mmsegmentation/projects/CAT-Seg

CAT-Seg

CAT-Seg: Cost Aggregation for Open-Vocabulary Semantic Segmentation

Introduction

Official Repo

Code Snippet

Abstract

Existing works on open-vocabulary semantic segmentation have utilized large-scale vision-language models, such as CLIP, to leverage their exceptional open-vocabulary recognition capabilities. However, the problem of transferring these capabilities learned from image-level supervision to the pixel-level task of segmentation and addressing arbitrary unseen categories at inference makes this task challenging. To address these issues, we aim to attentively relate objects within an image to given categories by leveraging relational information among class categories and visual semantics through aggregation, while also adapting the CLIP representations to the pixel-level task. However, we observe that direct optimization of the CLIP embeddings can harm its open-vocabulary capabilities. In this regard, we propose an alternative approach to optimize the imagetext similarity map, i.e. the cost map, using a novel cost aggregation-based method. Our framework, namely CATSeg, achieves state-of-the-art performance across all benchmarks. We provide extensive ablation studies to validate our choices. Project page.

CAT-Seg model structure

Usage

CAT-Seg model training needs pretrained CLIP model. We have implemented ViT-B and ViT-L based CLIP model. To further use ViT-bigG or ViT-H ones, you need additional dependencies. Please install open_clip first. The pretrained CLIP model state dicts are loaded from Huggingface-OpenCLIP. If you come up with ConnectionError when downloading CLIP weights, you can manually download them from the given repo and use custom_clip_weights=/path/to/you/folder of backbone in config file. Related tools are as shown in requirements/optional.txt:

pip install ftfy==6.0.1
pip install huggingface-hub
pip install regex

In addition to the necessary data preparation, you also need class texts for clip text encoder. Please download the class text json file first cls_texts and arrange the folder as follows:

mmsegmentation
├── mmseg
├── tools
├── configs
├── data
│   ├── VOCdevkit
│   │   ├── VOC2012
│   │   ├── VOC2010
│   │   ├── VOCaug
│   ├── ade
│   ├── coco_stuff164k
│   ├── coco.json
│   ├── pc59.json
│   ├── pc459.json
│   ├── ade150.json
│   ├── ade847.json
│   ├── voc20b.json
│   ├── voc20.json

# setup PYTHONPATH
export PYTHONPATH=`pwd`:$PYTHONPATH
# run evaluation
mim test mmsegmentation ${CONFIG} --checkpoint ${CHECKPOINT} --launcher pytorch --gpus=8

Results and models

ADE20K-150-ZeroShot

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
CAT-Seg	R-101 & ViT-B	384x384	80000	-	-	RTX3090	27.2	-	config	model

Note:

• All experiments of CAT-Seg are implemented with 4 RTX3090 GPUs, except the last one with pretrained ViT-bigG CLIP model (GPU Memory insufficient, you may need A100).
• Due to the feature size bottleneck of the CLIP image encoder, the inference and testing can only be done under slide mode, the inference time is longer since the test size is much more bigger that training size of (384, 384).
• The ResNet backbones utilized in CAT-Seg models are standard ResNet rather than ResNetV1c.
• The zero-shot segmentation results on PASCAL VOC and ADE20K are from the original paper. Our results are coming soon. We appreatiate your contribution!
• In additional to zero-shot segmentation performance results, we also provided the evaluation results on the val2017 set of COCO-stuff164k for reference, which is the training dataset of CAT-Seg. The testing was done without TTA.
• The number behind the dataset name is the category number for segmentation evaluation (except training data COCO-stuff 164k). PASCAL VOC-20b defines the "background" as classes present in PASCAL-Context-59 but not in PASCAL VOC-20.

Citation

@inproceedings{cheng2021mask2former,
  title={CAT-Seg: Cost Aggregation for Open-Vocabulary Semantic Segmentation},
  author={Seokju Cho and Heeseong Shin and Sunghwan Hong and Seungjun An and Seungjun Lee and Anurag Arnab and Paul Hongsuck Seo and Seungryong Kim},
  journal={CVPR},
  year={2023}
}