README.md

metadata

license: cc-by-4.0
task_categories:
  - other
size_categories:
  - 1K<n<10K
tags:
  - video-saliency-prediction
  - ntire-2026
  - computer-vision

CVPR-NTIRE Video Saliency Prediction Challenge 2026

This repository contains the dataset for the NTIRE 2026 Challenge on Video Saliency Prediction, as presented in the paper NTIRE 2026 Challenge on Video Saliency Prediction: Methods and Results.

Paper | GitHub | Project Page |

Dataset

We provide a novel audio-visual mouse saliency dataset with the following key-features:

• Diverse content: movie, sports, live, vertical videos, etc.;
• Large scale: 2000 videos with mean 18s duration;
• High resolution: all streams are FullHD;
• Audio track saved and played to observers;
• Mouse fixations from >5000 observers (>70 per video);
• License: CC-BY;

File structure:

1. Videos.zip — 2000 (1200 Train + 800 Test) .mp4 video (kindly reminder: videos contain an audio stream and users watched the video with the sound turned ON!)

2. TrainTestSplit.json — in this JSON we provide Train/Public Test/Private Test split of all videos

3. SaliencyTrain.zip/SaliencyTest.zip — almost losslessly (crf 0, 10bit, min-max normalized) compressed continuous saliency maps videos for Train/Test subset

4. FixationsTrain.zip/FixationsTest.zip — contains the following files for Train/Test subset:

• .../video_name/fixations.json — per-frame fixations coordinates, from which saliency maps were obtained, this JSON will be used for metrics calculation

• .../video_name/fixations_maps/ — binary fixation maps in '.png' format (since some fixations could share the same pixel, this is a lossy representation and is NOT used either in calculating metrics or generating Gaussians, however, we provide them for visualization and frames count checks)

5. VideoInfo.json — meta information about each video (e.g. license)

6. SampleSubmission.zip — example submission for the challenge, obtained from fitted Center Prior Gaussian over mean training saliency maps.

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

Archives with videos were accepted from challenge participants as submissions and scored using the same pipeline as in bench.py.

Usage example:

1. Check that your predictions match the structure and names of the baseline SampleSubmission.zip submission.
2. Install pip install -r requirements.txt, conda install ffmpeg
3. Download and extract SaliencyTest.zip, FixationsTest.zip, and TrainTestSplit.json files from the dataset page.
4. Run python bench.py (found in the GitHub repo) with flags:

• --model_video_predictions ./SampleSubmission — folder with predicted saliency videos
• --model_extracted_frames ./SampleSubmission-Frames — folder to store prediction frames (should not exist at launch time), requires ~170 GB of free space
• --gt_video_predictions ./SaliencyTest/Test — folder from dataset page with gt saliency videos
• --gt_extracted_frames ./SaliencyTest-Frames — folder to store ground-truth frames (should not exist at launch time), requires ~170 GB of free space
• --gt_fixations_path ./FixationsTest/Test — folder from dataset page with gt saliency fixations
• --split_json ./TrainTestSplit.json — JSON from dataset page with names splitting
• --results_json ./results.json — path to the output results json
• --mode public_test — public_test/private_test subsets

5. The result you get will be available following results.json path.