README.md

---
pretty_name: CountVideoBench
license: cc-by-4.0
language:
  - en
task_categories:
  - video-text-to-text
  - visual-question-answering
size_categories:
  - 1K<n<10K
tags:
  - video
  - counting
  - benchmark
  - video-language-models
  - synthetic
viewer: false
configs:
  - config_name: single_class
    default: true
    drop_labels: true
    data_files:
      - split: test
        path: "single_class/**"
---

# CountVideoBench

**A benchmark to evaluate prediction biases on the object counting task by video-language models.**

Marc Chapus, Carlos Crispim-Junior, Véronique Eglin, Atilla Baskurt
LIRIS, UMR 5205 — INSA Lyon, Université Lumière Lyon 2, CNRS, Université Claude Bernard Lyon 1

**BMVC 2026** · [Paper](https://arxiv.org/abs/2610.28356) · [Code](https://github.com/CountVideoBench/CountVideoBench) · [Project page](https://countvideobench.github.io)

## Dataset summary

CountVideoBench contains **2,600 synthetic videos** rendered with Blender, each showing a known number of identical objects. It is designed to isolate *cardinality* as the variable under study:

- **Balanced by cardinality**: 26 counts between 5 and 50, with exactly 100 videos per count. The counts are organised in clusters around six anchors (5, 10, 20, 30, 40, 50), each anchor with its neighbours at ±1 and ±2, so that accuracy can be compared across near-identical inputs:
  `5–12, 18–22, 28–32, 38–42, 48–50`.
- **Recognition kept simple**: each video contains a single object class among 10 (2 shapes × 5 colours), with exactly 10 videos per class and per count.
- **Static objects, moving camera**: objects are static and the camera rotates in place. The white background carries a regular line pattern, and object surfaces a low-amplitude procedural noise, so that densely packed objects stay visually distinct.
- **Exact ground truth**: counts come from the generation process, with no human annotation.

The benchmark is far from saturated: no model evaluated in the paper exceeds 27.6% exact-match accuracy. On a 104-video subset, human participants reach 90.8%, against 26.9% for the best models on the same videos.

## Example

![Preview of vid_001001: 20 red cubes, rotating camera](https://countvideobench.github.io/static/vid_001001.webp)

**Video**: [`single_class/20_entities/vid_001001.mp4`](single_class/20_entities/vid_001001.mp4) (preview above, downscaled) — red cubes, 150 frames.

**Question**: How many red cubes are in this video?

**Answer**: 20

The `question` field of each video holds the exact prompt used in the paper. Using it unchanged keeps results comparable with those reported there.

## Dataset structure

```
single_class/
├── metadata.csv
├── 05_entities/
│   ├── vid_000001.mp4
│   ├── ...
│   └── vid_000100.mp4
├── 06_entities/
├── ...
└── 50_entities/
```

Each folder holds the 100 videos of one count. `metadata.csv` has one row per video:

| Field       | Type   | Description                                      |
| ----------- | ------ | ------------------------------------------------ |
| `file_name` | string | Path to the video, relative to `single_class/`   |
| `video_id`  | string | Unique identifier (`vid_000001` to `vid_002600`) |
| `count`     | int    | Ground-truth number of objects                   |
| `shape`     | string | `cube` or `sphere`                               |
| `color`     | string | `red`, `green`, `blue`, `yellow` or `purple`     |
| `question`  | string | Question asked to the model                      |

Videos: MP4, 1920 × 1080, 30 fps, 5 seconds (150 frames). The full clips are provided so that users can choose their own frame sampling strategy.
They are encoded as H.264 High 4:4:4 Predictive in RGB (`pix_fmt=gbrp`), which avoids chroma subsampling. Chromium-based browsers (Chrome, Edge, Brave) display these files with a pink tint when opened directly on the Hub; this is a browser display issue, not a problem with the files. They decode correctly in Firefox, VLC and any FFmpeg-based reader (decord, PyAV, torchvision, torchcodec, OpenCV). To browse the videos, open them in Firefox or download a single count folder (see [Usage](#usage)) and play the files locally. Example clips are also shown on the [project page](https://countvideobench.github.io).

## Usage

Load the dataset in Python with the [`datasets`](https://huggingface.co/docs/datasets) library, or download the video files with the Hugging Face command-line tool [`hf`](https://huggingface.co/docs/huggingface_hub/guides/cli). The dataset is public: no account is needed.

```python
# pip install datasets
from datasets import load_dataset, Video

# Downloads the full dataset (139 GB) on first call
ds = load_dataset("countvideobench/CountVideoBench", "single_class", split="test")

# Keep file paths instead of decoded videos, to sample frames your own way
ds = ds.cast_column("video", Video(decode=False))

sample = ds[0]
print(sample["question"], sample["count"], sample["video"]["path"])
```

Without the `cast_column` line, accessing a row decodes its video, which requires [`torchcodec`](https://github.com/meta-pytorch/torchcodec) and FFmpeg.

To download the raw files instead:

```bash
# pip install -U huggingface_hub   (provides the hf command)

# All 2,600 videos
hf download countvideobench/CountVideoBench --repo-type dataset --local-dir CountVideoBench

# Or one count only, e.g. the 100 videos with 20 objects
hf download countvideobench/CountVideoBench --repo-type dataset --local-dir CountVideoBench \
  --include "single_class/metadata.csv" "single_class/20_entities/*"
```

## Intended uses

- Evaluating object counting in video-language models.
- Diagnosing prediction biases per cardinality (preferred values, peaks and collapses in accuracy).
- Comparing models under identical, controlled conditions.

CountVideoBench is an evaluation benchmark: please do not use it as training data, so that reported results remain comparable.

## Generation

Videos were rendered procedurally with Blender. All generation parameters (rendering, camera, objects, placement, physics, lighting) are documented in Appendix B of the paper.

## Citation

If you use CountVideoBench, please cite:

```bibtex
@inproceedings{chapus2026countvideobench,
  author    = {Chapus, Marc and Crispim-Junior, Carlos and Eglin, V{\'e}ronique and Baskurt, Atilla},
  title     = {{CountVideoBench}: A Benchmark to Evaluate Prediction Biases on the Object Counting task by Video-Language Models},
  booktitle = {37th British Machine Vision Conference 2026, {BMVC} 2026, Lancaster, UK, November 23-26, 2026},
  publisher = {BMVA},
  year      = {2026}
}
```

## Contact

Marc Chapus — marc.chapus@liris.cnrs.fr