Awesome

RClicks: Realistic Click Simulation for Benchmarking Interactive Segmentation

This repository provides the code to estimate performance and robustness of click-based interactive segmentation methods w.r.t. clicks positions.

Setting up an environment

Install dependencies:

# create conda env
conda create -n rclicks python==3.9.19
conda activate rclicks
# install python packages
pip install -r requirements.txt

# install mmsegmentation and transalnet
mim install mmcls==0.25.0 mmcv-full==1.7.2 mmengine==0.10.2
pip install -e mmsegmentation
pip install -e transalnet

# install our packages
pip install -e isegm
pip install -e rclicks

# For benchmarking Segment-Anything (SAM), SAM-HQ, MobileSAM, SAM2/2.1
# Please install it separately e.g:
# git clone https://github.com/facebookresearch/segment-anything.git
# pip install -e segment-anything

Prepare datasets & models checkpoints

This project is mostly developed based on RITM and use the same dataset structure and evaluation scripts. Thus, you should configure the paths to the datasets in config.yml. However, currently all paths for rclicks package are hard-coded in the rclicks/rclicks/paths.py, we will change it in the next release.

Please download all datasets and place them into datasets directory.

Checkpoints for a saliency model and our clickability model can be downloaded here CLICKABILITY_CHECKPOINTS.zip (445 MB). Please unzip it right into project directory. Make sure that clickability_model.pth is located in the root of the project directory; and resnet50-0676ba61.pth and TranSalNet_Res.pth are located in transalnet\transalnet\pretrained_models.

To download interactive segmentation methods checkpoints, please refer to the repositories of the relevant papers or download all checkpoints used in this work at once — MODELS_CHECKPONTS.zip (21.5 GB)

Run optimization

Clicking Groups Sampling Example

python3 scripts/evaluate_model_ritm.py NoBRS --checkpoint=coco_lvis_h18_itermask.pth --print-ious --save-ious --datasets=GrabCut,Berkeley,DAVIS,COCO_MVal,TETRIS --n-clicks=20 --n_workers=1 --iou-analysis --thresh=0.5 --clickability_model_pth clickability_model.pth --trajectory_sampling_count=1 --trajectory_selection=-1 --trajectory_sampling_prob_low=0.9 --trajectory_sampling_prob_high=1.0

All flags the same as in original models except following additional flags:

--n_workers — number of parallel workers for evaluation (the maximum number you can fit depends on your GPU)
--clickability_model_pth — path to clickability_model checkpoint
--trajectory_sampling_count — we used one sample
--trajectory_sampling_prob_low — lower bound to slice probability mass
--trajectory_sampling_prob_high — upper bound to slice probability mass

Real-User Evaluation Example

python3 scripts/evaluate_model_ritm.py NoBRS --checkpoint=coco_lvis_h18_itermask.pth --print-ious --save-ious --datasets=GrabCut,Berkeley,DAVIS,COCO_MVal,TETRIS --n-clicks=1 --n_workers=1 --iou-analysis --thresh=0.5 --user_inputs 

All flags the same as in original models except following additional flags:

--n_workers — number of parallel workers for evaluation (the maximum number you can fit depends on your GPU)
--user_inputs — user clicks benchmarking

Full Benchmarking

Some models (SAM, SAM-HQ, MobileSAM, SAM2/2.1) should be installed using separate package.

To benchmark all models (SAM, SAM2/2.1, SAM-HQ, MobileSAM, RITM, SimpleClick, GPCIS, CDNet, CFR-ICL) after setting up an environment and downloading all checkpoints to MODEL_CHECKPOINTS folder and just run:

bash run_clicking_groups.sh 0.0 0.1
bash run_clicking_groups.sh 0.1 0.2
bash run_clicking_groups.sh 0.2 0.3
bash run_clicking_groups.sh 0.3 0.4
bash run_clicking_groups.sh 0.4 0.5
bash run_clicking_groups.sh 0.5 0.6
bash run_clicking_groups.sh 0.6 0.7
bash run_clicking_groups.sh 0.7 0.8
bash run_clicking_groups.sh 0.8 0.9
bash run_clicking_groups.sh 0.9 1.0
bash run_base_user_sample.sh

All hyperparameters are set following author implementations.

We provide all obtained evaluation_results.zip (896 MB) folders and compute_benchmark_metrics.ipynb to reproduce all benchmark metrics from the main paper and supplementary.

Clicks and clickability model

For demonstration of usage rclicks package please refer to rclicks_demo.ipynb.

Display modes ablations

To obtain display modes ablation results use the following command:

python scripts/evaluate_previews.py

Results will be printed.

PC vs mobile

To obtain comparison results between PC and mobile clicks run:

python scripts/evaluate_mobile_pc.py

Results will be printed and saved into experiments/pc_vs_mobile.csv.

Training clickability models

To train clickability model use the following command:

python scripts/train_click_model.py --sigma 5

where sigma is hyperparameter (see paper). Checkpoints will be saved into experiments/train/sigma={sigma} directory.

Evaluate clickability models

From scratch

To evaluate and ablate clickability models for each sigma from scratch you need to download cm_ablation_checkpoints.zip (1.6 GB). Please unzip all clickability_model_*.pth files into cm_ablation_checkpoints directory in the project root.

To run evaluation script to calculate all metrics from scratch per dataset and per image use the following command:

bash eval_click_models.sh $NPROC_NUMBER

Results will be saved as .csv files in experiments/eval_cm directory. In our experiments we used NPROC_NUMBER=40 and evaluated on 8 A100 GPUs.

Tables from the paper

To process precalculated .csv files with per sample metrics into tables from the paper call:

python scripts/prepare_click_models_tables.py

• experiments/eval_cm/eval_cm_tetris.csv -- evaluation of clickability models for TETRIS (Val) (Table 3 in main paper).
• experiments/eval_cm/eval_cm_all.csv -- evaluation of clickability models for all datasets (Table 6 in Appendix B.2).
• experiments/eval_cm/ablation_cm_sigma_tetris.csv -- sigma-parameter ablation of our clickability models on TETRIS (Val) (Table 7 in Appendix B.3).

Citation

Please cite the paper if you find challenge materials useful for your research:

@article{antonov2024rclicks,
  title={RClicks: Realistic Click Simulation for Benchmarking Interactive Segmentation},
  author={Antonov, Anton and Moskalenko, Andrey and Shepelev, Denis and Krapukhin, Alexander and Soshin, Konstantin and Konushin, Anton and Shakhuro, Vlad},
  journal={arXiv preprint arXiv:2410.11722},
  year={2024}
}