levindabhi/cloth-segmentation

Built on the U2NET architecture, this project extends the salient object detection model to output 4-channel segmentation masks (upper body, lower body, full body, and background) trained on 45k images from the iMaterialist Fashion dataset. The implementation provides both training code with multi-GPU distributed support via PyTorch and inference capabilities, handling diverse backgrounds and poses without specialized preprocessing. Pre-trained weights are available for immediate use, and the codebase includes TensorBoard visualization for monitoring training metrics.

668 stars. No commits in the last 6 months.