License Scope (Research vs. Commercial):
- Default Access: Clicking "Agree" grants access under CC BY-NC-SA 4.0 for Non-Commercial Research only.
- Commercial Use: Requires a separate license. Contact deokyunKim@etri.re.kr.
Prohibited Military & Unethical Use:
- Development of Lethal Autonomous Weapons Systems (LAWS) or combat targeting.
- Military surveillance or reconnaissance for kinetic operations.
- Mass surveillance or identification of individuals without consent.
Humanitarian Exemption: Notwithstanding the above, use for peaceful humanitarian missions (e.g., Search and Rescue, disaster relief) is explicitly PERMITTED.
Termination: Any violation of these terms will result in the immediate and automatic termination of your rights to use this dataset.

Please provide your information below to request access.

ForestBelongings Dataset

Dataset Summary

ForestBelongings is an object-detection dataset of personal belongings for search-and-rescue scenarios, simulating items left behind by missing persons in forested and outdoor terrain. Clothes, hats, gloves, bags and shoes were placed on the ground and photographed in handheld sequences at heights and viewpoints designed to approximate low-altitude imagery from a micro air vehicle (MAV) flying beneath the forest canopy. People appearing alongside the belongings are annotated as well.


Images	31,897 (1920x1080: 27,663 · 1280x720: 4,234)
Sequences	250
Annotations	40,527 — 39,831 scored, 696 `iscrowd=1` (annotated, not scored)
Categories	7 — `person` `clothes` `hat` `gloves` `bag` `shoes` `etc`
Capture period	2023-09 to 2026-07
Terrain	forest, river, valley, beach, grass
Format	COCO, one JSON per split
Size	10.2 GiB

Supported Tasks

Object Detection
Open-Vocabulary / Zero-Shot Object Detection
Search and Rescue Benchmarking (belongings as evidence of a missing person)

Languages

Visual data only (category names are English)

Dataset Structure

images/<sequence>/<frame>.jpg
annotations/train.json  val.json  test.json     COCO, one per split
annotations/total.json                          all 250 sequences

file_name is <sequence>/<frame>.jpg, relative to images/.

Data Fields

image: RGB image (Electro-Optical modality)
annotations: COCO-style bounding boxes with the following attributes:
- Bounding box coordinates ([x, y, w, h])
- Category: person, clothes, hat, gloves, bag, shoes, etc
- iscrowd: 1 marks a box that is annotated but not scored (see below)

Data Splits

Split	# Sequences	# Images	# Annotations (scored)	# Annotations (`iscrowd=1`)
Train	177	22,984	27,802	436
Validation	35	4,486	6,225	174
Test	38	4,427	5,804	86

Total images: 31,897

Splits are at sequence level — consecutive frames are near-duplicates — and sequences captured over the same spot are kept in one split. The split is stratified by category, season, month, terrain, time of day, capture campaign and zero-shot difficulty. The 17 background sequences (501–517, 2,161 frames) are all in train.

Annotation Conventions

iscrowd=1 means annotated but not scored. 696 boxes carry it, all small: sqrt(area) < 16px after the usual detector resize (ResizeShortestEdge(800, max 1333)) — under 23.0px at 1920x1080, under 15.4px at 1280x720.
etc is an ordinary annotation in this release. The six-class benchmark in the accompanying paper treats etc as an ignore region because its vocabulary has no term for it; that is a benchmark choice, not a property of the data.
Empty frames are kept. 3,423 images carry no scored annotation: 2,144 frames of the background sequences 501–517 and 1,279 frames elsewhere (190 of the 3,423 hold only iscrowd=1 boxes). Use filter_empty=False or the equivalent.

Dataset Creation

Collection Process

Belongings were placed on the ground in forest, river, valley, beach and grass terrain and captured in handheld sequences approximating under-canopy MAV flight, across three campaigns (2023, 2025–26, and a 2026 summer re-shoot) plus 17 background sequences captured in 2026-07.

Personal Information

Every frame was screened for visible faces before release; 21 face regions in 20 frames were mosaicked (block size 12). No bounding box was changed.

Usage Example

Access is gated: accept the terms on this page first, then authenticate with a Hugging Face token (hf auth login).

(Recommended) Full Download — COCO Format Ready

# Clone the dataset repo (images + COCO annotations)
git lfs install
git clone https://huggingface.co/datasets/etri/ForestBelongings
cd ForestBelongings

# Resulting directory:
# ├── images/<sequence>/<frame>.jpg
# └── annotations/
#     ├── train.json
#     ├── val.json
#     ├── test.json
#     └── total.json

Or with huggingface_hub:

from huggingface_hub import snapshot_download

root = snapshot_download("etri/ForestBelongings", repo_type="dataset")

Read the Annotations

import json
from pathlib import Path

root = Path("ForestBelongings")
data = json.loads((root / "annotations" / "test.json").read_text())
images = {i["id"]: i for i in data["images"]}
names  = {c["id"]: c["name"] for c in data["categories"]}

for ann in data["annotations"]:
    if ann.get("iscrowd", 0):             # annotated, not scored
        continue
    image = images[ann["image_id"]]
    print(root / "images" / image["file_name"], names[ann["category_id"]], ann["bbox"])

Visualize One Sample

import json
from pathlib import Path
from PIL import Image
import matplotlib.pyplot as plt
import matplotlib.patches as patches

root = Path("ForestBelongings")
data = json.loads((root / "annotations" / "val.json").read_text())
names = {c["id"]: c["name"] for c in data["categories"]}
image = data["images"][0]
anns = [a for a in data["annotations"] if a["image_id"] == image["id"]]

fig, ax = plt.subplots()
ax.imshow(Image.open(root / "images" / image["file_name"]))
for a in anns:
    x, y, w, h = a["bbox"]
    ax.add_patch(patches.Rectangle((x, y), w, h, linewidth=2, edgecolor="red", facecolor="none"))
    ax.text(x, y - 5, names[a["category_id"]], fontsize=10, color="white", backgroundcolor="red")
plt.axis("off")
plt.show()

License

The ForestBelongings Dataset is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

Under this license, you may use, share, and adapt the dataset for non-commercial purposes, provided you give appropriate credit and distribute any derivatives under the same license.

For full license terms, please refer to the LICENSE file.

If you have questions regarding the dataset or its usage, please contact:

deokyunkim@etri.re.kr

Additional Terms Regarding Trained Models

Any AI models, algorithms, or systems trained, fine-tuned, or developed using the ForestBelongings Dataset are strictly limited to non-commercial use.

Disclaimer

The ForestBelongings Dataset is provided "as is" without any warranty of any kind, either express or implied, including but not limited to warranties of merchantability, fitness for a particular purpose, and non-infringement.

The authors and affiliated institutions shall not be held liable for any damages arising from the use of the dataset.

Citation Information

If you are using this dataset, please cite

@inproceedings{kim2026querycomposition,
  title     = {Query-Composition Sensitivity of Open-Vocabulary Detectors in Under-Canopy Search and Rescue},
  author    = {Deokyun Kim and Giyoung Lee and Yookyung Kim and Myungseok Ki and Jihun Cha},
  booktitle = {NeurIPS 2026 Workshop VLM4RWD},
  year      = {2026},
  url       = {https://huggingface.co/datasets/etri/ForestBelongings},
}

Deokyun Kim and Giyoung Lee contributed equally to this work.

Acknowledgments

This work was supported by the Institute of Information & communications Technology Planning Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-II220021, Development of Core Technologies for Autonomous Searching Drones)

Downloads last month: 30

Total file size:

10.9 GB

id	name	train	val	test	total
1	`person`	179	47	28	254
2	`clothes`	15,197	3,472	3,228	21,897
3	`hat`	2,685	603	600	3,888
4	`gloves`	2,193	513	498	3,204
5	`bag`	3,757	857	701	5,315
6	`shoes`	3,514	714	696	4,924
7	`etc`	277	19	53	349
	total	27,802	6,225	5,804	39,831