DroneCATS-Pilot-54K

The depth and drone-action data used to fine-tune DroneCATS-Cosmos3-Edge-2B: 53,976 single-turn question-answer rows over frames from NYU Depth v2, SUN RGB-D and TartanAir v2.

This repository holds the annotations only. No image is included. The images are rebuilt from the official source downloads with a reconstruction script, which will be released in the DroneCATS GitHub repository soon. The rebuilt files are byte-identical to the ones we trained on, and reference/SHA256SUMS lets you check them.

Tasks

Task	The question asks for	train	val
depth	Depth at given image coordinates	12,530	393
point_depth	The point and depth of a described object	25,910	732
action_point_depth	One drone action (go, rotate, ...) toward a described object	13,039	1,372
Total		51,479	2,497

The rows cover 25,727 images in 89 zips. Depth labels come from the ground-truth depth maps of the source datasets. The referring expressions and object tags were generated with OpenAI GPT models and can contain errors.

Contents

Path	What
rows/	{train,val}_{depth,point_depth,action_point_depth}.jsonl, one JSON row per line. Images are referenced as images/<zip>#<member>.
ledger/derivations.jsonl	One line per derived image: path, source, frame, ops, sha256, bytes, pixel_sha256.
ledger/zip_layout.jsonl	One line per image zip: zip, sha256, bytes, and members in archive order (name, path, date_time).
ledger/frames_*.jsonl	Where each frame is in the official download (NYU Depth v2 .mat index, SUN RGB-D zip member, TartanAir v2 environment / trajectory / frame).
ledger/archives.json	Official URL, size and SHA-256 of every source archive needed.
reference/SHA256SUMS	SHA-256 of the row files, the 89 zips and every zip member.
LICENSES/	Annotation license and source dataset notices.

Row format

Each row is {"meta": {...}, "tools": [], "messages": [...]}. messages is the chat used for training; meta describes the target:

meta field	Meaning
source	nyuv2, sunrgbd or tartanair
oid	Target id <frame>-<x>x<y>: source frame and target pixel in the original frame
desc	Referring expression of the target (used in point_depth and action_point_depth questions)
point	Target [x, y], normalized to [0, 1] over the image shown in the row. action_point_depth uses a 0–1000 grid, so a go answer's point is round(1000 * point).
gt_depth_m	Metric depth of the target in metres (median over the target region of the source depth map)
target_kind	canonical_fov: the answer is canonical depth. metric_paired: a second row for the same oid that asks for metric depth.
gt_depth_target	Depth the answer is built from: gt_depth_m / focal_ratio for canonical_fov (a reference camera with fx = width, about 53° horizontal FOV), gt_depth_m for metric_paired
focal_ratio	fx / width of the image shown in the row

meta objects inside messages (such as {"trainable": false}) are part of the training format; trainable marks the target turns.

Derivation paths

A derived image path encodes how the file was made:

images/<source>/<kind>/[<W>x<H>/<base>/]<frame>[__<cell>__<x0>_<x1>]__(fx<fx>_fy<fy>_<W>x<H> | nofx).<ext>

ops in derivations.jsonl lists the same steps, applied in order:

op	Meaning
source	Original frame bytes: SUN RGB-D JPEG and TartanAir PNG as stored in the archive; NYU Depth v2 RGB array saved by Pillow as PNG (compress_level=1)
crop_x	Decode with OpenCV (IMREAD_COLOR) and keep columns x0:x1
fit_width_jpeg	If wider than max_width, resize to that width with cv2.INTER_AREA (height round(h * s)), then encode JPEG with cv2.imencode at quality
resize_lanczos	Pillow resize((width, height), LANCZOS); PNG saved with optimize=True, JPEG saved with the source quantization tables, subsampling and progressive flag
camera_exif	Add EXIF tags for fx, fy (pixels) of a width x height image. PNG is re-saved by Pillow with an eXIf chunk; JPEG gets an APP1 segment right after SOI, without re-encoding.

Files marked nofx carry no added EXIF. Several paths can have identical bytes; the ledger lists the one used for each zip member.

License

The annotations (rows, ledger, reference hashes) are licensed under CC BY-NC 4.0, for non-commercial research only. As parts were generated with OpenAI GPT models, users must also comply with OpenAI's Terms of Use. No image is redistributed: rebuilt images stay under the terms of NYU Depth v2, SUN RGB-D and TartanAir v2 (download links and required citations in LICENSES/THIRD_PARTY_DATA.md).

Citation

@article{park2026dronecats,
  title   = {Evaluating Multimodal LLMs as Generalist Vision-Language-Action Agents
             for Drone Control: Commanding, Approaching, Tracking and Searching},
  author  = {Park, Jaewoo and Lee, Minyoung and Seo, Sukmin and Yim, Moonbin and
             Yoon, Hyunwook and Ryu, Dohoon and Kim, Daehee and Song, Myungseo and
             Byun, Jihyuk and Chang, Seunggyu and Kil, Taeho and Kim, Jiseob and
             Lee, Bado and Kim, Geewook},
  journal = {arXiv preprint arXiv:2609.01404},
  year    = {2026}
}

Please also cite the source datasets listed in LICENSES/THIRD_PARTY_DATA.md.