# MDVD-108K Annotation Format

MDVD-108K keeps the original VisDrone-style detection labels in the release package. YOLO and COCO labels are not stored by default; they can be generated from the provided scripts when needed. The conversion scripts use only the Python standard library.

## Directory Layout

Synthetic subsets keep degradation folders:

```text
train/{degradation}/input/*.jpg
train/{degradation}/gt/*.jpg
train/{degradation}/label/*.txt
val/{degradation}/input/*.jpg
test/{degradation}/input/*.jpg
```

The real subset is flattened in the public `MDVD-108K` package:

```text
real/input/00001.jpg
real/label/00001.txt
```

## Label Fields

Each label file is a comma-separated text file. One line represents one object or ignored region:

```text
bbox_left,bbox_top,bbox_width,bbox_height,score,object_category,truncation,occlusion
```

Field meanings:

| Field | Meaning |
| --- | --- |
| `bbox_left` | left pixel coordinate |
| `bbox_top` | top pixel coordinate |
| `bbox_width` | box width in pixels |
| `bbox_height` | box height in pixels |
| `score` | `1` for valid object labels; `0` marks ignored regions |
| `object_category` | VisDrone category id |
| `truncation` | truncation flag/level from VisDrone |
| `occlusion` | occlusion flag/level from VisDrone |

Ignored regions are lines with `score <= 0` or `object_category <= 0`. They are kept in the native labels for fidelity, but are skipped by default when generating YOLO/COCO labels.

## Categories

| id | name |
| ---: | --- |
| 0 | ignored_regions |
| 1 | pedestrian |
| 2 | people |
| 3 | bicycle |
| 4 | car |
| 5 | van |
| 6 | truck |
| 7 | tricycle |
| 8 | awning_tricycle |
| 9 | bus |
| 10 | motor |
| 11 | others |

## Generate YOLO Labels

YOLO uses normalized `class x_center y_center width height` labels. Class ids are converted from VisDrone `1-11` to YOLO `0-10`.

```bash
python tools/convert_mdvd_to_yolo.py \
  --dataset-root /path/to/MDVD-108K \
  --output-dir /path/to/MDVD-108K_yolo \
  --image-mode symlink
```

The generated YOLO directory contains:

```text
MDVD-108K_yolo/
  images/{train,val,test,real}/...
  labels/{train,val,test,real}/...
  data.yaml
```

Use `--image-mode none` to generate labels only.

## Generate COCO Labels

COCO labels are generated as one JSON file per split. Bounding boxes remain in pixel-space `[x, y, width, height]`. The original `score`, `truncation`, and `occlusion` fields are preserved under each annotation's `attributes`.

```bash
python tools/convert_mdvd_to_coco.py \
  --dataset-root /path/to/MDVD-108K \
  --output-dir /path/to/MDVD-108K_coco
```

Outputs:

```text
MDVD-108K_coco/
  instances_train.json
  instances_val.json
  instances_test.json
  instances_real.json
```

By default ignored regions are skipped. Add `--include-ignored` if you want to keep them as `category_id=0` annotations with `ignore=1`.