MST: Adaptive Multi-Scale Tokens Guided Interactive Segmentation

Interactive segmentation has gained significant attention due to its applications in human-computer interaction and data annotation. To address the challenge of target scale variations in interactive segmentation, we propose a novel multi-scale token fusion algorithm. This algorithm selectively fuses only the most important tokens, enabling the model to better capture multi-scale characteristics in important regions. To further enhance the robustness of multi-scale token selection, we introduce a token learning algorithm based on contrastive loss. This algorithm fully utilizes the discriminative information between target and background multi-scale tokens, effectively improving the quality of selected tokens. Extensive benchmark testing demonstrates the effectiveness of our approach in addressing multi-scale issues.

Environment

Training and evaluation environment: Python 3.9+, PyTorch > 1.0, CUDA. Run the following command to install required packages.

pip install -r requirements.txt

You need to configue the paths to the datasets in config.yml before training or testing. If you have any questions, please feel free to open an Issue.

Dataset

A script download_datasets.sh is prepared to download and organize required datasets.

Dataset	Description	Download Link
MS COCO	118k images with 1.2M instances (train)	official site
LVIS v1.0	100k images with 1.2M instances (total)	official site
COCO+LVIS*	99k images with 1.5M instances (train)	original LVIS images + combined annotations
SBD	8498 images with 20172 instances for (train) 2857 images with 6671 instances for (test)	official site
Grab Cut	50 images with one object each (test)	GrabCut.zip (11 MB)
Berkeley	96 images with 100 instances (test)	Berkeley.zip (7 MB)
DAVIS	345 images with one object each (test)	DAVIS.zip (43 MB)
Pascal VOC	1449 images with 3417 instances (test)	official site
COCO_MVal	800 images with 800 instances (test)	COCO_MVal.zip (127 MB)
Multi-Scale LoveDA	1666 images with 31,908 instances (test)	LoveDAMS.zip (1.4 GB)

Demo

Online Demo.

An example script to run the demo.

python demo.py --checkpoint=weights/mst-3s-slim-448-base.pth --gpu 0

Evaluation

Before evaluation, please download the datasets and models, and then configure the path in config.yml.

Download our model, please download below zipped files and extract before use:

• Model Zoo

Use the following code to evaluate the MST-3s model.

python evaluate_model.py NoBRS\ 
--gpu=0\
--checkpoint=weights/mst-3s-slim-448-base.pth\
--datasets=DAVIS,LoveDASMALL,LoveDAMEDIUM,LoveDALARGE,LoveDAHUGE\
--cf-n=3\
--acf\
--n-clicks=20 \
--target-iou=0.9\
--inference_size=448

Training

Before training, please download the MAE pretrained weights (click to download: ViT-Base, ViT-Large, ViT-Huge) and configure the dowloaded path in config.yml

Please also download the pretrained SimpleClick models from here.

Use the following code to train a huge model on C+L:

python dis_train.py models/adaptivevit_base448_cclvs.py\ 
    --batch-size 60\
    --model_base_name cclvs_adaptivevit_base448\
    --exp-name 3s_mst\
    --gpus 0,1,2,3,4,5\
    --dport 29500\
    --workers 2\
    --img_size 448\
    --amp

Experiment

Res	Train Data	Arch		3s		6s
448*448	COCO-LVIS	ViT-B	+CL	github	weight	-	-
			+MST	github	weight	-	-
			+MST+CL	github	weight	-	-
224*224	COCO-LVIS	ViT-B	+UpConv	github	weight	github	weight
			+AT	github	weight	github	weight
			+IT	github	weight	github	weight
			+MST	github	weight	github	weight
			+MST+CL	github	weight	github	weight
		ViT-L	+MST	github	weight	-	-
			+MST+CL	github	weight	-	-
1024*1024	COCO-LVIS+HQ	ViT-B	+MST-3s	github	weight	-	-
Multi-Scale Datasets				Download Link

Citation

@article{xu2024mst,
  title={MST: Adaptive Multi-Scale Tokens Guided Interactive Segmentation},
  author={Xu, Long and Li, Shanghong and Chen, Yongquan and Luo, Jun and Lai, Shiwu},
  journal={arXiv preprint arXiv:2401.04403},
  year={2024}
}

Acknowledgement

Our project is developed based on RITM and SimpleClick