README.md

---
license: other
license_name: academic-non-commercial
task_categories:
- visual-question-answering
- video-text-to-text
language:
- en
tags:
- video
- human-motion
- 3d-pose
- multiple-choice
- benchmark
size_categories:
- 100K<n<1M
configs:
- config_name: default
  data_files:
  - split: train
    path: train.jsonl
  - split: validation
    path: val.jsonl
  - split: test_emdb
    path: test_emdb.jsonl
  - split: test_rich
    path: test_rich.jsonl
  - split: test_egobody
    path: test_egobody.jsonl
  - split: test_3dpw
    path: test_3dpw.jsonl
  - split: test_sloper4d
    path: test_sloper4d.jsonl
---

# HumanMoveVQA

Multiple-choice question answering about human motion in video. Questions are
generated from 3D human pose estimated with
[PromptHMR](https://github.com/yufu-wang/PromptHMR) and captioned with MotionScript.

This is the **v6.3** release. It covers five source datasets: EMDB, RICH, EgoBody,
3DPW, and SLOPER4D.

## Files

| File | Rows | Source dataset |
|------|-----:|----------------|
| `train.jsonl` | 86,797 | all five |
| `val.jsonl` | 11,081 | all five |
| `test_emdb.jsonl` | 768 | EMDB |
| `test_rich.jsonl` | 2,218 | RICH |
| `test_egobody.jsonl` | 6,994 | EgoBody |
| `test_3dpw.jsonl` | 1,294 | 3DPW |
| `test_sloper4d.jsonl` | 1,008 | SLOPER4D |

`val` is the held-out in-domain split used for checkpoint selection. The `test_*`
files are the evaluation benchmark.

Train, val, and test share no videos and no questions, in every source dataset.

## Schema

One row per question:

```json
{
  "id": "q1",
  "source_dataset": "EMDB",
  "source_video_id": "00_mvs_a",
  "source_video_start": "0:00",
  "source_video_end": "0:15.875",
  "video_id": "00_mvs_a",
  "data": "P0/00_mvs_a/images/",
  "category": "comparative",
  "question": "In the first quarter of the video, is the person's first move left faster or slower than their first move right?",
  "A": "left",
  "B": "same category",
  "C": "right",
  "D": null,
  "answer": "A. left"
}
```

| Field | Meaning |
|---|---|
| `id` | Row identifier, unique within a file |
| `source_dataset` | One of `EMDB`, `RICH`, `EgoBody`, `3DPW`, `SLOPER4D` |
| `source_video_id` | Full recording. Differs from `video_id` for EgoBody `_partNN` segments |
| `source_video_start` | Segment start within the full recording, `M:SS.mmm` |
| `source_video_end` | Segment end within the full recording, `M:SS.mmm` |
| `video_id` | The video the question was generated from |
| `data` | Frame directory relative to the dataset root |
| `category` | Question type, see below |
| `question` | Question stem |
| `A`, `B`, `C`, `D` | Options. `C` and `D` are `null` for two-option questions |
| `answer` | `"<letter>. <option text>"` |

`source_video_start` and `source_video_end` are read from the video file duration,
not from the caption metadata.

### Categories

| Category | Count (train + val) | What it asks |
|---|---:|---|
| `existence` | 20,891 | Whether a motion event occurred |
| `numerical` | 16,070 | How many times an event occurred |
| `comparative` | 14,021 | Which of two directions has greater magnitude |
| `dominant` | 13,977 | Which direction dominates overall |
| `temporal` | 13,939 | When an event occurred, or its speed |
| `trajectory_affordance` | 10,946 | Path shape and what the motion affords |
| `ordering` | 8,034 | The order in which events occurred |

## Changes in v6.3

- Adds 3DPW and SLOPER4D. The previous release covered EMDB, RICH, and EgoBody only.
- Replaces all train, val, and test files with the v6.3 build.
- Fixes a duplicate-question bug in the previous `test_egobody.jsonl`, which held 265
  repeated `(video_id, question)` pairs, 34 of them self-contradictory.
- Recomputes `source_video_start` and `source_video_end` for every row.

The previous release is available at the `v6.2-legacy` tag.

## Source datasets

All five are academic, non-commercial use only. Each requires its own request or
download step. The videos are **not** redistributed here; only the questions and
answers are.

| Dataset | Videos (train / val / test) | Frame rate | Downsample | Access |
|---|---|---|---|---|
| RICH | 701 / 83 / 32 | 30 fps | 8 | [rich.is.tue.mpg.de](https://rich.is.tue.mpg.de/) |
| EgoBody | 300 / 39 / 54 | 30 fps | 4 | [sanweiliti.github.io/egobody](https://sanweiliti.github.io/egobody/egobody.html) |
| EMDB | 57 / 7 / 11 | 30 fps | 2 | [emdb.ait.ethz.ch](https://emdb.ait.ethz.ch/) |
| 3DPW | 40 / 4 / 15 | 30 fps | 1 | [3DPW project page](https://virtualhumans.mpi-inf.mpg.de/3DPW/) |
| SLOPER4D | 13 / 2 / 15 | 30 fps | 1 | [SLOPER4D project page](https://github.com/climbingdaily/SLOPER4D) |

Frame rate is the rate at which image frames are assembled into video. Downsample is
the factor applied to the source image width and height when rendering the video.

Split properties:

- EgoBody and SLOPER4D splits are group-disjoint. EgoBody groups by recording, so
  views of the same recording never straddle a split. SLOPER4D groups by subject, and
  the benchmark holds three distinct subjects.
- SLOPER4D clips pass a horizontal alignment gate (XZ RMS <= 2 m against LiDAR GT).
  Clips failing the gate, and RGB-only clips with no GT, are excluded.
- 3DPW is 30 fps while the other three original datasets are 24 fps, so 3DPW results
  are reported separately in the paper.

## Video sources and path mapping

The `data` field points at the frame directory, not the video file. The layout
differs per dataset.

### EMDB

**Request access:** [https://emdb.ait.ethz.ch/](https://emdb.ait.ethz.ch/)
Fill out the form with an institutional email address. Access is granted after approval.

**Download:** Once approved, download all partition folders `P0`, `P1`, ..., `P9`.

```
"data": "P0/07_outdoor_push_ups/images/"
  ->  <EMDB_ROOT>/P0/07_outdoor_push_ups/images/
```

> **Partition nesting:** Some archives unzip into a doubled folder, so the frames
> land at `<EMDB_ROOT>/P2/P2/<sequence>/images/`. Check your local layout and adjust
> the prefix if needed.

### RICH

**Request access:** [https://rich.is.tue.mpg.de/](https://rich.is.tue.mpg.de/)

**Download:** From the download page, download the JPEG image files for all three
splits: `train`, `val`, and `test`. The `data` field indicates which split a sequence
belongs to, so only the relevant split files are strictly required.

```
"data": "val/Pavallion_003_yoga1/cam_10/"
  ->  <RICH_ROOT>/val/Pavallion_003_yoga1/cam_10/
```

> **Train prefix:** Some copies of the RICH train split are nested one level deeper,
> so the frames land at `<RICH_ROOT>/train2/train/<scene>/<camera>/`. The `data` field
> reflects the layout the frames were read from. Adjust the prefix to match your copy.

### EgoBody

**Request access & download:** [https://sanweiliti.github.io/egobody/egobody.html](https://sanweiliti.github.io/egobody/egobody.html)
Sign the license and download the `kinect_color` folder, which holds per-recording
image frames organised by camera (`master`, `sub_1`, `sub_2`).

```
"data": "recording_20211004_S19_S06_01/sub_1/"
  ->  <EGOBODY_ROOT>/kinect_color/recording_20211004_S19_S06_01/sub_1/
```

> **Split videos:** Some entries have a `video_id` with a `_partNN` suffix, for
> example `recording_20211004_S19_S06_01_sub1_part02`. This means the questions were
> generated on a temporal segment of the full recording. `source_video_id` gives the
> full recording, and `source_video_start` / `source_video_end` give the time range
> within it. All image frames are still found under the same `data` path.

### 3DPW

**Download:** Download the image sequence archive (`imageFiles.zip`) and extract it.

```
"data": "courtyard_backpack_00/"
  ->  <3DPW_ROOT>/imageFiles/courtyard_backpack_00/
```

> **Clip videos:** Some entries have a `video_id` with a `_clipN` suffix, for example
> `downtown_bus_00_clip0`. The frames live under the unsuffixed sequence directory,
> so `data` is `downtown_bus_00/`.

### SLOPER4D

**Download:** Download the SLOPER4D dataset and extract the RGB clips.

```
"data": "seq002_football_001/seq002_football_001_c000/"
  ->  <SLOPER4D_ROOT>/clips/seq002_football_001/seq002_football_001_c000/
```

## Loading

```python
from datasets import load_dataset

ds = load_dataset("humanmovevqa/HumanMoveVQA", "default")
print(ds["test_emdb"][0])
```

## Citation

If you use this dataset, please cite the HumanMoveVQA paper and the source datasets
you use. See the paper for the full reference list.

## Generation pipeline

The code that builds this dataset is at
[github.com/humanmovevqa/HumanMove_VQA](https://github.com/humanmovevqa/HumanMove_VQA).