __init__(self, model_path: str, input_size: Tuple[int, int]):
|
| 9 |
+
"""初始化ONNX模型基类
|
| 10 |
+
|
| 11 |
+
Args:
|
| 12 |
+
model_path (str): ONNX模型路径
|
| 13 |
+
input_size (tuple): 模型输入尺寸 (width, height)
|
| 14 |
+
"""
|
| 15 |
+
self.session = onnxruntime.InferenceSession(model_path)
|
| 16 |
+
self.input_name = self.session.get_inputs()[0].name
|
| 17 |
+
self.input_size = input_size
|
| 18 |
+
|
| 19 |
+
def load_image(self, image: Union[cv2.UMat, str]) -> cv2.UMat:
|
| 20 |
+
"""加载图像
|
| 21 |
+
|
| 22 |
+
Args:
|
| 23 |
+
image (Union[cv2.UMat, str]): 图像路径或cv2图像对象
|
| 24 |
+
|
| 25 |
+
Returns:
|
| 26 |
+
cv2.UMat: 加载的图像
|
| 27 |
+
"""
|
| 28 |
+
if isinstance(image, str):
|
| 29 |
+
return cv2.imread(image)
|
| 30 |
+
return image.copy()
|
| 31 |
+
|
| 32 |
+
@abstractmethod
|
| 33 |
+
def preprocess_image(self, img_bgr: cv2.UMat, *args, **kwargs) -> np.ndarray:
|
| 34 |
+
"""图像预处理抽象方法
|
| 35 |
+
|
| 36 |
+
Args:
|
| 37 |
+
img_bgr (cv2.UMat): BGR格式的输入图像
|
| 38 |
+
|
| 39 |
+
Returns:
|
| 40 |
+
np.ndarray: 预处理后的图像
|
| 41 |
+
"""
|
| 42 |
+
pass
|
| 43 |
+
|
| 44 |
+
@abstractmethod
|
| 45 |
+
def run_inference(self, image: np.ndarray) -> Any:
|
| 46 |
+
"""运行推理的抽象方法
|
| 47 |
+
|
| 48 |
+
Args:
|
| 49 |
+
image (np.ndarray): 预处理后的输入图像
|
| 50 |
+
|
| 51 |
+
Returns:
|
| 52 |
+
Any: 模型输出结果
|
| 53 |
+
"""
|
| 54 |
+
pass
|
| 55 |
+
|
| 56 |
+
@abstractmethod
|
| 57 |
+
def pred(self, image: Union[cv2.UMat, str], *args, **kwargs) -> Any:
|
| 58 |
+
"""预测的抽象方法
|
| 59 |
+
|
| 60 |
+
Args:
|
| 61 |
+
image (Union[cv2.UMat, str]): 输入图像或图像路径
|
| 62 |
+
|
| 63 |
+
Returns:
|
| 64 |
+
Any: 预测结果
|
| 65 |
+
"""
|
| 66 |
+
pass
|
| 67 |
+
|
| 68 |
+
@abstractmethod
|
| 69 |
+
def draw_pred(self, img: cv2.UMat, *args, **kwargs) -> cv2.UMat:
|
| 70 |
+
"""绘制预测结果的抽象方法
|
| 71 |
+
|
| 72 |
+
Args:
|
| 73 |
+
img (cv2.UMat): 要绘制的图像
|
| 74 |
+
|
| 75 |
+
Returns:
|
| 76 |
+
cv2.UMat: 绘制结果后的图像
|
| 77 |
+
"""
|
| 78 |
+
pass
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
def check_images_list(self, images: List[Union[cv2.UMat, str, np.ndarray]]):
|
| 82 |
+
"""
|
| 83 |
+
检查图像列表是否有效
|
| 84 |
+
"""
|
| 85 |
+
for image in images:
|
| 86 |
+
if not isinstance(image, cv2.UMat) and not isinstance(image, str) and not isinstance(image, np.ndarray):
|
| 87 |
+
raise ValueError("The images must be a list of cv2.UMat or str or np.ndarray.")
|
| 88 |
+
|
core/runonnx/full_classifier.py
ADDED
|
@@ -0,0 +1,203 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import onnxruntime
|
| 2 |
+
import numpy as np
|
| 3 |
+
import cv2
|
| 4 |
+
from typing import Tuple, List, Union
|
| 5 |
+
from .base_onnx import BaseONNX
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
def center_crop(image: np.ndarray, target_size: Tuple[int, int]) -> np.ndarray:
|
| 9 |
+
"""
|
| 10 |
+
Center crop the image to the target size.
|
| 11 |
+
|
| 12 |
+
Args:
|
| 13 |
+
image (np.ndarray): The input image.
|
| 14 |
+
target_size (Tuple[int, int]): The desired output size (height, width).
|
| 15 |
+
|
| 16 |
+
Returns:
|
| 17 |
+
np.ndarray: The cropped image.
|
| 18 |
+
"""
|
| 19 |
+
h, w, _ = image.shape
|
| 20 |
+
target_w, target_h = target_size
|
| 21 |
+
|
| 22 |
+
center_x = w // 2
|
| 23 |
+
center_y = h // 2
|
| 24 |
+
|
| 25 |
+
start_x = int(center_x - target_w // 2)
|
| 26 |
+
start_y = int(center_y - target_h // 2)
|
| 27 |
+
|
| 28 |
+
cropped_image = image[start_y:start_y + target_h, start_x:start_x + target_w]
|
| 29 |
+
|
| 30 |
+
return cropped_image
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
dict_cate_names = {
|
| 34 |
+
'point': '.',
|
| 35 |
+
'other': 'x',
|
| 36 |
+
'red_king': 'K',
|
| 37 |
+
'red_advisor': 'A',
|
| 38 |
+
'red_bishop': 'B',
|
| 39 |
+
'red_knight': 'N',
|
| 40 |
+
'red_rook': 'R',
|
| 41 |
+
'red_cannon': 'C',
|
| 42 |
+
'red_pawn': 'P',
|
| 43 |
+
'black_king': 'k',
|
| 44 |
+
'black_advisor': 'a',
|
| 45 |
+
'black_bishop': 'b',
|
| 46 |
+
'black_knight': 'n',
|
| 47 |
+
'black_rook': 'r',
|
| 48 |
+
'black_cannon': 'c',
|
| 49 |
+
'black_pawn': 'p',
|
| 50 |
+
}
|
| 51 |
+
|
| 52 |
+
class FULL_CLASSIFIER_ONNX(BaseONNX):
|
| 53 |
+
|
| 54 |
+
label_2_short = dict_cate_names
|
| 55 |
+
|
| 56 |
+
classes_labels = list(dict_cate_names.keys())
|
| 57 |
+
|
| 58 |
+
def __init__(self,
|
| 59 |
+
model_path,
|
| 60 |
+
# 输入图片大小
|
| 61 |
+
input_size=(280, 315), # (w, h)
|
| 62 |
+
# 图片裁剪大小
|
| 63 |
+
crop_size=(400, 450), # (w, h)
|
| 64 |
+
):
|
| 65 |
+
super().__init__(model_path, input_size)
|
| 66 |
+
|
| 67 |
+
self.crop_size = crop_size
|
| 68 |
+
|
| 69 |
+
|
| 70 |
+
def preprocess_image(self, img_bgr: cv2.UMat, is_rgb: bool = True):
|
| 71 |
+
|
| 72 |
+
if not is_rgb:
|
| 73 |
+
img_rgb = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2RGB)
|
| 74 |
+
else:
|
| 75 |
+
img_rgb = img_bgr
|
| 76 |
+
|
| 77 |
+
if img_rgb.shape[:2] != self.crop_size:
|
| 78 |
+
# 调整图片大小 执行 center crop
|
| 79 |
+
img_rgb = center_crop(img_rgb, self.crop_size) # dst_size = (w, h)
|
| 80 |
+
|
| 81 |
+
# resize 到 input_size
|
| 82 |
+
img_rgb = cv2.resize(img_rgb, self.input_size)
|
| 83 |
+
|
| 84 |
+
# normalize mean and std
|
| 85 |
+
img = (img_rgb - np.array([ 123.675, 116.28, 103.53])) / np.array([58.395, 57.12, 57.375])
|
| 86 |
+
|
| 87 |
+
img = img.astype(np.float32)
|
| 88 |
+
# 转换为浮点型并归一化
|
| 89 |
+
# img = img.astype(np.float32) / 255.0
|
| 90 |
+
|
| 91 |
+
# 调整维度顺序 (H,W,C) -> (C,H,W)
|
| 92 |
+
img = np.transpose(img, (2, 0, 1))
|
| 93 |
+
|
| 94 |
+
# 添加 batch 维度
|
| 95 |
+
img = np.expand_dims(img, axis=0)
|
| 96 |
+
|
| 97 |
+
return img
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
def run_inference(self, image: np.ndarray) -> np.ndarray:
|
| 101 |
+
"""
|
| 102 |
+
Run inference on the image.
|
| 103 |
+
|
| 104 |
+
Args:
|
| 105 |
+
image (np.ndarray): The image to run inference on.
|
| 106 |
+
|
| 107 |
+
Returns:
|
| 108 |
+
tuple: A tuple containing the detection results and labels.
|
| 109 |
+
"""
|
| 110 |
+
# 运行推理
|
| 111 |
+
outputs, = self.session.run(None, {self.input_name: image})
|
| 112 |
+
|
| 113 |
+
return outputs
|
| 114 |
+
|
| 115 |
+
def pred(self, image: List[Union[cv2.UMat, str]], is_rgb: bool = True) -> Tuple[List[List[str]], List[List[str]], List[List[float]], str]:
|
| 116 |
+
"""
|
| 117 |
+
Predict the detection results of the image.
|
| 118 |
+
|
| 119 |
+
Args:
|
| 120 |
+
image (cv2.UMat, str): The image to predict.
|
| 121 |
+
|
| 122 |
+
Returns:
|
| 123 |
+
|
| 124 |
+
"""
|
| 125 |
+
if isinstance(image, str):
|
| 126 |
+
img_bgr = cv2.imread(image)
|
| 127 |
+
is_rgb = False
|
| 128 |
+
else:
|
| 129 |
+
img_bgr = image.copy()
|
| 130 |
+
|
| 131 |
+
image = self.preprocess_image(img_bgr, is_rgb)
|
| 132 |
+
|
| 133 |
+
labels = self.run_inference(image)
|
| 134 |
+
|
| 135 |
+
# 校验 labels 的 shape
|
| 136 |
+
assert labels.shape[1:] == (90, 16)
|
| 137 |
+
|
| 138 |
+
# shape (90, 16)
|
| 139 |
+
first_batch_labels = labels[0]
|
| 140 |
+
|
| 141 |
+
# 获取置信度最高的标签
|
| 142 |
+
# list[int]
|
| 143 |
+
label_indexes = np.argmax(first_batch_labels, axis=-1).tolist()
|
| 144 |
+
|
| 145 |
+
# 将标签索引转换为标签
|
| 146 |
+
# list[str]
|
| 147 |
+
label_names = [self.classes_labels[index] for index in label_indexes]
|
| 148 |
+
|
| 149 |
+
# list[str]
|
| 150 |
+
label_short = [self.label_2_short[name] for name in label_names]
|
| 151 |
+
|
| 152 |
+
# 获取置信度, 根据 first_batch_labels 和 label_indexes
|
| 153 |
+
confidence = first_batch_labels[np.arange(first_batch_labels.shape[0]), label_indexes]
|
| 154 |
+
|
| 155 |
+
label_names_10x9 = [label_names[i*9:(i+1)*9] for i in range(10)]
|
| 156 |
+
label_short_10x9 = [label_short[i*9:(i+1)*9] for i in range(10)]
|
| 157 |
+
confidence_10x9 = [confidence[i*9:(i+1)*9] for i in range(10)]
|
| 158 |
+
|
| 159 |
+
|
| 160 |
+
layout_str = "\n".join(["".join(row) for row in label_short_10x9])
|
| 161 |
+
|
| 162 |
+
return label_names_10x9, label_short_10x9, confidence_10x9, layout_str
|
| 163 |
+
|
| 164 |
+
def draw_pred(self, image: cv2.UMat, label_index: int, label_name: str, label_short: str, confidence: float) -> cv2.UMat:
|
| 165 |
+
|
| 166 |
+
# 在图像上绘制预测结果
|
| 167 |
+
cv2.putText(image, f"{label_short} {confidence:.2f}", (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
|
| 168 |
+
|
| 169 |
+
return image
|
| 170 |
+
|
| 171 |
+
|
| 172 |
+
def draw_pred_with_result(self, image: cv2.UMat, results: List[Tuple[int, str, str, float]], cells_xyxy: np.ndarray, is_rgb: bool = True) -> cv2.UMat:
|
| 173 |
+
|
| 174 |
+
assert len(results) == cells_xyxy.shape[0]
|
| 175 |
+
|
| 176 |
+
if not is_rgb:
|
| 177 |
+
image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
|
| 178 |
+
|
| 179 |
+
for i, (label_index, label_name, label_short, confidence) in enumerate(results):
|
| 180 |
+
# 确保坐标是整数类型
|
| 181 |
+
x1, y1, x2, y2 = map(int, cells_xyxy[i])
|
| 182 |
+
|
| 183 |
+
if label_name.startswith('red'):
|
| 184 |
+
color = (180, 105, 255) # 粉红色
|
| 185 |
+
elif label_name.startswith('black'):
|
| 186 |
+
color = (0, 100, 50) # 黑色
|
| 187 |
+
else:
|
| 188 |
+
color = (0, 0, 255) # 蓝色
|
| 189 |
+
|
| 190 |
+
|
| 191 |
+
if confidence < 0.5:
|
| 192 |
+
# yellow
|
| 193 |
+
color = (255, 255, 0)
|
| 194 |
+
|
| 195 |
+
# confidence:.2f 仅保留两位小数 移除
|
| 196 |
+
|
| 197 |
+
label_str = f"{label_short} {confidence:.2f}" if confidence < 0.9 else f"{label_short}"
|
| 198 |
+
|
| 199 |
+
cv2.putText(image, label_str, (x1 + 8, y2 - 5), cv2.FONT_HERSHEY_SIMPLEX, 0.3, color, 1)
|
| 200 |
+
|
| 201 |
+
return image
|
| 202 |
+
|
| 203 |
+
|
core/runonnx/rtmdet.py
ADDED
|
@@ -0,0 +1,117 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
|
| 3 |
+
import numpy as np
|
| 4 |
+
import cv2
|
| 5 |
+
from typing import Tuple, List, Union
|
| 6 |
+
from .base_onnx import BaseONNX
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
class RTMDET_ONNX(BaseONNX):
|
| 10 |
+
|
| 11 |
+
def __init__(self, model_path, input_size=(640, 640)):
|
| 12 |
+
super().__init__(model_path, input_size)
|
| 13 |
+
|
| 14 |
+
def preprocess_image(self, img_bgr: cv2.UMat):
|
| 15 |
+
# 调整图片大小
|
| 16 |
+
img_bgr = cv2.resize(img_bgr, self.input_size)
|
| 17 |
+
|
| 18 |
+
# normalize mean and std
|
| 19 |
+
img = (img_bgr - np.array([103.53, 116.28, 123.675])) / np.array([57.375, 57.12, 58.395])
|
| 20 |
+
|
| 21 |
+
img = img.astype(np.float32)
|
| 22 |
+
# 转换为浮点型并归一化
|
| 23 |
+
# img = img.astype(np.float32) / 255.0
|
| 24 |
+
|
| 25 |
+
# 调整维度顺序 (H,W,C) -> (C,H,W)
|
| 26 |
+
img = np.transpose(img, (2, 0, 1))
|
| 27 |
+
|
| 28 |
+
# 添加 batch 维度
|
| 29 |
+
img = np.expand_dims(img, axis=0)
|
| 30 |
+
|
| 31 |
+
return img
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
def run_inference(self, image: np.ndarray):
|
| 35 |
+
"""
|
| 36 |
+
Run inference on the image.
|
| 37 |
+
|
| 38 |
+
Args:
|
| 39 |
+
image (np.ndarray): The image to run inference on.
|
| 40 |
+
|
| 41 |
+
Returns:
|
| 42 |
+
tuple: A tuple containing the detection results and labels.
|
| 43 |
+
"""
|
| 44 |
+
# 运行推理
|
| 45 |
+
outputs = self.session.run(None, {self.input_name: image})
|
| 46 |
+
|
| 47 |
+
"""
|
| 48 |
+
dets: 检测框 [batch, num_dets, [x1, y1, x2, y2, conf]] ([batch, num_dets, Reshape(dets_dim_2)])
|
| 49 |
+
labels: 标签 [batch,num_dets]
|
| 50 |
+
"""
|
| 51 |
+
dets, labels = outputs
|
| 52 |
+
|
| 53 |
+
return dets, labels
|
| 54 |
+
|
| 55 |
+
def pred(self, image: List[Union[cv2.UMat, str]]) -> Tuple[List[int], float]:
|
| 56 |
+
"""
|
| 57 |
+
Predict the detection results of the image.
|
| 58 |
+
|
| 59 |
+
Args:
|
| 60 |
+
image (cv2.UMat, str): The image to predict.
|
| 61 |
+
|
| 62 |
+
Returns:
|
| 63 |
+
xyxy (list[int, int, int, int]): The detection results.
|
| 64 |
+
conf (float): The confidence of the detection results.
|
| 65 |
+
"""
|
| 66 |
+
if isinstance(image, str):
|
| 67 |
+
img_bgr = cv2.imread(image)
|
| 68 |
+
else:
|
| 69 |
+
img_bgr = image.copy()
|
| 70 |
+
|
| 71 |
+
original_w, original_h = img_bgr.shape[1], img_bgr.shape[0]
|
| 72 |
+
|
| 73 |
+
image = self.preprocess_image(img_bgr)
|
| 74 |
+
dets, labels = self.run_inference(image)
|
| 75 |
+
|
| 76 |
+
# 获取置信度最高的检测框
|
| 77 |
+
# dets = dets[0][0]
|
| 78 |
+
# labels = labels[0][0]
|
| 79 |
+
|
| 80 |
+
x1, y1, x2, y2, conf = dets[0][0]
|
| 81 |
+
|
| 82 |
+
xyxy = [x1, y1, x2, y2]
|
| 83 |
+
|
| 84 |
+
xyxy = self.transform_xyxy_to_original(xyxy, original_w, original_h)
|
| 85 |
+
|
| 86 |
+
return xyxy, conf
|
| 87 |
+
|
| 88 |
+
def transform_xyxy_to_original(self, xyxy, original_w, original_h) -> List[int]:
|
| 89 |
+
"""
|
| 90 |
+
将检测框从输入图像的尺寸转换为原始图像的尺寸
|
| 91 |
+
"""
|
| 92 |
+
x1, y1, x2, y2 = xyxy
|
| 93 |
+
|
| 94 |
+
input_w, input_h = self.input_size
|
| 95 |
+
ratio_w, ratio_h = original_w / input_w, original_h / input_h
|
| 96 |
+
|
| 97 |
+
x1, y1, x2, y2 = x1 * ratio_w, y1 * ratio_h, x2 * ratio_w, y2 * ratio_h
|
| 98 |
+
# 转换为整数
|
| 99 |
+
x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)
|
| 100 |
+
|
| 101 |
+
return [x1, y1, x2, y2]
|
| 102 |
+
|
| 103 |
+
def draw_pred(self, img: cv2.UMat, xyxy: List[int], conf: float, is_rgb: bool = True) -> cv2.UMat:
|
| 104 |
+
"""
|
| 105 |
+
Draw the detection results on the image.
|
| 106 |
+
"""
|
| 107 |
+
|
| 108 |
+
if not is_rgb:
|
| 109 |
+
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
| 110 |
+
|
| 111 |
+
x1, y1, x2, y2 = xyxy
|
| 112 |
+
|
| 113 |
+
cv2.rectangle(img, (x1, y1), (x2, y2), (0, 0, 255), 2)
|
| 114 |
+
cv2.putText(img, f"{conf:.2f}", (x1, y1), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 2)
|
| 115 |
+
|
| 116 |
+
return img
|
| 117 |
+
|
core/runonnx/rtmpose.py
ADDED
|
@@ -0,0 +1,356 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import numpy as np
|
| 2 |
+
import cv2
|
| 3 |
+
from typing import Tuple, List, Union
|
| 4 |
+
from .base_onnx import BaseONNX
|
| 5 |
+
|
| 6 |
+
class RTMPOSE_ONNX(BaseONNX):
|
| 7 |
+
|
| 8 |
+
bone_names = [
|
| 9 |
+
"A0", "A1", "A2", "A3", "A4", "A5", "A6", "A7", "A8",
|
| 10 |
+
"J0", "J1", "J2", "J3", "J4", "J5", "J6", "J7", "J8",
|
| 11 |
+
"B0", "C0", "D0", "E0", "F0", "G0", "H0", "I0",
|
| 12 |
+
"B8", "C8", "D8", "E8", "F8", "G8", "H8", "I8",
|
| 13 |
+
]
|
| 14 |
+
|
| 15 |
+
def __init__(self, model_path, input_size=(256, 256), padding=1.25):
|
| 16 |
+
super().__init__(model_path, input_size)
|
| 17 |
+
self.padding = padding
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def get_bbox_center_scale(self, bbox: List[int]):
|
| 21 |
+
"""Convert bounding box to center and scale.
|
| 22 |
+
|
| 23 |
+
The center is the coordinates of the bbox center, and the scale is the
|
| 24 |
+
bbox width and height normalized by the padding factor.
|
| 25 |
+
|
| 26 |
+
Args:
|
| 27 |
+
bbox: Bounding box in format [x1, y1, x2, y2]
|
| 28 |
+
|
| 29 |
+
Returns:
|
| 30 |
+
tuple: A tuple containing:
|
| 31 |
+
- center (numpy.ndarray): Center coordinates [x, y]
|
| 32 |
+
- scale (numpy.ndarray): Scale [width, height]
|
| 33 |
+
"""
|
| 34 |
+
|
| 35 |
+
# Get bbox center
|
| 36 |
+
x1, y1, x2, y2 = bbox
|
| 37 |
+
center = np.array([(x1 + x2) / 2.0, (y1 + y2) / 2.0])
|
| 38 |
+
|
| 39 |
+
# Get bbox scale (width and height)
|
| 40 |
+
w = x2 - x1
|
| 41 |
+
h = y2 - y1
|
| 42 |
+
|
| 43 |
+
# Convert to scaled width/height
|
| 44 |
+
scale = np.array([w, h]) * self.padding
|
| 45 |
+
|
| 46 |
+
return center, scale
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
@staticmethod
|
| 50 |
+
def _rotate_point(pt: np.ndarray, angle_rad: float) -> np.ndarray:
|
| 51 |
+
"""Rotate a point by an angle.
|
| 52 |
+
|
| 53 |
+
Args:
|
| 54 |
+
pt (np.ndarray): 2D point coordinates (x, y) in shape (2, )
|
| 55 |
+
angle_rad (float): rotation angle in radian
|
| 56 |
+
|
| 57 |
+
Returns:
|
| 58 |
+
np.ndarray: Rotated point in shape (2, )
|
| 59 |
+
"""
|
| 60 |
+
|
| 61 |
+
sn, cs = np.sin(angle_rad), np.cos(angle_rad)
|
| 62 |
+
rot_mat = np.array([[cs, -sn], [sn, cs]])
|
| 63 |
+
return rot_mat @ pt
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
@staticmethod
|
| 67 |
+
def _get_3rd_point(a: np.ndarray, b: np.ndarray):
|
| 68 |
+
"""To calculate the affine matrix, three pairs of points are required. This
|
| 69 |
+
function is used to get the 3rd point, given 2D points a & b.
|
| 70 |
+
|
| 71 |
+
The 3rd point is defined by rotating vector `a - b` by 90 degrees
|
| 72 |
+
anticlockwise, using b as the rotation center.
|
| 73 |
+
|
| 74 |
+
Args:
|
| 75 |
+
a (np.ndarray): The 1st point (x,y) in shape (2, )
|
| 76 |
+
b (np.ndarray): The 2nd point (x,y) in shape (2, )
|
| 77 |
+
|
| 78 |
+
Returns:
|
| 79 |
+
np.ndarray: The 3rd point.
|
| 80 |
+
"""
|
| 81 |
+
direction = a - b
|
| 82 |
+
c = b + np.r_[-direction[1], direction[0]]
|
| 83 |
+
return c
|
| 84 |
+
|
| 85 |
+
|
| 86 |
+
@staticmethod
|
| 87 |
+
def get_warp_matrix(
|
| 88 |
+
center: np.ndarray,
|
| 89 |
+
scale: np.ndarray,
|
| 90 |
+
rot: float,
|
| 91 |
+
output_size: Tuple[int, int],
|
| 92 |
+
shift: Tuple[float, float] = (0., 0.),
|
| 93 |
+
inv: bool = False,
|
| 94 |
+
fix_aspect_ratio: bool = True,
|
| 95 |
+
) -> np.ndarray:
|
| 96 |
+
"""Calculate the affine transformation matrix that can warp the bbox area
|
| 97 |
+
in the input image to the output size.
|
| 98 |
+
|
| 99 |
+
Args:
|
| 100 |
+
center (np.ndarray[2, ]): Center of the bounding box (x, y).
|
| 101 |
+
scale (np.ndarray[2, ]): Scale of the bounding box
|
| 102 |
+
wrt [width, height].
|
| 103 |
+
rot (float): Rotation angle (degree).
|
| 104 |
+
output_size (np.ndarray[2, ] | list(2,)): Size of the
|
| 105 |
+
destination heatmaps.
|
| 106 |
+
shift (0-100%): Shift translation ratio wrt the width/height.
|
| 107 |
+
Default (0., 0.).
|
| 108 |
+
inv (bool): Option to inverse the affine transform direction.
|
| 109 |
+
(inv=False: src->dst or inv=True: dst->src)
|
| 110 |
+
fix_aspect_ratio (bool): Whether to fix aspect ratio during transform.
|
| 111 |
+
Defaults to True.
|
| 112 |
+
|
| 113 |
+
Returns:
|
| 114 |
+
np.ndarray: A 2x3 transformation matrix
|
| 115 |
+
"""
|
| 116 |
+
assert len(center) == 2
|
| 117 |
+
assert len(scale) == 2
|
| 118 |
+
assert len(output_size) == 2
|
| 119 |
+
assert len(shift) == 2
|
| 120 |
+
|
| 121 |
+
shift = np.array(shift)
|
| 122 |
+
src_w, src_h = scale[:2]
|
| 123 |
+
dst_w, dst_h = output_size[:2]
|
| 124 |
+
|
| 125 |
+
rot_rad = np.deg2rad(rot)
|
| 126 |
+
src_dir = RTMPOSE_ONNX._rotate_point(np.array([src_w * -0.5, 0.]), rot_rad)
|
| 127 |
+
dst_dir = np.array([dst_w * -0.5, 0.])
|
| 128 |
+
|
| 129 |
+
src = np.zeros((3, 2), dtype=np.float32)
|
| 130 |
+
src[0, :] = center + scale * shift
|
| 131 |
+
src[1, :] = center + src_dir + scale * shift
|
| 132 |
+
|
| 133 |
+
dst = np.zeros((3, 2), dtype=np.float32)
|
| 134 |
+
dst[0, :] = [dst_w * 0.5, dst_h * 0.5]
|
| 135 |
+
dst[1, :] = np.array([dst_w * 0.5, dst_h * 0.5]) + dst_dir
|
| 136 |
+
|
| 137 |
+
if fix_aspect_ratio:
|
| 138 |
+
src[2, :] = RTMPOSE_ONNX._get_3rd_point(src[0, :], src[1, :])
|
| 139 |
+
dst[2, :] = RTMPOSE_ONNX._get_3rd_point(dst[0, :], dst[1, :])
|
| 140 |
+
else:
|
| 141 |
+
src_dir_2 = RTMPOSE_ONNX._rotate_point(np.array([0., src_h * -0.5]), rot_rad)
|
| 142 |
+
dst_dir_2 = np.array([0., dst_h * -0.5])
|
| 143 |
+
src[2, :] = center + src_dir_2 + scale * shift
|
| 144 |
+
dst[2, :] = np.array([dst_w * 0.5, dst_h * 0.5]) + dst_dir_2
|
| 145 |
+
|
| 146 |
+
if inv:
|
| 147 |
+
warp_mat = cv2.getAffineTransform(np.float32(dst), np.float32(src))
|
| 148 |
+
else:
|
| 149 |
+
warp_mat = cv2.getAffineTransform(np.float32(src), np.float32(dst))
|
| 150 |
+
return warp_mat
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
def get_warp_size_with_input_size(self,
|
| 154 |
+
bbox_center: List[int],
|
| 155 |
+
bbox_scale: List[int],
|
| 156 |
+
inv: bool = False,
|
| 157 |
+
):
|
| 158 |
+
"""
|
| 159 |
+
获取仿射变换矩阵的输出尺寸
|
| 160 |
+
"""
|
| 161 |
+
|
| 162 |
+
w, h = self.input_size
|
| 163 |
+
warp_size = self.input_size
|
| 164 |
+
|
| 165 |
+
# 修正长宽比
|
| 166 |
+
scale_w, scale_h = bbox_scale
|
| 167 |
+
aspect_ratio = w / h
|
| 168 |
+
if scale_w > scale_h * aspect_ratio:
|
| 169 |
+
bbox_scale = [scale_w, scale_w / aspect_ratio]
|
| 170 |
+
else:
|
| 171 |
+
bbox_scale = [scale_h * aspect_ratio, scale_h]
|
| 172 |
+
|
| 173 |
+
# 计算仿射变换矩阵 确保数据类型正确
|
| 174 |
+
center = np.array(bbox_center, dtype=np.float32)
|
| 175 |
+
scale = np.array(bbox_scale, dtype=np.float32)
|
| 176 |
+
|
| 177 |
+
rot = 0.0 # 不考虑旋转
|
| 178 |
+
|
| 179 |
+
warp_mat = self.get_warp_matrix(center, scale, rot, output_size=warp_size, inv=inv)
|
| 180 |
+
|
| 181 |
+
return warp_mat
|
| 182 |
+
|
| 183 |
+
def topdown_affine(self, img: cv2.UMat, bbox_center: List[int], bbox_scale: List[int]):
|
| 184 |
+
"""简化版的 top-down 仿射变换函数
|
| 185 |
+
|
| 186 |
+
Args:
|
| 187 |
+
img: 输入图像
|
| 188 |
+
|
| 189 |
+
Returns:
|
| 190 |
+
变换后的图像
|
| 191 |
+
"""
|
| 192 |
+
|
| 193 |
+
warp_mat = self.get_warp_size_with_input_size(bbox_center, bbox_scale)
|
| 194 |
+
|
| 195 |
+
# 应用仿射变换
|
| 196 |
+
dst_img = cv2.warpAffine(img, warp_mat, self.input_size, flags=cv2.INTER_LINEAR)
|
| 197 |
+
|
| 198 |
+
return dst_img
|
| 199 |
+
|
| 200 |
+
|
| 201 |
+
# 获取每个关键点的最优预测位置
|
| 202 |
+
def get_simcc_maximum(self, simcc_x, simcc_y):
|
| 203 |
+
|
| 204 |
+
# 在最后一维上找到最大值的索引
|
| 205 |
+
x_indices = np.argmax(simcc_x[0], axis=1) # (34,)
|
| 206 |
+
y_indices = np.argmax(simcc_y[0], axis=1) # (34,)
|
| 207 |
+
|
| 208 |
+
|
| 209 |
+
input_w, input_h = self.input_size
|
| 210 |
+
|
| 211 |
+
# 将索引转换为实际坐标 (0-1之间)
|
| 212 |
+
x_coords = x_indices / (input_w * 2) # 归一化到0-1
|
| 213 |
+
y_coords = y_indices / (input_h * 2)
|
| 214 |
+
|
| 215 |
+
# 组合成坐标对
|
| 216 |
+
keypoints = np.stack([x_coords, y_coords], axis=1) # (34, 2)
|
| 217 |
+
|
| 218 |
+
# 获取每个点的置信度分数
|
| 219 |
+
scores = np.max(simcc_x[0], axis=1) * np.max(simcc_y[0], axis=1)
|
| 220 |
+
|
| 221 |
+
return keypoints, scores
|
| 222 |
+
|
| 223 |
+
|
| 224 |
+
|
| 225 |
+
def preprocess_image(self, img_bgr: cv2.UMat, bbox_center: List[int], bbox_scale: List[int]):
|
| 226 |
+
|
| 227 |
+
"""
|
| 228 |
+
预处理图像
|
| 229 |
+
|
| 230 |
+
Args:
|
| 231 |
+
img_bgr (cv2.UMat): 输入图像
|
| 232 |
+
bbox_center (list[int, int]): 边界框中心坐标 [x, y]
|
| 233 |
+
bbox_scale (list[int, int]): 边界框尺度 [w, h]
|
| 234 |
+
"""
|
| 235 |
+
|
| 236 |
+
affine_img_bgr = self.topdown_affine(img_bgr, bbox_center, bbox_scale)
|
| 237 |
+
|
| 238 |
+
# 转RGB并进行归一化
|
| 239 |
+
affine_img_rgb = cv2.cvtColor(affine_img_bgr, cv2.COLOR_BGR2RGB)
|
| 240 |
+
# normalize mean and std
|
| 241 |
+
affine_img_rgb_norm = (affine_img_rgb - np.array([123.675, 116.28, 103.53])) / np.array([58.395, 57.12, 57.375])
|
| 242 |
+
# 转换为浮点型并归一化
|
| 243 |
+
img = affine_img_rgb_norm.astype(np.float32)
|
| 244 |
+
# 调整维度顺序 (H,W,C) -> (C,H,W)
|
| 245 |
+
img = np.transpose(img, (2, 0, 1))
|
| 246 |
+
# 添加 batch 维度
|
| 247 |
+
img = np.expand_dims(img, axis=0)
|
| 248 |
+
|
| 249 |
+
return img
|
| 250 |
+
|
| 251 |
+
|
| 252 |
+
def run_inference(self, image: np.ndarray):
|
| 253 |
+
"""
|
| 254 |
+
Run inference on the image.
|
| 255 |
+
|
| 256 |
+
Args:
|
| 257 |
+
image (np.ndarray): The image to run inference on.
|
| 258 |
+
|
| 259 |
+
Returns:
|
| 260 |
+
tuple: A tuple containing the detection results and labels.
|
| 261 |
+
"""
|
| 262 |
+
# 运行推理
|
| 263 |
+
outputs = self.session.run(None, {self.input_name: image})
|
| 264 |
+
"""
|
| 265 |
+
simcc_x: float32[batch,MatMulsimcc_x_dim_1,512]
|
| 266 |
+
simcc_y: float32[batch,MatMulsimcc_x_dim_1,512]
|
| 267 |
+
"""
|
| 268 |
+
simcc_x, simcc_y = outputs
|
| 269 |
+
|
| 270 |
+
return simcc_x, simcc_y
|
| 271 |
+
|
| 272 |
+
def pred(self, image: List[Union[cv2.UMat, str]], bbox: List[int]) -> Tuple[np.ndarray, np.ndarray]:
|
| 273 |
+
"""
|
| 274 |
+
Predict the keypoints results of the image.
|
| 275 |
+
|
| 276 |
+
Args:
|
| 277 |
+
image (str | cv2.UMat): The image to predict.
|
| 278 |
+
bbox (list[int, int, int, int]): The bounding box to predict.
|
| 279 |
+
|
| 280 |
+
Returns:
|
| 281 |
+
keypoints (np.ndarray): The predicted keypoints.
|
| 282 |
+
scores (np.ndarray): The predicted scores.
|
| 283 |
+
"""
|
| 284 |
+
if isinstance(image, str):
|
| 285 |
+
img_bgr = cv2.imread(image)
|
| 286 |
+
else:
|
| 287 |
+
img_bgr = image.copy()
|
| 288 |
+
|
| 289 |
+
bbox_center, bbox_scale = self.get_bbox_center_scale(bbox)
|
| 290 |
+
|
| 291 |
+
image = self.preprocess_image(img_bgr, bbox_center, bbox_scale)
|
| 292 |
+
simcc_x, simcc_y = self.run_inference(image)
|
| 293 |
+
|
| 294 |
+
# 获取SimCC预测的最大值位置,返回关键点坐标��置信度分数
|
| 295 |
+
# 对应 width 和 height 为 input_size 的归一化,即 (256,256)
|
| 296 |
+
keypoints, scores = self.get_simcc_maximum(simcc_x, simcc_y)
|
| 297 |
+
|
| 298 |
+
# 将预测的关键点坐标从模型输出尺寸映射回原图尺寸
|
| 299 |
+
keypoints = self.transform_keypoints_to_original(keypoints, bbox_center, bbox_scale, self.input_size)
|
| 300 |
+
|
| 301 |
+
return keypoints, scores
|
| 302 |
+
|
| 303 |
+
def transform_keypoints_to_original(self, keypoints, center, scale, output_size):
|
| 304 |
+
"""
|
| 305 |
+
将预测的关键点坐标从模型输出尺寸映射回原图尺寸
|
| 306 |
+
|
| 307 |
+
Args:
|
| 308 |
+
keypoints: 预测的关键点坐标 [N, 2]
|
| 309 |
+
center: bbox中心点 [x, y]
|
| 310 |
+
scale: bbox尺度 [w, h]
|
| 311 |
+
output_size: 模型输入尺寸 (w, h)
|
| 312 |
+
|
| 313 |
+
Returns:
|
| 314 |
+
np.ndarray: 转换后的关键点坐标 [N, 2]
|
| 315 |
+
"""
|
| 316 |
+
target_coords = keypoints.copy()
|
| 317 |
+
|
| 318 |
+
# 将0-1的预测坐标转换为像素坐标, 256*256
|
| 319 |
+
target_coords[:, 0] = target_coords[:, 0] * output_size[0]
|
| 320 |
+
target_coords[:, 1] = target_coords[:, 1] * output_size[1]
|
| 321 |
+
|
| 322 |
+
# 计算仿射变换矩阵
|
| 323 |
+
warp_mat = self.get_warp_size_with_input_size(center, scale, inv=True)
|
| 324 |
+
|
| 325 |
+
# 转换为齐次坐标
|
| 326 |
+
ones = np.ones((len(target_coords), 1))
|
| 327 |
+
target_coords_homogeneous = np.hstack([target_coords, ones])
|
| 328 |
+
|
| 329 |
+
# 应用逆变换
|
| 330 |
+
original_keypoints = target_coords_homogeneous @ warp_mat.T
|
| 331 |
+
|
| 332 |
+
return original_keypoints
|
| 333 |
+
|
| 334 |
+
def draw_pred(self, img: cv2.UMat, keypoints: np.ndarray, scores: np.ndarray, is_rgb: bool = True) -> cv2.UMat:
|
| 335 |
+
"""
|
| 336 |
+
Draw the keypoints results on the image.
|
| 337 |
+
"""
|
| 338 |
+
|
| 339 |
+
if not is_rgb:
|
| 340 |
+
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
|
| 341 |
+
|
| 342 |
+
# 获取 随机的 34 中颜色
|
| 343 |
+
colors = np.random.randint(0, 256, (34, 3))
|
| 344 |
+
|
| 345 |
+
for i, (point, score) in enumerate(zip(keypoints, scores)):
|
| 346 |
+
if score > 0.3: # 设置置信度阈值
|
| 347 |
+
x, y = map(int, point)
|
| 348 |
+
# 使用不同颜色标注不同的关键点
|
| 349 |
+
color = colors[i]
|
| 350 |
+
|
| 351 |
+
cv2.circle(img, (x, y), 5, (int(color[0]), int(color[1]), int(color[2])), -1)
|
| 352 |
+
# 添加关键点索引标注
|
| 353 |
+
cv2.putText(img, self.bone_names[i], (x+5, y+5),
|
| 354 |
+
cv2.FONT_HERSHEY_SIMPLEX, 1.0, (int(color[0]), int(color[1]), int(color[2])), 1)
|
| 355 |
+
return img
|
| 356 |
+
|
requirements.txt
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
opencv-python
|