CrowdSAL: Video Saliency Dataset and Benchmark

Dataset

CrowdSAL is the largest video saliency dataset with the following key features:

• Large scale: 5000 videos with mean 18.4s duration, 2.7M+ frames;
• Mouse fixations from >19000 observers (>75 per video);
• Audio track saved and played to observers;
• High resolution: all streams are FullHD;
• Diverse content from YouTube, Shorts, Vimeo;
• License: CC-BY;

File Structure

1. Train/Test folders — dataset splits, ids 0001-3000 are from Train, 3001-5000 from Test subset;

2. Videos — 5000 mp4 FullHD, 30 FPS videos with audio streams;

3. Saliency — 5000 mp4 almost losslessly (crf 0, 10bit, min-max normalized) compressed continuous saliency maps videos;

4. Fixations — 5000 json files with per-frame fixation coordinates, from which saliency maps were obtained;

5. metadata.jsonl — meta information about each video (e.g. license, source URL, etc.);

Benchmark Evaluation

Environment Setup

conda create -n saliency python=3.10.19
conda activate saliency
pip install numpy==2.2.6 opencv-python-headless==4.12.0.88 tqdm==4.67.1
conda install ffmpeg=4.4.2 -c conda-forge

Run Evaluation

Usage example:

1. Check that your predictions match the structure and names of the Test dataset subset;
2. Install all dependencies from Environment Setup;
3. Download and extract all CrowdSAL files from the dataset page;
4. Run python bench.py with flags:

• --model_video_predictions — folder with predicted saliency videos
• --model_extracted_frames — folder to store prediction frames (should not exist at launch time)
• --gt_video_predictions — folder from dataset page with gt saliency videos
• --gt_extracted_frames — folder to store ground-truth frames (should not exist at launch time)
• --gt_fixations_path — folder from dataset page with gt saliency fixations
• --mode — Train/Test subsets split
• --results_json — path to the output results json

5. The result you get will be available following results_json path.