mehmetkeremturkcan/constellation_urban_intersection_dataset

Mehmet Kerem Turkcan, Chengbo Zang, Sanjeev Narasimhan, Gyung Hyun Je, Bo Yu, Mahshid Ghasemi, Javad Ghaderi, Gil Zussman, Zoran Kostic
NSF Center for Smart Streetscapes  |  Department of Electrical Engineering (AIDL Lab)  |  COSMOS Testbed

Paper (IJCV 2026)  |  arXiv  |  Website  |  Code & Models

Abstract

As smart cities evolve, privacy-preserving edge processing at traffic intersections has become essential for real-time safety applications while reducing data transmission and centralized computation. High-altitude cameras with on-device inference provide an optimal solution that respects privacy while delivering low-latency results. We introduce Constellation, a dataset of 13K images for research on object detection in dense urban streetscapes from high-elevation cameras across varied temporal conditions. The dataset addresses challenges in small object detection, particularly for pedestrians observed from elevated positions with limited pixel footprints. Our evaluation of contemporary object detection architectures reveals a 10% lower average precision (AP) for small pedestrians compared to vehicles. Pretraining models on structurally similar datasets increases mean AP by 1.8%. Domain-specific data augmentations and pseudo-labeled data from top-performing models further enhance performance. We evaluate deployment viability on resource-constrained edge devices including Jetson Orin, Raspberry Pi 5, and mobile platforms, demonstrating feasibility of privacy-preserving on-device processing. Comparing models trained on data collected across different time intervals reveals performance drift due to changing intersection conditions. The best-performing model achieves 92.0% pedestrian AP with 7.08 ms inference time on A100 machines, and 95.4% mAP. The best-performing edge model achieves a similar performance, with Jetson Orin Nano achieving 94.5% mAP and 27.5ms inference time using TensorRT.

The same camera under different conditions. (a–d) Weather and time of day; (e–h) changes to the scene background.

Dataset Files

File	Contents	Size
constellation_dataset.zip	Main dataset: 13,314 annotated frames, 10,230 train and 3,084 test	3.3 GB
constellation_scenarios.zip	Rare scenarios such as snow and rain (day, night, fog, rain, heavy rain and snow subsets)	1.4 GB
constellation_drift.zip	Time-drift experiments: 4,731 frames from 2020 and 4,716 frames from 2023	2.5 GB
constellation_pre_transformation.zip	Pre-transformation frames: the original camera view, before the perspective transformation to a top-down view	1.9 GB
constellation_boxmask.zip	BoxMask dataset: 50,556 pseudo-labeled frames used for pretraining	15.6 GB
2019_sample.mp4, 2025_sample.mp4	Sample videos	1.3 GB

Annotations are YOLO-format bounding boxes with two classes, vehicle and pedestrian.

Usage

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="mehmetkeremturkcan/constellation_urban_intersection_dataset",
    filename="constellation_dataset.zip",
    repo_type="dataset",
)

Extract the archive and set the dataset path in configs/constellation.yaml.

Pretrained Models

Pretrained PyTorch and TensorRT models, together with training and evaluation code, are available on GitHub.

Acknowledgements

This data was collected at the PAWR COSMOS testbed at Columbia University. This work began while the first author was a member of the Department of Electrical Engineering (AIDL Lab) at Columbia University.

Citation

@article{turkcan2026constellation,
  title={Constellation dataset: Benchmarking high-altitude object detection for an urban intersection},
  author={Turkcan, Mehmet Kerem and Zang, Chengbo and Narasimhan, Sanjeev and Je, Gyung Hyun and Yu, Bo and Ghasemi, Mahshid and Ghaderi, Javad and Zussman, Gil and Kostic, Zoran},
  journal={International Journal of Computer Vision},
  volume={134},
  number={10},
  pages={429},
  year={2026},
  publisher={Springer}
}