---
pretty_name: SAVAM — Semiautomatic Visual-Attention Modeling
configs:
- config_name: clips
  default: true
  data_files:
  - split: train
    path: viewer/clips/*.parquet
- config_name: filtered_gaze
  data_files:
  - split: train
    path: viewer/filtered_gaze/*.parquet
- config_name: raw_gaze
  data_files:
  - split: train
    path: viewer/raw_gaze/*.parquet
tags:
- saliency
- video
- fixations
- gaze
size_categories:
- n<1K
---

# SAVAM — Semiautomatic Visual-Attention Modeling

**Official dataset page:** [videoprocessing.ai/datasets/savam.html](https://videoprocessing.ai/datasets/savam.html)

SAVAM contains human eye-movement recordings collected while viewing videos, including static and dynamic scenes, film excerpts, and sequences from research video databases.

## Official dataset description

- 41 video fragments from feature films, commercials, and stereoscopic video databases.
- Approximately 13 minutes of video, ~20000 frames.
- 50 observers, predominantly aged 18–27.
- Full HD and 4K UHDTV stereoscopic video sequences.
- Eye tracking with an SMI iViewX Hi-Speed 1250 at 500 Hz.
- Additional postprocessing to improve recording accuracy.

## Gaze data

The descriptions below follow the original [`GazeData/README.txt`](GazeData/README.txt).

| Directory | Contents |
| --- | --- |
| `GazeData/raw_gaze_data/` | Original eye-tracking device output |
| `GazeData/filtered_gaze_data/` | Filtered eye-tracking data |
| `GazeData/gaussian_vizualizations/` | Gaze-location distributions visualized with multiple Gaussians |

The original processing procedure described in the [Data post-processing section](https://videoprocessing.ai/datasets/savam.html#data-post-processing) of the official page.

### Metadata

- [`list_video.txt`](GazeData/list_video.txt): video name, source sequence, and starting frame in that sequence.
- [`list_user.txt`](GazeData/list_user.txt): observer identifier, sex (`m` or `f`), age, viewing number, and presentation order (`bwd` or `fwd`).
- Gaze filenames contain the video name, source sequence, starting frame, observer identifier, sex, age, trial number, and presentation order.

Some observers participated more than once. For `bwd` recordings, the order of clips was reversed, not the order of frames within a clip. The original documentation states that these data do not need to be reversed.

### Gaze coordinates

The original format documentation lists the following fields:

| Position | Field |
| --- | --- |
| 1 | Timestamp; 1,000,000 units correspond to one second |
| 2 | Left X coordinate, documented range 0–1920 |
| 3 | Left Y coordinate, documented range 0–1080 |
| 4 | Right X coordinate, documented range 0–1920 |
| 5 | Right Y coordinate, documented range 0–1080 |

If both X and Y are zero, the gaze position is unknown. This usually means the observer's eyes were closed.

## License files

[`GazeData/LICENSE.txt`](GazeData/LICENSE.txt) specifies Creative Commons Attribution 4.0 International for gaze data and requests citation of the SAVAM paper.

The original video notices are available here:

- Source videos: [VQEG](VideoSources/VQEG_sources/LICENSE.txt), [LIVE](VideoSources/LIVE_sources/LICENSE.txt), [film and commercial excerpts](VideoSources/COPYRIGHTED_sources/LICENSE.txt).
- Gaze-dot visualizations: [VQEG](DotsVisualisation/VQEG_sources_with_dots/LICENSE.txt), [LIVE](DotsVisualisation/LIVE_sources_with_dots/LICENSE.txt), [film and commercial excerpts](DotsVisualisation/COPYRIGHTED_sources_with_dots/LICENSE.txt).

## Our related saliency datasets and papers

| Dataset | Description | Related paper |
| --- | --- | --- |
| [AudioVisualMouseSaliency (AViMoS)](https://huggingface.co/datasets/ANDRYHA/AudioVisualMouseSaliency) | 1,500 Full HD videos with audio and crowdsourced mouse-tracking saliency annotations, used for the AIM 2024 challenge. | Andrey Moskalenko et al. (2024). [AIM 2024 Challenge on Video Saliency Prediction: Methods and Results](https://arxiv.org/abs/2409.14827). |
| [VideoSaliencyChallenge](https://huggingface.co/datasets/ANDRYHA/VideoSaliencyChallenge) | 2,000 Full HD videos with audio and mouse-tracking saliency annotations, used for the NTIRE 2026 challenge. | Andrey Moskalenko et al. (2026). [NTIRE 2026 Challenge on Video Saliency Prediction: Methods and Results](https://arxiv.org/abs/2604.14816). |
| [OpenSAL360](https://huggingface.co/datasets/ANDRYHA/OpenSAL360) | 500 omnidirectional videos with audio and crowdsourced saliency annotations from more than 2,000 observers. | Alexey Bryncev et al. (2026). [OpenSAL360: Open-Source Crowdsourcing Platform for Omnidirectional Video Saliency Collection](https://arxiv.org/abs/2609.21480). |

## Citation

```
 @INPROCEEDINGS {
    Gitm1410:Semiautomatic,
    AUTHOR    = "Yury Gitman and Mikhail Erofeev and Dmitriy Vatolin
                 and Andrey Bolshakov and Alexey Fedorov",
    TITLE     = "Semiautomatic {Visual-Attention} Modeling and Its 
                 Application to Video Compression",
    BOOKTITLE = "2014 IEEE International Conference on Image Processing
                 (ICIP) (ICIP 2014)",
    ADDRESS   = "Paris, France",
    PAGES     = "1105-1109",
    DAYS      =  27,
    MONTH     =  oct,
    YEAR      =  2014,
    KEYWORDS  = "Saliency;Visual attention;Eye-tracking;Saliencyaware 
                 compression;H.264",
  }
```

[Accepted manuscript (PDF)](https://compression.ru/video/savam/pdf/Semiautomatic_visual_attention_modeling_and_its_application_to_video_compression.pdf) · [Published paper (IEEE)](https://ieeexplore.ieee.org/document/7025220)