Mask-Consistent Contrastive Learning

Introduction

The repository contains official Pytorch implementations of training and evaluation codes for Mask-Consistent Contrastive Learning.

Installation

1. Follow tutorial to install pytorch1.11.0(or newer version) and cudatookit.
2. Install mmcv, mmengine, mmcls and mmdet using MIM.

pip install -U openmim
mim install mmengine
mim install "mmcv==2.0.0rc3"
mim install "mmcls==1.0.0rc4"
mim install "mmdet==3.0.0rc4"

3. Install mmsegv1.0.0rc2 from source.

git clone https://github.com/CVIU-CSU/MaskCC.git
cd MaskCC
pip install -v -e .

4. Prepare datasets following tutorial.

Training

Cityscapes

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 PORT=29513 bash tools/dist_train.sh configs/maskcc/maskcc_r50_8xb2-90k_cityscapes-512x1024.py 8

ADE20K:

CUDA_VISIBLE_DEVICES=0,1,2,3 PORT=29513 bash tools/dist_train.sh configs/maskcc/maskcc_r50_4xb4-160k_ade20k-512x512.py 4

PASCAL-Context

CUDA_VISIBLE_DEVICES=0,1,2,3 PORT=29513 bash tools/dist_train.sh configs/maskcc/maskcc_r50_4xb4-60k_pascal-context_480x480.py 4

COCO-Stuff10k

CUDA_VISIBLE_DEVICES=0,1,2,3 PORT=29513 bash tools/dist_train.sh configs/maskcc/maskcc_r50_4xb4-60k_cocostuff_480x480.py 4

Evalution

For single-scale test:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 PORT=29513 bash tools/dist_test.sh configs/maskcc/maskcc_r50_8xb2-90k_cityscapes-512x1024.py 8

For multi-scale test: The multi-scale test for mask2former is not supported in mmsegv1.0.0rc2.

Results

Method	Backbone	Train Set	Eval Set	Batch	Iters	mIoU	Config
Mask2Former	ResNet50	Cityscapes train	Cityscapes val	8x2	90K	79.4	config
Mask2Former+MaskCC	ResNet50	Cityscapes train	Cityscapes val	8x2	90K	80.9	config
Mask2Former	ResNet50	ADE20K train	ADE20K val	4x4	160K	47.2	config
Mask2Former+MaskCC	ResNet50	ADE20K train	ADE20K val	4x4	160K	48.4	config
Mask2Former	ResNet50	PASCAL-Context train	PASCAL-Context val	4x4	60K	54.8	config
Mask2Former+MaskCC	ResNet50	PASCAL-Context train	PASCAL-Context val	4x4	60K	55.3	config
Mask2Former	ResNet50	COCO-Stuff10k train	COCO-Stuff10k val	4x4	60K	40.0	config
Mask2Former+MaskCC	ResNet50	COCO-Stuff10k train	COCO-Stuff10k val	4x4	60K	41.2	config

Todo

• MaskCC code
• Configs
• Detailed readme
• Citation