Clothing Segmentation — SegFormer-B0
Binary clothing segmentation model fine-tuned on the iMaterialist Fashion 2020 (FGVC7) dataset. Separates worn clothing from images of people at the pixel level — built for use cases like virtual fitting rooms, fashion cataloging, and outfit-swap applications.
Full training code, technical report, and documentation: GitHub repository
Performance
| Metric | Score |
|---|---|
| Dice Score | 0.926 |
| Mean IoU | 0.878 |
| Pixel Accuracy | 0.976 |
Model Details
- Architecture: SegFormer-B0 (
nvidia/segformer-b0-finetuned-ade-512-512), fine-tuned with a binary (1-channel) segmentation head - Task framing: Binary segmentation (clothing vs. background) rather than multi-class, to match real-world virtual-fitting-room use cases
- Loss function: Combined Dice + Binary Cross-Entropy (BCE) loss, to handle background-pixel-dominant masks
- Input resolution: 512×512
- Training data: 3,000-image subset of iMaterialist Fashion 2020 (FGVC7), 85/15 train/validation split
- Training hardware: Kaggle free-tier T4 GPU, mixed precision training
Available Checkpoints
| File | Contains | Use case |
|---|---|---|
best_model_inference_only.pth |
Model weights only | Evaluation and inference (recommended) |
best_model_full.pth |
Model weights + optimizer state | Resuming training from this checkpoint |
Usage
Download the checkpoint:
wget https://huggingface.co/UseItOrLoseIt/clothing-segmentation-segformer-b0/resolve/main/best_model_inference_only.pth -O checkpoints/best_model.pth
Then, using the inference code from the GitHub repo:
python inference.py --image_path your_photo.jpg --output_path result.png
Limitations
- Performance degrades in low-light or backlit scenes
- Under-segments non-standard garment shapes (e.g., flared or trailing hemlines)
- Minor boundary imprecision at low-contrast garment-to-garment transitions (e.g., dark pants meeting dark boots)
- Edge sharpness is good but not pixel-perfect, a tradeoff of SegFormer-B0's lightweight decoder
Full analysis with annotated visual examples: report.md
Citation
If you use this model, please reference the GitHub repository.