DATA_DICTIONARY.md

# MVOT data dictionary

## Video records

Each line in `data/metadata.jsonl` describes one video. Hugging Face
`VideoFolder` exposes the same fields and loads `file_name` as the video
feature.

| Field | Type | Description |
|---|---|---|
| `file_name` | string | Path from `data/metadata.jsonl` to the MP4 file. |
| `sample_id` | string | Unique identifier for the sequence. |
| `subset` | string | `regular` or `occlusion`. |
| `lighting` | string | `day`, `dusk`, or `night`. |
| `camera_tilt` | string | `0degree`, `10degree`, or `20degree`. All dusk and night sequences use `0degree`. |
| `view` | string | MatrixCity camera-trajectory region: `bottom`, `left`, `right`, or `top`. |
| `segment_id` | integer | Numeric segment identifier. It is not globally unique. |
| `annotation_path` | string | Repository-relative path to the canonical CSV annotation. |
| `has_annotations` | boolean | Whether the sequence includes canonical frame annotations. This is `true` for every published sequence. |
| `frame_count` | integer | Number of video frames. |
| `annotation_frame_count` | integer | Number of frame annotation records. |
| `bbox_frame_count` | integer | Number of frames with all four box coordinates. |
| `usable_bbox_frame_count` | integer | Number of frames with a box and no active invalid marker. |
| `width`, `height` | integer | Video dimensions in pixels. |
| `fps` | number | Video frame rate. |
| `duration_seconds` | number | Video duration in seconds. |
| `codec` | string | Video codec. |
| `occlusion_values` | list[string] | Distinct non-null frame-level occlusion labels in the sequence. |
| `similar_values` | list[string] | Distinct non-null similar-object labels in the sequence. |
| `sha256` | string | SHA-256 digest of the MP4 file. |
| `annotations` | list[object] | Ordered frame-level annotation records. |

## Frame annotations

| Field | Type | Description |
|---|---|---|
| `frame_id` | integer | One-based frame index. |
| `bbox_xywh` | list[number] or null | Bounding box as `[left, top, width, height]` in pixels. |
| `has_bbox` | boolean | `true` when all four box coordinates are present. |
| `source_marked_invalid` | boolean | `true` when an annotation-level invalid marker or range covers the frame. |
| `invalid_note` | string or null | Original invalid marker on the row, such as `12-100` or `invalid`. |
| `occlusion_raw` | string or null | Original frame-level occlusion value: `no`, `yes`, `1`, or `2`. |
| `similar_raw` | string or null | Original similar-object value: `no` or `yes`. |

## Canonical CSV annotations

Each file under `annotations/` contains one row per frame. The CSV columns are:

```text
frame_id,bbox_left,bbox_top,bbox_width,bbox_height,has_bbox,
source_marked_invalid,invalid_note,occlusion_raw,similar_raw
```

The line break above is for readability only. Empty bounding-box cells indicate
a missing box. Boolean values are written as lowercase `true` and `false`.

## Coordinate convention

```text
(0, 0) --------------------> x / bbox_left
  |
  |       +-------------+
  |       | bounding box|
  |       +-------------+
  v
y / bbox_top
```

Bounding boxes use `[left, top, width, height]` in pixel coordinates. Audited
boxes lie within the 1024 x 1024 frame bounds.

## Value interpretation

- Missing boxes remain explicit as `null` values.
- Invalid ranges are expanded into `source_marked_invalid` while the original
  text remains available in `invalid_note`.
- `occlusion_raw` preserves the supplied labels without assigning an
  undocumented meaning to numeric values.
- `subset` is the sequence-level grouping and is independent of
  `occlusion_raw`.
- All published videos contain 100 frames and 100 ordered annotation records.

## Provenance

The associated paper describes MVOT as a synthetic dataset rendered from
MatrixCity in Unreal Engine. It uses regularly sampled camera positions, fixed
camera height, controlled scene pitch and illumination, and bounding boxes
obtained through the Unreal Engine actor-tracking API.