bethgelab/video_saliency

SalTempto — Saliency Eye-Tracking Dataset

A video eye-tracking dataset of 224 naturalistic video stimuli with gaze recordings from multiple subjects, designed for training and evaluating visual saliency models.

Dataset Overview

Videos	224 (1920×1080, ~30 fps)
Train / Val / Test	204 / 10 / 10
Subjects per video	Train: 1–3 (mean ≈ 2.5) · Val: 15–16 · Test: held out

Gaze recordings for the test split are intentionally held out as a hidden benchmark — only the test video stimuli are public.

Repository Structure

video_saliency/
├── dataloader.py          # PyTorch dataset and dataloader
├── data/
│   ├── train_videos.csv       # List of training video IDs
│   ├── val_videos.csv         # List of validation video IDs
│   ├── test_videos.csv        # List of test video IDs
│   ├── videos/                # MP4 stimulus videos (video_0000.mp4 … video_0223.mp4)
│   └── eyetracking/           # Per-video gaze recordings
│       └── video_XXXX/
│           └── video_XXXX_subject_Y.csv

Eyetracking Data Format

Each video_XXXX_subject_Y.csv file contains raw gaze samples at ~500 Hz with the following columns:

Column	Description
time	Timestamp in milliseconds
left_x, left_y	Left eye gaze position in pixels (1920×1080)
left_p	Left eye pupil size
right_x, right_y	Right eye gaze position in pixels
right_p	Right eye pupil size
seconds	Tracker timestamp in seconds (= time / 1000)
timestamp_start	Seconds elapsed since video onset (first row of each CSV is 0.0)
frames	Corresponding video frame index (fractional)

Missing gaze samples (e.g. blinks, lost tracking) are encoded as empty cells in the CSV (≈3.6% of left-eye and ≈4.7% of right-eye coordinates across the dataset). Out-of-screen fixations are not separately marked but can be identified by coordinates outside the valid range (x ≤ 0 or x ≥ 1920, y ≤ 0 or y ≥ 1080). The dataloader filters both cases automatically.

Installation

pip install torch decord pandas numpy

Or with uv:

uv add torch decord pandas numpy

Usage

from dataloader import EyeTrackingDataset

# Instantiate — paths are auto-derived from data_dir
dataset = EyeTrackingDataset(
    data_dir="data",
    split="train",   # "train", "val", or "test"
    use_gpu=False,
)

# Load an entire video (returns all frames)
video, fixations = dataset[0]
# video:    Tensor (num_frames, 3, H, W), float32 in [0, 1]
# fixations: Tensor (num_frames, max_fixations, 2), padded with -1
#   max_fixations = max gaze samples in any frame (≈500 Hz / 30 fps × n_subjects)

# Load a frame range
video, fixations = dataset[(0, 10, 40)]  # video 0, frames 10–40

# Negative indexing: last 50 frames
video, fixations = dataset[(0, -50, None)]

DataLoader with batching

from torch.utils.data import DataLoader
from dataloader import EyeTrackingDataset, EyeTrackingCollator

class ClipDataset(EyeTrackingDataset):
    def __getitem__(self, idx):
        return super().__getitem__((idx, 0, 30))  # 30-frame clips

dataset = ClipDataset(data_dir="data", split="train", use_gpu=False)
loader = DataLoader(dataset, batch_size=4, collate_fn=EyeTrackingCollator())

for videos, fixations, lengths in loader:
    # videos:    (B, max_frames, 3, H, W)
    # fixations: (B, max_frames, max_fixations, 2)
    # lengths:   (B,) actual frame counts per clip
    break

Parameters

Parameter	Default	Description
data_dir	required	Path to the data/ directory
split	"train"	One of "train", "val", "test"
use_gpu	True	GPU-accelerated video decoding via Decord
gpu_id	0	GPU device index
subject_mode	"combined"	"combined" for all subjects, or "subject_N" for a single subject
cumulative_fixations	False	If True, return all fixations from frame 0 up to end_frame
video_dir	data_dir/videos	Override video directory
csv_base_dir	data_dir/eyetracking	Override eyetracking directory

Croissant metadata

The dataset ships with a Croissant metadata file at croissant.json. To iterate over per-sample gaze records via mlcroissant:

from mlcroissant import Dataset

ds = Dataset(jsonld="https://huggingface.co/datasets/anonymous-neurips-submission/video_saliency/resolve/main/croissant.json")
for record in ds.records("eyetracking-data"):
    # keys: video_id, subject_id, time, left_x, left_y, left_p, right_x, right_y, right_p, seconds, timestamp_start, frames
    ...

Other record sets exposed by the metadata: train-split, val-split, test-split (each yields the list of video_ids for that split).