---
license: cc-by-4.0
language:
- en
tags:
- monocular
- 3d
- wildlife
- drone
task_categories:
- object-detection
pretty_name: WildBox
---

# WildBox

A 6-class monocular 3D wildlife detection dataset with KITTI / Omni3D-compatible
labels, derived from drone footage of African megafauna. This README describes
the contents of this release and the conventions used.

Release v1.1: annotations are exported from the human-corrected annotation tracks,
and the KITTI label channel is written in a per-frame upright rectified reference
frame with standard calibration files (see `kitti_labels` section below).

Paper: [WildBox: A Dataset and Benchmark for Aerial Monocular 3D Detection of
African Savanna Wildlife](https://arxiv.org/abs/2606.21309) — NeurIPS 2026,
Evaluations and Datasets Track.

---

## Contents of this release

```
README.md                                ← this file
WildBox_train.json                       ← Omni3D-format train split
WildBox_val.json                         ← Omni3D-format val   split
metadata/                                ← per-segment / per-video / per-track statistics (CSV)
checkpoints/                             ← paper baseline checkpoints (see checkpoints/CHECKPOINTS_README.md)
<group>/                                 ← 11 group folders (see table below)
└── <video>.zip                          ← 64 per-video archives total
```

Each per-video archive expands to `<video>/seg{N}/` with frames + cameras +
KITTI labels + calibration + tracking metadata (see "Per-segment contents" below).

To use the dataset, place all of the above into a single working directory
and extract every `<video>.zip` *in place* (i.e. inside its group folder)
so that the `file_path` fields in the two JSONs resolve:

```bash
for z in */*.zip; do unzip -d "$(dirname "$z")" "$z"; done
```

Each `file_path` in the JSONs is relative to that working directory, e.g.
`gira1/DJI_20240118140338_0007_V/seg1/frame_000001.jpg`.

---

## Headline statistics

| Split | Images | Annotations | Categories |
|-------|-------:|------------:|-----------:|
| train | 45 979 | 170 554 | 6 |
| val   | 13 779 |  66 951 | 6 |

The 6 categories are: `giraffe`, `grevys_zebra`, `elephant`, `plains_zebra`,
`rhino`, `gazelle` (canonical IDs 1000 – 1005).

Each group folder contains annotations of a single species:

| Group | Species (category) |
|-------|---|
| gira1, gira2 | giraffe |
| zebr1, zebr3 | plains_zebra |
| zebr2 | grevys_zebra |
| elep1, elep2, elep3 | elephant |
| rhin1, rhin2 | rhino |
| gaze1 | gazelle |

Frames are 1920 × 1080 RGB JPEGs.

---

## Per-segment contents (inside each `<video>.zip`)

```
<video>/
└── seg{N}/
    ├── frame_NNNNNN.jpg          ← RGB 1920×1080
    ├── cameras.json              ← per-frame intrinsics K + extrinsic [R|t]
    ├── kitti_labels/
    │   └── frame_NNNNNN.txt      ← KITTI 3D-detection format (rectified frame)
    ├── calib/
    │   └── frame_NNNNNN.txt      ← per-frame KITTI calibration (P2, R0_rect)
    └── tracking_summary.json     ← annotator-native (full 3×3 R, centroid loc)
```

---

## kitti_labels/*.txt — KITTI 3D detection format (rectified frame)

One line per object, space-separated, 16 fields:

```
class trunc occ alpha x1 y1 x2 y2 h w l X Y_bot Z ry score
```

| Field         | Meaning |
|---------------|---------|
| class         | species name (e.g. `giraffe`, `plains_zebra`) |
| trunc         | truncation [0, 1] |
| occ           | occlusion {0, 1, 2, 3} |
| alpha         | observation angle, radians |
| x1 y1 x2 y2   | 2D bbox in image pixels at 1920 × 1080 |
| h w l         | cuboid extents along rectified Y / Z / X axes |
| X Y_bot Z     | **bottom-center** of cuboid in RECTIFIED camera coords (Y points down) |
| ry            | yaw rotation around the rectified Y axis (radians) |
| score         | 1.0 for ground-truth |

**The 3D fields live in a per-frame upright rectified camera frame**, exactly as
in original KITTI (where labels live in the `R0_rect` frame). The rectified
frame is chosen per frame so that its −Y axis aligns with the animals' up
direction; standing animals are therefore yaw-only and the single `ry` angle
represents the full annotated orientation without loss. To project a 3D point
`X_rect` to pixels, use the per-frame calibration:

```
x_img ~ P2 · [X_rect; 1]          (P2 from calib/frame_NNNNNN.txt)
```

Standard KITTI tooling that projects boxes through `P2` renders these labels
correctly. Do NOT project rectified coordinates with the raw camera K from
`cameras.json` — use `P2`, or rotate back to the camera frame first
(`X_cam = R_cam_from_rect · X_rect`, where `P2 = [K · R_cam_from_rect | 0]`).

Dimensions order here is KITTI-standard `(h, w, l)`; the Omni3D JSONs and
`tracking_summary.json` use the annotator-native order (see below).

---

## Coordinate conventions

* Image coordinates: 1920 × 1080, origin top-left, X right, Y down.
* Camera frame: X right, Y down, Z forward (OpenCV / Omni3D / KITTI).
* Per-segment scale normalisation: every 3D coordinate is divided by the
  median |z_cam| of valid GT cuboids in that segment, so median |z_cam| ≈ 1
  per segment. Per-segment scale factors are listed in
  `WildBox_{train,val}.json` under `info.scene_scales`. The KITTI labels and
  the Omni3D JSONs share this same per-segment gauge.

---

## kitti_labels vs. Omni3D JSON / tracking_summary.json

| Property         | kitti_labels                        | tracking_summary.json / Omni3D JSON |
|------------------|-------------------------------------|-------------------------------------|
| Reference frame  | per-frame upright rectified (calib) | camera frame                        |
| Location         | bottom-center of cuboid             | centroid of cuboid                  |
| Dimensions order | (h, w, l) — KITTI standard          | (l, w, h) — annotator native        |
| Rotation         | yaw `ry` in the rectified frame     | full 3 × 3 matrix                   |

Both channels are exported from the same human-corrected annotation tracks and
agree exactly; use whichever format your pipeline expects.

---

## cameras.json (per segment)

```json
{
  "cameras": [
    {
      "image_name": "frame_000001.jpg",
      "image_width": 1920,
      "image_height": 1080,
      "intrinsic": [[fx, 0, cx], [0, fy, cy], [0, 0, 1]],
      "extrinsic": [[r11, r12, r13, tx], ...]    // 3 × 4
    },
    ...
  ]
}
```

Intrinsics K are calibrated for 1920 × 1080. Extrinsics map world → camera
coords (`X_cam = R · X_world + t`), with the segment's first frame as the
world origin.

---

## Omni3D JSONs (`WildBox_{train,val}.json`)

Top-level keys: `info`, `categories`, `images`, `annotations` — standard
Omni3D / COCO layout.

`images[i]` keys:
* `id` (int)
* `dataset_id` (int)
* `file_path` (str, relative to this directory)
* `height`, `width` (int — 1920 × 1080)
* `K` (3 × 3, intrinsic at native resolution)

`annotations[i]` follows the Omni3D 3D-cuboid schema (centroid location,
full 3 × 3 rotation `R_cam`, full-precision dimensions `[W, H, L]`,
tight 2D box `bbox` in XYWH), plus `track_id` for per-animal identity
within a segment. `info.scene_scales` holds the per-segment scale factors
and `info.tracking_provenance` records the annotation source per segment.

---

## metadata/ — summary statistics (CSV)

Derived purely from the two Omni3D JSONs (no flight telemetry). All three
files share the keys `split,group,video[,segment]`.

* `metadata/segments.csv` — one row per segment (345):
  `species, n_frames, n_annotations, n_tracks, herd_density`.
* `metadata/videos.csv` — one row per video (64): same columns aggregated
  over the video's segments.
* `metadata/tracks.csv` — one row per track (1,478):
  `track_id, species, n_annotations, first_frame, last_frame`
  (frame numbers as in the released `frame_NNNNNN.jpg` names).

`herd_density` is the mean number of annotated animals per frame
(`n_annotations / n_frames`). A track is one animal instance followed
through one segment; `track_id` is numbered within a segment, so tracks
are keyed by `(video, segment, track_id)`.

---

## Source / provenance

Each segment was tracked and 3D-annotated via the WildLIFT-assisted annotator
pipeline and human-corrected in the annotation tool. Release v1.1 exports the
human-corrected tracks for both label channels. Intrinsics and 2D boxes are at
native frame resolution (1920 × 1080); the KITTI channel stores 3D in the
per-frame rectified frame described above, with calibration shipped per frame.

---

## Citation

```bibtex
@inproceedings{shukla2026wildbox,
  title     = {{WildBox}: A Dataset and Benchmark for Aerial Monocular {3D} Detection of {African} Savanna Wildlife},
  author    = {Shukla, Vandita and Meier, Kilian and Laporte-Devylder, Lucie and Rondeau Saint-Jean, Camille and Kline, Jenna M. and Costelloe, Blair R. and Tuia, Devis and Remondino, Fabio and Risse, Benjamin},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026},
  url       = {https://openreview.net/forum?id=AXawAJRwDi},
  eprint    = {2606.21309},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}
```