matclint/vip-corpus-level

VIP Corpus-Level

VIP: Finding Important People in Images (CVPR 2015)

Dataset Summary

Pairwise ground-truth comparisons of the same person's importance across two video frames.

• Comparisons: 1000
• Data file: data/comparisons.jsonl

Each JSONL row is a complete comparison. The two relative_importance values are normalized within that row and sum to 1. They are not absolute scores and should not be compared directly across unrelated comparisons.

Data Fields

Field	Description
comparison_id	Stable identifier for the frame pair.
frame_a.frame_id, frame_b.frame_id	VIP frame identifiers.
frame_a.media_path, frame_b.media_path	Relative paths expected by the local viewer.
frame_a.width, frame_a.height	Frame dimensions in pixels; the same fields exist for B.
frame_a.bbox_xywh, frame_b.bbox_xywh	Face bounding boxes as [x, y, width, height] in frame pixels.
frame_a.episode_number, frame_b.episode_number	Source episode numbers.
frame_a.frame_number, frame_b.frame_number	Frame numbers within their episodes.
frame_a.timestamp_seconds, frame_b.timestamp_seconds	Extraction timestamps when available.
frame_a.relative_importance, frame_b.relative_importance	Normalized vote fractions that sum to 1.
label	More important side: a, b, or tie.

Bounding boxes identify faces. They preserve the original release coordinates and may extend beyond an image boundary when a face is partially outside the frame. Clip boxes to the corresponding image dimensions before cropping.

Example

{
  "comparison_id": "TR001",
  "frame_a": {
    "frame_id": "l1000600",
    "media_path": "task_bbt/l1000600.jpg",
    "width": 512,
    "height": 288,
    "bbox_xywh": [
      234.0,
      57.0,
      25.0,
      25.0
    ],
    "episode_number": 1,
    "frame_number": 600,
    "relative_importance": 0.0,
    "timestamp_seconds": 24.054
  },
  "frame_b": {
    "frame_id": "l3025830",
    "media_path": "task_bbt/l3025830.jpg",
    "width": 512,
    "height": 288,
    "bbox_xywh": [
      162.0,
      68.0,
      41.0,
      41.0
    ],
    "episode_number": 3,
    "frame_number": 25830,
    "relative_importance": 1.0,
    "timestamp_seconds": 1033.25
  },
  "label": "b"
}

Media Access

Raw source media is not redistributed in this dataset release. The vip package provides helpers for preparing locally sourced media and building the viewer.

License

VIP comparison labels, relative-importance scores, face bounding boxes, timestamps, and release metadata are licensed under CC BY 4.0. Upstream source media is not covered by this license and is not redistributed with the dataset.

Tools

See mclint/vip for dataset loaders, evaluation utilities, media helpers, and local viewers.

Citation

If you use this dataset, please cite the CVPR 2015 paper:

@InProceedings{Mathialagan_2015_CVPR,
author = {Solomon Mathialagan, Clint and Gallagher, Andrew C. and Batra, Dhruv},
title = {VIP: Finding Important People in Images},
booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2015},
doi = {10.1109/CVPR.2015.7299119}
}