Improve model card: add pipeline tag, paper link, and GitHub repository link

README.md

@@ -1,3 +1,66 @@
----
-license: mit
----
+---
+license: mit
+pipeline_tag: image-feature-extraction
+---
+
+# Semantic Concentration for Self-Supervised Dense Representations Learning (CoTAP)
+
+[![arXiv](https://img.shields.io/badge/arXiv-Paper-b31b1b.svg?style=flat)](https://arxiv.org/abs/2509.09429) [![GitHub](https://img.shields.io/badge/GitHub-Code-blue.svg?logo=github)](https://github.com/KID-7391/CoTAP)
+
+This repository contains the official code for the paper [Semantic Concentration for Self-Supervised Dense Representations Learning](https://huggingface.co/papers/2509.09429), accepted by IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI 2025).
+
+## Abstract
+
+Recent advances in image-level self-supervised learning (SSL) have made significant progress, yet learning dense representations for patches remains challenging. Mainstream methods encounter an over-dispersion phenomenon that patches from the same instance/category scatter, harming downstream performance on dense tasks. This work reveals that image-level SSL avoids over-dispersion by involving implicit semantic concentration. Specifically, the non-strict spatial alignment ensures intra-instance consistency, while shared patterns, i.e., similar parts of within-class instances in the input space, ensure inter-image consistency. Unfortunately, these approaches are infeasible for dense SSL due to their spatial sensitivity and complicated scene-centric data. These observations motivate us to explore explicit semantic concentration for dense SSL. First, to break the strict spatial alignment, we propose to distill the patch correspondences. Facing noisy and imbalanced pseudo labels, we propose a noise-tolerant ranking loss. The core idea is extending the Average Precision (AP) loss to continuous targets, such that its decision-agnostic and adaptive focusing properties prevent the student model from being misled. Second, to discriminate the shared patterns from complicated scenes, we propose the object-aware filter to map the output space to an object-based space. Specifically, patches are represented by learnable prototypes of objects via cross-attention. Last but not least, empirical studies across various tasks soundly support the effectiveness of our method.
+
+## Overview
+
+<p align="center">
+  <img src="https://github.com/KID-7391/CoTAP/raw/main/overview-v2.png" alt="CoTAP Pipeline" width="100%"/>
+</p>
+
+## Checkpoints
+
+| Method | Backbone | Linear Seg (mIoU) | kNN Cls (Acc) | Linear Cls (Acc) | Download |
+| ------ | -------- | ----------------- | ------------- | ---------------- | -------- |
+| DINO+CoTAP    | VIT-S/16 | 50.5 | 78.0 | 79.1 | [link](https://huggingface.co/KID-7391/CoTAP/blob/main/checkpoints/ours/vits16/dino%2Bcotap.ckpt) |
+| Leopart+CoTAP | VIT-S/16 | 50.4 | 77.3 | 78.6 | [link](https://huggingface.co/KID-7391/CoTAP/blob/main/checkpoints/ours/vits16/leopart%2Bcotap.ckpt) |
+| iBOT+CoTAP    | VIT-S/16 | 51.6 | 78.4 | 79.4 | [link](https://huggingface.co/KID-7391/CoTAP/blob/main/checkpoints/ours/vits16/ibot%2Bcotap.ckpt) |
+| Mugs+CoTAP    | VIT-S/16 | 51.6 | 78.9 | 79.4 | [link](https://huggingface.co/KID-7391/CoTAP/blob/main/checkpoints/ours/vits16/mugs%2Bcotap.ckpt) |
+| DINO+CoTAP    | VIT-B/16 | 52.2 | 79.2 | 80.2 | [link](https://huggingface.co/KID-7391/CoTAP/blob/main/checkpoints/ours/vitb16/dino%2Bcotap.ckpt) |
+| iBOT+CoTAP    | VIT-B/16 | 53.9 | 80.1 | 80.8 | [link](https://huggingface.co/KID-7391/CoTAP/blob/main/checkpoints/ours/vitb16/ibot%2Bcotap.ckpt) |
+| Mugs+CoTAP    | VIT-B/16 | 52.1 | 80.8 | 81.7 | [link](https://huggingface.co/KID-7391/CoTAP/blob/main/checkpoints/ours/vitb16/mugs%2Bcotap.ckpt) |
+| DINO+CoTAP    | VIT-B/8  | 54.7 | 81.6 | 82.1 | [link](https://huggingface.co/KID-7391/CoTAP/blob/main/checkpoints/ours/vitb8/dino%2Bcotap.ckpt) |
+
+## Installation
+
+To set up the environment and install dependencies, please refer to the [Installation section on the GitHub repository](https://github.com/KID-7391/CoTAP#installation). This typically involves cloning the repository, creating a Python environment, and installing `requirements.txt`.
+
+## Training
+
+Detailed instructions for dataset preparation (ImageNet-1k, COCO, COCOStuff164k, NNS) and running training scripts can be found in the [Training section on the GitHub repository](https://github.com/KID-7391/CoTAP#training).
+
+## Evaluation
+
+For evaluation on datasets such as COCOStuff27, PASCAL-VOC, ADE20k, and Cityscapes, including kNN classification and semantic segmentation, please consult the [Evaluation section on the GitHub repository](https://github.com/KID-7391/CoTAP#evaluation).
+
+## Citation
+
+If you find this repository useful in your research, please cite the following paper:
+
+```bibtex
+@article{wen2025semantic,
+  title={Semantic Concentration for Self-Supervised Dense Representations Learning},
+  author={Wen, Peisong and Xu, Qianqian and Dai, Siran and Cong, Runmin and Huang, Qingming},
+  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
+  year={2025}
+}
+```
+
+## Contact us
+
+If you have any detailed questions or suggestions, you can email us: wenpeisong@ucas.ac.cn. We will reply in 1-2 business days. Thanks for your interest in our work!
+
+## Acknowledgements
+
+Our code is based on the official PyTorch implementation of [DINO](https://github.com/facebookresearch/dino), [Leopart](https://github.com/MkuuWaUjinga/leopart), and [STEGO](https://github.com/mhamilton723/STEGO).