Dataset Card for IndoorCrowd

Dataset Summary

IndoorCrowd is a multi-scene dataset designed for indoor human detection, instance segmentation, and multi-object tracking. It captures diverse challenges such as viewpoint variation, partial occlusion, and varying crowd density across four distinct campus locations (ACS-EC, ACS-EG, IE-Central, R-Central). Faces are explicitly blurred to preserve privacy, making it suitable for safe research into intelligent crowd management and behaviour tracking.

The dataset consists of 31 videos sampled at 5 FPS, totalling 9,913 frames.

Subsets

1. Object Detection and Segmentation: 9,913 frames featuring bounding boxes and instance segmentation masks. Includes a rigorously annotated 620-frame pure-human control subset for foundation-model benchmarking.
2. Multi-Object Tracking (MOT): A 2,552-frame tracking subset providing continuous identity tracks following the MOTChallenge format.

Supported Tasks

• object-detection: Detecting human bounding boxes (Baselines benchmarked: YOLOv8n, YOLOv26n, RT-DETR-L).
• image-segmentation: Generating instance-level masks for people in crowded indoor geometries.
• video-object-tracking: Maintaining human identity across consecutive frames via tracking algorithms (Baselines benchmarked: ByteTrack, BoT-SORT, OC-SORT).

Dataset Creation

Curation Rationale

Outdoor datasets currently dominate development. Indoor environments introduce a new set of challenges like camera view obstructions (pillars, furniture), structural occlusions, near-to-distal scale variance, and abrupt density fluctuations.

Annotations

Annotations were produced using a semi-automated pipeline:

1. Auto-labelling: Uses foundation models such as SAM3, GroundingSAM, and EfficientGroundingSAM to generate initial candidate masks and tracklets.
2. Human Correction: Expert human reviewers used SAM 2.1 to manually delete false positives, append missing masks, correct identity switches, and linearly interpolate gaps, ensuring high-fidelity ground truth.

Data Splits

The dataset provides varied crowd density regimes:

• ACS-EC: A dense multi-level atrium setting with small instance scales and high occlusion ($79.3%$ dense frames).
• ACS-EG: A narrow ground-level corridor with substantial person scale variations lengthways.
• IE-Central: An intermediate seating/entrance hall environment.
• R-Central: An overhead-view atrium with prominent structural columns causing regular occlusions.

Personal and Sensitive Information

All human faces in the raw footage have been strictly blurred by an automated de-identification pipeline prior to release. No audio, demographic attributes, or personal identifiers are collected.

Additional Information

Licensing Information

The dataset is released under a license restricting its use strictly to non-commercial computer vision research. It prohibits surveillance and any re-identification of individuals.

Citation

If you use IndoorCrowd, please cite:

@article{nae2026indoorcrowd,
  title   = {IndoorCrowd: A Multi-Scene Dataset for Human Detection, Segmentation, and Tracking with an Automated Annotation Pipeline},
  author  = {Nae, Sebastian-Ion and Moldoveanu, Radu and Ghita, Alexandra Stefania and Florea, Adina Magda},
  journal = {arXiv preprint arXiv:2604.02032},
  year    = {2026},
  url     = {https://arxiv.org/abs/2604.02032}
}

Hugging Face loading

Parquet shards are organized by config (subset) and scene (split).

Hub config	Labels
obj_det_seg	SAM3 (default)
obj_det_seg_grounded_sam	Grounded SAM
obj_det_seg_efficient_grounded_sam	Efficient Grounded SAM
tracking	MOT boxes + track IDs (no masks)

Splits are scene names: acs_ec, acs_eg, ie_central, r_central (not a single train split).

Each row includes image, type, scene, objects, and metadata. For detection/segmentation, objects contains bbox, category, area, id, score, track_id, and rle_mask (COCO RLE JSON per instance).

from datasets import load_dataset
import json
from pycocotools import mask as mask_utils

ds = load_dataset("sebnae/IndoorCrowd", "obj_det_seg", split="acs_ec")
row = ds[0]
rle = json.loads(row["objects"]["rle_mask"][0])
mask = mask_utils.decode(rle)