Awesome
<h1 align="left">RegionCL: Can Simple Region Swapping Contribute to Contrastive Learning? <a href="https://arxiv.org/abs/2111.12309"><img src="https://img.shields.io/badge/arXiv-Paper-<COLOR>.svg" ></a> </a> </h1> <p align="center"> <a href="#introduction">Introduction</a> | <a href="#results">Results</a> | <a href="#installation">Installation</a> | <a href="#news">News</a> </p>Introduction
<p align="justify">This repository contains the code, models, test results for the paper <a href="https://arxiv.org/abs/2111.12309">RegionCL: Can Simple Region Swapping Contribute to Contrastive Learning?</a>. It contains a simple region swapping module to construct both global- and region-level contrastive pairs with minor modifications to current self-supervised methods, e.g., MoCov2, DenseCL, SimSiam, and so on. <img src="demo/NetworkStructure.png">Results
We test RegionCL-M/D/S on the ImageNet dataset for classification, MS COCO dataset for detection, instance segmentation and human pose estimation, Cityscapes dataset for instance and semantic segmentation, and AP-10K dataset for animal pose estimation. The results and training logs are available below.
Pretrain & Classification on ImageNet-1K
| Model | Pretrain | IN1K Linear |
|---|---|---|
| RegionCL-M | log | config | 69.4 | log |
| RegionCL-D | N/A | 68.5 | N/A |
| RegionCL-S | N/A | 71.3 | N/A |
MaskRCNN detection and segmentation on MS COCO
| Model | MS COCO Det&Seg C4 1x | MS COCO Det&Seg FPN 1x |
|---|---|---|
| RegionCL-M | 39.8&34.8 | log | 40.1&36.3 | log |
| RegionCL-D | 40.3&35.2 | N/A | 40.4&36.7 | N/A |
| RegionCL-S | 38.7&33.7 | N/A | 38.8&35.2 | N/A |
| Model | MS COCO Det&Seg C4 2x | MS COCO Det&Seg FPN 2x |
|---|---|---|
| RegionCL-M | 41.5&35.9 | log | 41.6&37.7 | log |
| RegionCL-D | 41.8&36.4 | N/A | 42.1&38.0 | N/A |
| RegionCL-S | 40.7&35.4 | N/A | 41.0&37.1 | N/A |
RetinaNet detection on MS COCO
| Model | MS COCO Det 1x | MS COCO Det 2x |
|---|---|---|
| RegionCL-M | 38.4 | log | 40.1 | log |
| RegionCL-D | 38.8 | N/A | 40.6 | N/A |
| RegionCL-S | 36.8 | N/A | 39.1 | N/A |
Instance and semantic segmentation on Cityscapes
| Model | MaskRCNN Inst-Seg | UperNet Sem-Seg 40K | UperNet Sem-Seg 80K |
|---|---|---|---|
| RegionCL-M | 34.9 | log | 78.1 | log | 79.0 | log |
| RegionCL-D | 34.8 | N/A | 78.7 | N/A | 79.5 | N/A |
| RegionCL-S | 34.9 | N/A | 77.8 | N/A | 78.7 | N/A |
SimpleBaseline pose estimation on MS COCO and AP-10K
| Model | Human Pose | Animal Pose |
|---|---|---|
| RegionCL-M | 72.3 | 70.6 |
| RegionCL-D | 73.6 | 72.1 |
| RegionCL-S | 72.2 | 71.6 |
Installation
The code is based on Openselfsup, thanks for their wonderful work!
Requirements:
- Python 3.6.5+
- Pytorch (version 1.7.0)
- mmcv (version 1.0.3)
-
Install mmcv following the requirements as in link
-
Clone this repository
git clone https://github.com/Annbless/RegionCL.git -
Go into the repository
cd RegionCL -
Install this repository
pip install -v -e .
News
2021/12/30 Release the code and results for RegionCL-M
To do
- Release the code and logs for RegionCL-D and RegionCL-S
- Release the pretrained models
- Release the RegionCL with more self-supervised methods