LaRI / tools.py
ruili3's picture
init LaRI demo
860c6b0
raw
history blame
11.3 kB
import argparse
import os
import torch
import numpy as np
import trimesh
from scipy.spatial.transform import Rotation
import torch.backends.cudnn as cudnn
import torch.nn.functional as F
from PIL import Image
from src.utils.vis import (
prob_to_mask,
colorize,
denormalize,
)
import numpy as np
from src.lari.model import LaRIModel, DinoSegModel
from rembg import remove
from plyfile import PlyData, PlyElement
import torchvision.transforms as transforms
LAYER_COLOR = [
[255, 190, 11], # FFFF0B
[251, 86, 7], # FB5607
[241, 91, 181], # F15BB5
[131, 56, 236], # 8338EC
[58, 134, 255], # 3A86FF
]
OPENGL = np.array([[1, 0, 0, 0], [0, -1, 0, 0], [0, 0, -1, 0], [0, 0, 0, 1]])
def save_point_cloud(pcd, rgb, filename, binary=True):
"""Save an RGB point cloud as a PLY file.
:paras
@pcd: Nx3 matrix, the XYZ coordinates
@rgb: Nx3 matrix, the rgb colors for each 3D point
"""
if rgb is None:
gray_concat = np.tile(np.array([128], dtype=np.uint8),
(pcd.shape[0], 3))
points_3d = np.hstack((pcd, gray_concat))
else:
assert pcd.shape[0] == rgb.shape[0]
points_3d = np.hstack((pcd, rgb))
python_types = (float, float, float, int, int, int)
npy_types = [('x', 'f4'), ('y', 'f4'), ('z', 'f4'), ('red', 'u1'),
('green', 'u1'), ('blue', 'u1')]
if binary is True:
# Format into Numpy structured array
vertices = []
for row_idx in range(points_3d.shape[0]):
cur_point = points_3d[row_idx]
vertices.append(
tuple(
dtype(point)
for dtype, point in zip(python_types, cur_point)))
vertices_array = np.array(vertices, dtype=npy_types)
el = PlyElement.describe(vertices_array, 'vertex')
# write
PlyData([el]).write(filename)
else:
x = np.squeeze(points_3d[:, 0])
y = np.squeeze(points_3d[:, 1])
z = np.squeeze(points_3d[:, 2])
r = np.squeeze(points_3d[:, 3])
g = np.squeeze(points_3d[:, 4])
b = np.squeeze(points_3d[:, 5])
ply_head = 'ply\n' \
'format ascii 1.0\n' \
'element vertex %d\n' \
'property float x\n' \
'property float y\n' \
'property float z\n' \
'property uchar red\n' \
'property uchar green\n' \
'property uchar blue\n' \
'end_header' % r.shape[0]
# ---- Save ply data to disk
np.savetxt(filename, np.column_stack[x, y, z, r, g, b], fmt='%f %f %f %d %d %d', header=ply_head, comments='')
def load_model(model_info, ckpt_path, device):
model = eval(model_info)
model.to(device)
model.eval()
# Load pretrained weights
ckpt = torch.load(ckpt_path, map_location=device, weights_only=False)
if "model" in ckpt:
model.load_state_dict(ckpt["model"], strict=False)
else:
model.load_state_dict(ckpt, strict=False)
return model
def process_image_custom(pil_image, resolution=512):
"""
Read an image, resize the long side to `resolution` and pad the short side with gray,
so that the final image is (resolution x resolution).
Returns:
padded_img (PIL.Image): The processed image.
crop_coords (tuple): (top, left, bottom, right) coordinates of the valid region.
original_size (tuple): (width, height) of the original image.
"""
pil_image = pil_image.convert("RGB")
original_width, original_height = pil_image.size
# If already at fixed resolution, no processing is needed.
if original_width == resolution and original_height == resolution:
crop_coords = (0, 0, resolution, resolution)
return pil_image, crop_coords, (original_width, original_height), pil_image
# Compute scaling factor based on the long side.
if original_width >= original_height:
# Width is the long side.
scale = resolution / float(original_width)
new_width = resolution
new_height = int(round(original_height * scale))
resized_img = pil_image.resize((new_width, new_height), Image.BILINEAR)
# Compute vertical padding.
pad_top = (resolution - new_height) // 2
pad_bottom = resolution - new_height - pad_top
pad_left, pad_right = 0, 0
else:
# Height is the long side.
scale = resolution / float(original_height)
new_height = resolution
new_width = int(round(original_width * scale))
resized_img = pil_image.resize((new_width, new_height), Image.BILINEAR)
# Compute horizontal padding.
pad_left = (resolution - new_width) // 2
pad_right = resolution - new_width - pad_left
pad_top, pad_bottom = 0, 0
# Create new image filled with black
padded_img = Image.new("RGB", (resolution, resolution), (0, 0, 0))
padded_img.paste(resized_img, (pad_left, pad_top))
# The valid region (crop) is where the resized image was pasted.
crop_coords = (pad_top, pad_left, pad_top + new_height, pad_left + new_width)
return padded_img, crop_coords, (original_width, original_height), pil_image
def process_image(pil_image, resolution=512):
"""
Process the image: apply custom resize/pad then convert to normalized tensor.
Returns:
img_tensor (torch.Tensor): Tensor of shape (1, 3, resolution, resolution).
crop_coords (tuple): (top, left, bottom, right) coordinates of the valid region.
original_size (tuple): (width, height) of the original image.
"""
padded_img, crop_coords, original_size, ori_img = process_image_custom(
pil_image, resolution
)
transform = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
]
)
img_tensor = transform(padded_img).unsqueeze(0)
ori_img_tensor = transform(ori_img).unsqueeze(0)
return img_tensor, ori_img_tensor, crop_coords, original_size
def post_process_output(input_tensor, crop_coords, original_size):
"""
Crop the input tensor using the crop_coords and then resize to the original image size.
Args:
input_tensor (torch.Tensor): Input with shape (H, W, L, C) where C is 1 or 3.
crop_coords (tuple): (top, left, bottom, right) coordinates for cropping.
original_size (tuple): (width, height) of the original image.
Returns:
processed_output (torch.Tensor): Output with shape (original_height, original_width, L, C).
"""
top, left, bottom, right = crop_coords
# Crop the input spatially: resulting shape (crop_h, crop_w, L, C)
cropped = input_tensor[top:bottom, left:right, ...]
crop_h, crop_w, L, C = cropped.shape
# New shape becomes (1, L * C, crop_h, crop_w)
reshaped = cropped.permute(2, 3, 0, 1).reshape(1, L * C, crop_h, crop_w)
# Unpack the original size (width, height) and use bilinear interpolation.
new_width, new_height = original_size
mode = "nearest" if L == 1 else "bilinear"
resized = F.interpolate(
reshaped, size=(new_height, new_width), mode=mode, align_corners=False
)
resized = resized.reshape(L, C, new_height, new_width)
# Permute to the output shape: (new_height, new_width, L, C)
processed_output = resized.permute(2, 3, 0, 1)
return processed_output
def get_masked_depth(lari_map, valid_mask, layer_id):
layer_id = max(0, layer_id)
lari_depth = lari_map[:, :, layer_id, 2].cpu().numpy() # H W
valid_mask = valid_mask[:, :, layer_id, 0].cpu().numpy() # H W
valid_values = lari_depth[valid_mask]
# Handle empty valid values
if valid_values.size == 0:
vis_depth_range = [0, 1]
else:
vis_depth_range = [valid_values.min(), valid_values.max()]
depth_image = Image.fromarray(
colorize(
lari_depth,
vis_depth_range[0],
vis_depth_range[1],
invalid_mask=~valid_mask,
cmap="Spectral",
)
).convert("RGB")
return depth_image
def save_to_glb(pts3d, color3d, path):
scene = trimesh.Scene()
pct = trimesh.PointCloud(pts3d, colors=color3d)
scene.add_geometry(pct)
rot_y = np.eye(4)
rot_y[:3, :3] = Rotation.from_euler("y", np.deg2rad(180)).as_matrix()
scene.apply_transform(np.linalg.inv(OPENGL @ rot_y))
outfile = os.path.join(path, "res.glb")
scene.export(file_obj=outfile)
return outfile
def get_point_cloud(pred, img, mask, first_layer_color="image", target_folder=None):
"""
pred h w l 3 - the point cloud
img: 3 h w - the colored image
mask: h w l - indicating the valid layers
n_samples: int - n of pts to sample and save
"""
ori_shape = pred.shape
pred = pred.cpu().numpy()
pred = pred.reshape(-1, 3) # M 3
color_palette = LAYER_COLOR[: min(len(LAYER_COLOR), ori_shape[-2])]
assert first_layer_color in ["image", "pseudo"]
# assign color to point clouds: [M,3] -> [M, 6]
img = torch.clip(denormalize(img).squeeze(0), 0.0, 1.0)
img = img.permute(1, 2, 0).unsqueeze(2).cpu().numpy() # H W 1 3
img = (img * 255.0).astype(np.uint8)
layered_color = np.array([[color_palette]]).astype(np.uint8) # 1 1 n_layer 3
layered_color = np.broadcast_to(
layered_color, (img.shape[0], img.shape[1], ori_shape[2], 3)
) # H W n_layer 3
if first_layer_color == "image":
layered_color[:, :, :1, :] = img
layered_color = layered_color.reshape(-1, 3)
valid_mask_arr = mask.squeeze().reshape(-1).cpu().numpy() # [H,W,layers] -> [M]
pred = pred[valid_mask_arr.astype(bool)]
layered_color = layered_color[valid_mask_arr.astype(bool)] # V,3
save_folder = target_folder if target_folder is not None else os.path.dirname(__file__)
ply_path = os.path.join(save_folder, "res.ply")
save_point_cloud(pred, layered_color, filename=ply_path)
glb_path = save_to_glb(pred, layered_color, save_folder)
return glb_path, ply_path
def removebg_crop(pil_input):
pil_input = remove(pil_input.convert("RGB"))
pil_np = np.array(pil_input)
alpha = pil_np[:, :, 3]
is_crop = (
False
if np.sum(alpha > 0.8 * 255) > 0.1 * (alpha.shape[0] * alpha.shape[1])
else True
)
# adjust object size to fit the image resolution
if is_crop:
width, height = pil_input.size
# adjust object size
output_np = np.array(pil_input)
alpha = output_np[:, :, 3]
bbox = np.argwhere(alpha > 0.8 * 255)
bbox = (
np.min(bbox[:, 1]),
np.min(bbox[:, 0]),
np.max(bbox[:, 1]),
np.max(bbox[:, 0]),
)
center = (bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2
size = max(bbox[2] - bbox[0], bbox[3] - bbox[1])
size = int(size * 1.5)
bbox = (
max(center[0] - size // 2, 0),
max(center[1] - size // 2, 0),
min(center[0] + size // 2, width),
min(center[1] + size // 2, height),
)
pil_input = pil_input.crop(bbox) # type: ignore
return pil_input