File size: 7,240 Bytes

---
library_name: transformers
tags:
- vision
- image-segmentation
- ecology
datasets:
- coralscapes
metrics:
- mean_iou
license: apache-2.0
---

# Model Card for Model ID

SegFormer model with a MiT-B5 backbone fine-tuned on Coralscapes at resolution 1024x1024, as introduced in [The Coralscapes Dataset: Semantic Scene Understanding in Coral Reefs](https://arxiv.org/abs/2503.20000).


## Model Details

### Model Description

- **Model type:** SegFormer
- **Finetuned from model:** [SegFormer (b5-sized) encoder pre-trained-only (`nvidia/mit-b5`)](https://huggingface.co/nvidia/mit-b5)

### Model Sources 

- **Repository:** [coralscapesScripts](https://github.com/eceo-epfl/coralscapesScripts/)
- **Demo** [Hugging Face Spaces](https://huggingface.co/spaces/EPFL-ECEO/coralscapes_demo):

## How to Get Started with the Model

The simplest way to use this model to segment an image of the Coralscapes dataset is as follows:

```python
from transformers import SegformerImageProcessor, SegformerForSemanticSegmentation
from PIL import Image
from datasets import load_dataset

# Load an image from the coralscapes dataset or load your own image 
dataset = load_dataset("EPFL-ECEO/coralscapes") 
image = dataset["test"][42]["image"]

preprocessor = SegformerImageProcessor.from_pretrained("EPFL-ECEO/segformer-b5-finetuned-coralscapes-1024-1024")
model = SegformerForSemanticSegmentation.from_pretrained("EPFL-ECEO/segformer-b5-finetuned-coralscapes-1024-1024")

inputs = preprocessor(image, return_tensors = "pt")
outputs = model(**inputs)
outputs = preprocessor.post_process_semantic_segmentation(outputs, target_sizes=[(image.size[1], image.size[0])])
label_pred = outputs[0].numpy()
```

While using the above approach should still work for images of different sizes and scales, for images that are not close to the training size of the model (1024x1024), 
we recommend using the following approach using a sliding window to achieve better results:

```python
import torch 
import torch.nn.functional as F
from transformers import SegformerImageProcessor, SegformerForSemanticSegmentation
from PIL import Image
import numpy as np
from datasets import load_dataset
device = 'cuda' if torch.cuda.is_available() else 'cpu'

def resize_image(image, target_size=1024):
    """
    Used to resize the image such that the smaller side equals 1024
    """
    h_img, w_img = image.size
    if h_img < w_img:
        new_h, new_w = target_size, int(w_img * (target_size / h_img))
    else:
        new_h, new_w  = int(h_img * (target_size / w_img)), target_size
    resized_img = image.resize((new_h, new_w))
    return resized_img

def segment_image(image, preprocessor, model, crop_size = (1024, 1024), num_classes = 40, transform=None):
    """
    Finds an optimal stride based on the image size and aspect ratio to create
    overlapping sliding windows of size 1024x1024 which are then fed into the model.  
    """ 
    h_crop, w_crop = crop_size
    
    img = torch.Tensor(np.array(resize_image(image, target_size=1024)).transpose(2, 0, 1)).unsqueeze(0)
    batch_size, _, h_img, w_img = img.size()
    
    if transform:
        img = torch.Tensor(transform(image = img.numpy())["image"]).to(device)    
        
    h_grids = int(np.round(3/2*h_img/h_crop)) if h_img > h_crop else 1
    w_grids = int(np.round(3/2*w_img/w_crop)) if w_img > w_crop else 1
    
    h_stride = int((h_img - h_crop + h_grids -1)/(h_grids -1)) if h_grids > 1 else h_crop
    w_stride = int((w_img - w_crop + w_grids -1)/(w_grids -1)) if w_grids > 1 else w_crop
    
    preds = img.new_zeros((batch_size, num_classes, h_img, w_img))
    count_mat = img.new_zeros((batch_size, 1, h_img, w_img))
    
    for h_idx in range(h_grids):
        for w_idx in range(w_grids):
            y1 = h_idx * h_stride
            x1 = w_idx * w_stride
            y2 = min(y1 + h_crop, h_img)
            x2 = min(x1 + w_crop, w_img)
            y1 = max(y2 - h_crop, 0)
            x1 = max(x2 - w_crop, 0)
            crop_img = img[:, :, y1:y2, x1:x2]
            with torch.no_grad():
                if(preprocessor):
                    inputs = preprocessor(crop_img, return_tensors = "pt")
                    inputs["pixel_values"] = inputs["pixel_values"].to(device)
                else:
                    inputs = crop_img.to(device)
                outputs = model(**inputs)

            resized_logits = F.interpolate(
                outputs.logits[0].unsqueeze(dim=0), size=crop_img.shape[-2:], mode="bilinear", align_corners=False
            )
            preds += F.pad(resized_logits,
                            (int(x1), int(preds.shape[3] - x2), int(y1),
                            int(preds.shape[2] - y2))).cpu()
            count_mat[:, :, y1:y2, x1:x2] += 1
        
    assert (count_mat == 0).sum() == 0
    preds = preds / count_mat
    preds = preds.argmax(dim=1)
    preds = F.interpolate(preds.unsqueeze(0).type(torch.uint8), size=image.size[::-1], mode='nearest')
    label_pred = preds.squeeze().cpu().numpy()
    return label_pred

# Load an image from the coralscapes dataset or load your own image 
dataset = load_dataset("EPFL-ECEO/coralscapes") 
image = dataset["test"][42]["image"]

preprocessor = SegformerImageProcessor.from_pretrained("EPFL-ECEO/segformer-b5-finetuned-coralscapes-1024-1024")
model = SegformerForSemanticSegmentation.from_pretrained("EPFL-ECEO/segformer-b5-finetuned-coralscapes-1024-1024")

label_pred = segment_image(image, preprocessor, model)
```

## Training & Evaluation Details

### Data

The model is trained and evaluated on the [Coralscapes dataset](https://huggingface.co/datasets/EPFL-ECEO/coralscapes) which is a general-purpose dense semantic segmentation dataset for coral reefs. 


### Procedure

Training mostly follows the Segformer original [implementation](https://proceedings.neurips.cc/paper_files/paper/2021/file/64f1f27bf1b4ec22924fd0acb550c235-Paper.pdf), outside of using stronger augmentations during training. 
We used a batch size of 4 for 100 epochs with the AdamW optimizer with an initial learning rate of 6e-5, weight decay of 1e-2 and polynomial learning rate scheduler with a power of 1. 
During training, images are randomly scaled within a range of 1.02 and 2 with aspect ratio 3/4 to 4/3, include random rotations of up to 15 degrees (cropped to exclude non-image areas), 
as well as random color jitters of contrast, saturation, and brightness between 0.8 and 1.2, and hue changes between -0.05 and 0.05. 
Input images are normalized using the ImageNet mean and standard deviation. For evaluation, a non-overlapping sliding window strategy is employed, 
using a window size of 1024x1024. 

### Results

- Test Accuracy:  82.761
- Test Mean IoU: 57.800

## Citation
If you find this project useful, please consider citing:
```bibtex
@misc{sauder2025coralscapesdatasetsemanticscene,
        title={The Coralscapes Dataset: Semantic Scene Understanding in Coral Reefs}, 
        author={Jonathan Sauder and Viktor Domazetoski and Guilhem Banc-Prandi and Gabriela Perna and Anders Meibom and Devis Tuia},
        year={2025},
        eprint={2503.20000},
        archivePrefix={arXiv},
        primaryClass={cs.CV},
        url={https://arxiv.org/abs/2503.20000}, 
  }
```