Video Saliency Challenge: Audio-Visual Mouse Fixation Data for 2000 FullHD Videos

2000 FullHD videos with audio tracks and mouse fixation data from over 5000 observers form a novel audio-visual saliency dataset. The collection includes diverse content such as movies, sports, and live videos, with a mean duration of 18 seconds. This dataset was created by ANDRYHA for the CVPR-NTIRE Video Saliency Prediction Challenge 2026.

Use Cases

• Training video saliency prediction models based on mouse fixation data from over 5000 observers.
• Developing multimodal attention models based on the synchronized audio and visual streams.
• Benchmarking saliency algorithms on diverse content types like movies, sports, and vertical videos.
• Analyzing human attention patterns in high-resolution (FullHD) video content.

Strengths

• Large scale with 2000 videos, split into 1200 for training and 800 for testing.
• High-resolution video content, with all streams in FullHD.
• Includes audio tracks, enabling audio-visual saliency research.
• Mouse fixations collected from a large pool of over 5000 observers, averaging more than 70 per video.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect bias inherent to the specific video content and observer pool used.

Provenance

Source

ANDRYHA via Hugging Face, for the CVPR-NTIRE Video Saliency Prediction Challenge 2026.

Collection Method

Mouse fixations recorded from over 5000 observers while they watched videos with audio.

Time Range

The dataset page was last updated in 2026; the temporal coverage of the video content is unknown.

Freshness

Last updated 2026-03-20 13:12:33; freshness should be verified.

Geography

Region tag indicates 'us', but specific geographic coverage of the video content is unknown.