ForestBelongings

Under-canopy imagery of dropped personal belongings, annotated for object detection. Handheld capture at heights and viewpoints designed to approximate low-altitude imagery from a micro air vehicle (MAV) flying beneath the forest canopy.

Anonymous release for peer review at the NeurIPS 2026 VLM4RWD workshop. Author, institution and funding information will be added after the review period.

Overview

Images	31,897 (1920x1080: 27,663 · 1280x720: 4,234)
Sequences	250
Annotations	40,527 — 39,831 scored, 696 iscrowd=1 ignore
Categories	7 — person clothes hat gloves bag shoes etc
Splits	train 22,984 · val 4,486 · test 4,427 images
Capture period	2023-09 to 2026-07
Terrain	forest, river, valley, beach, grass
Format	COCO, one JSON per split
Size	10.2 GiB

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/. Splits are at sequence level — consecutive frames are near-duplicates — and sequences flown over the same spot are kept in one split.

Categories

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

etc is an annotated belonging outside the five belongings types. person covers people appearing alongside the belongings.

Annotation conventions

iscrowd=1 means annotated but not scored. 696 boxes carry it, and every one of them is a small box: below sqrt(area) < 16px after the usual detector resize (ResizeShortestEdge(800, max 1333)) — under 23.0px at 1920x1080, under 15.4px at 1280x720. By class: clothes 205, shoes 183, gloves 132, bag 104, hat 71, person 1. No etc box is flagged. All counts above exclude them.

etc is scored here, and ignored downstream. Nothing in this release marks etc, so its 349 boxes are ordinary annotations. The six-class benchmark built from this corpus does force them to iscrowd=1, because its vocabulary has no term for them — that is a choice made when building the benchmark, not a property of the data you are downloading.

Empty frames are kept. 3,423 images carry no scored annotation, of which 2,161 are object-free background sequences (train only, sequences 501-517). Set filter_empty=False or the equivalent.

Face anonymisation

Faces were reviewed frame by frame and mosaicked (block size 12): 20 images, 21 face boxes, out of 236 frames flagged for review. Affected frames fall in train 13 / val 6 / test 1. See ANONYMISATION.md.

Provenance

Images, boxes, category ids, category names, iscrowd flags and image references are the ones the accompanying paper's experiments ran on, verified element by element. The 20 face-anonymised frames are the exception.

Rebuilding the paper's benchmark

The paper pools this corpus with ForestPersons into a six-class benchmark. Building it from this release reproduces the published splits exactly — 90,670 train / 22,729 val / 14,980 test images, matching annotation for annotation. etc is carried through as an ignore region. See the accompanying code repository for prepare.py.

Loading

import json
from pathlib import Path

root = Path("ForestBelongings")           # after snapshot_download
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"])

STATISTICS.md breaks the corpus down by capture month, season, terrain, time of day, object size and zero-shot difficulty.

Licence

CC BY-NC-SA 4.0 — https://creativecommons.org/licenses/by-nc-sa/4.0/ Non-commercial research use; attribution required; derivatives under the same licence.

Citation

Anonymous during review.