CLIP-VIS: Adapting CLIP for Open-Vocabulary Video Instance Segmentation

This repo is the official implementation of CLIP-VIS: Adapting CLIP for Open-Vocabulary Video Instance Segmentation

Introduction

We present a simple encoder-decoder to adapt CLIP for open-vocabulary video instance segmentation, called CLIP-VIS. Based on frozen CLIP, our CLIP-VIS retains strong the zero-shot classification ability to various instance categories.
We design a temporal topK-enhanced matching strategy, which adaptively selects K mostly matching frames to perform query matching.
We further introduce a weighted open-vocabulary classification module, which refines mask classification by correlating mask prediction and classification.
Our CLIP-VIS achieves superior performance on multiple datasets. On validation of LV-VIS dataset, our CLIP-VIS outperforms OV2Seg by 4.6% and 11.0% AP using the backbone ResNet50 and ConNext-B. When evaluating novel categories, our CLIP-VIS outperforms OV2Seg by 10.9% and 24.0% AP, which demonstrates the good zero-shot classification ability to novel categories. When using the ResNet50 as backbone, our CLIP-VIS outperforms OpenVIS by 1.7% AP on the validation set of BURST dataset.

For further details, please check out our paper.

Installation

Please follow installation.

Data Preparation

Please follow dataset preperation.

Training

We provide shell scripts for training on image datasets and video datasets. scripts/train.sh trains the model on LVIS or COCO dataset. scripts/train_video.sh fine-tune the model on YTVIS2019 dataset.

To train or evaluate the model in different environments, modify the given shell script and config files accordingly.

Training script

sh scripts/train.sh [CONFIG] [NUM_GPUS] [BATCH_SIZE] [OUTPUT_DIR] [OPTS]
sh scripts/train_video.sh [CONFIG] [NUM_GPUS] [BATCH_SIZE] [OUTPUT_DIR] [OPTS]

# Training on LVIS dataset with ResNet50 backbone
sh scripts/train.sh configs/clipvis_R50.yaml 4 8 output/lvis MODEL.MASK_FORMER.DEC_LAYERS 7
#Training on COCO dataset with ResNet50 backbone
sh scripts/train.sh configs/clipvis_R50.yaml 4 8 output/coco MODEL.MASK_FORMER.DEC_LAYERS 10 DATASETS.TRAIN '("coco_2017_train",)' DATASETS.TEST '("coco_2017_val",)'
#Fine-tune on YTVIS2019 dataset with ResNet50 backbone
sh scripts/train_video.sh configs/clipvis_video_R50.yaml 4 8 output/ytvis MODEL.MASK_FORMER.DEC_LAYERS 10 MODEL.WEIGHTS models/coco/model_final.pth

Evaluation

We provide shell scripts scripts/eval_video.sh for Evaluation on various video datasets.

Evaluation script

sh scripts/eval_video.sh [CONFIG] [NUM_GPUS] [VAL_DATA] [TEST_NUM_CLASS] [OUTPUT_DIR] [WEIGHTS] [OPTS]

#Evaluation on validation set of LV-VIS datset
sh scripts/eval_video.sh configs/clipvis_video_R50.yaml 4 '("lvvis_val",)' 1196 output/lvvis models/clipvis_lvis_r50_7.pth MODEL.MASK_FORMER.DEC_LAYERS 7

Results

We train our network on training set of LVIS dataset and evaluate our network on multiple video datasets. We provide pretrained weights for our models reported in the paper. All of the models were evaluated with 4 NVIDIA 3090 GPUs.

Training Data	Backbone	LV-VIS val	LV-VIS test	OVIS	YTVIS19	YTVIS21	BURST	Download
LVIS	ResNet-50	18.8	13.8	12.3	30.7	28.7	5.1	ckpt
LVIS	ConvNeXt-B	32.1	25.2	18.2	42.3	39.5	8.5	ckpt

Citation

@article{zhu2024clip,
  title={CLIP-VIS: Adapting CLIP for Open-Vocabulary Video Instance Segmentation},
  author={Zhu, Wenqi and Cao, Jiale and Xie, Jin and Yang, Shuangming and Pang, Yanwei},
  journal={arXiv preprint arXiv:2403.12455},
  year={2024}
}

Acknowledgement

We would like to acknowledge the contributions of public projects, such as Mask2Former, LVVIS and fc-clip whose code has been utilized in this repository.

zwq456/CLIP-VIS