Projects visual datasets

Constellation

A 13,000-image benchmark for detecting pedestrians and vehicles from a camera mounted high above a Manhattan intersection, with detector, domain transfer, drift, and edge deployment evaluations.

Role: Lead author and dataset lead
Published: International Journal of Computer Vision, 2026
Access: Open-access paper, public dataset, and pretrained models

A Constellation test frame with its ground-truth labels and a magnified crosswalk, where pedestrians cover only a few pixels. Right, the same camera at dawn, in daytime, and at night, and across changes to the street surface between 2019 and 2023.

Constellation was recorded by a single camera mounted high above a busy intersection in Manhattan. Its 13,314 annotated frames span 28 time intervals between 2019 and 2023, covering dawn, daylight, rain, fog, and night, as well as physical changes to the street itself: faded markings, an unpaved surface, and repaving. The training and test sets are separated in time, so no interval appears in both.

The camera is deliberately placed high enough that faces and license plates cannot be resolved. The same privacy-preserving vantage point makes pedestrians very small, which puts the benchmark squarely in the small-object regime where contemporary detectors are weakest.

Findings

Small pedestrians are the bottleneck. Vehicle AP is close to saturation for most architectures, while pedestrian AP varies by more than 25 points between them.
Off-the-shelf models transfer poorly. The strongest zero-shot open-vocabulary detector, SAM3, reaches 64.6 pedestrian AP. A YOLOv8x trained on VisDrone reaches 27.3, compared with 87.4 when trained on Constellation.
Domain-aware training closes the gap. Scene-specific augmentations such as synthetic shadows, combined with VisDrone pretraining, raise YOLOv8x to 92.0 pedestrian AP and 95.4 mAP@0.5.
Intersections drift over time. A model trained only on 2020 data loses 7.1 mAP@0.5 on 2023 footage, and heavy snow (73.7) and heavy rain (82.7) remain the hardest conditions.

Model	Training	Pedestrian	Vehicle	Mean
SAM3	Zero-shot	64.6	97.8	81.2
YOLOv8x	VisDrone only	27.3	88.6	57.9
YOLOv8x	Constellation	87.4	98.6	93.0
YOLOv8x	+ domain-specific augmentations	90.7	98.6	94.7
YOLOv8x	+ VisDrone pretraining	92.0	98.8	95.4

AP@0.5 on the Constellation test set, as reported in the IJCV paper.

Edge deployment

The camera is meant to process video on site, so the paper also measures latency on embedded and mobile hardware. On a Jetson Orin Nano with TensorRT, a VisDrone-pretrained YOLOv8n reaches 94.5 mAP@0.5 at 27.5 ms per frame, within a point of the best server model.

Platform	Runtime	YOLOv8n	YOLOv8x
NVIDIA A100	TensorRT	3.4 ms	7.1 ms
Jetson Orin Nano	TensorRT	27.5 ms	111.6 ms
Rubik Pi 3	LiteRT	55.1 ms	819.0 ms
Raspberry Pi 5	NCNN	353.2 ms	2,325.7 ms
iPhone 13 Pro Max	WebAssembly	372.6 ms	10,233.1 ms

Mean inference latency per frame over 100 Constellation images.

Release

The dataset is available in YOLO format on Hugging Face under CC BY-NC-SA 3.0, together with the training and evaluation scripts and a model zoo of seven pretrained detectors, including the 95.4 mAP YOLOv8x and the edge-ready YOLOv8n. The benchmark toolkit reproduces the latency measurements across GPU, embedded, and mobile runtimes.

Project details

Artifacts: Open-access journal paper, public dataset in YOLO format, training and evaluation code, seven pretrained models, cross-dataset evaluations, and hardware latency benchmarks
Keywords: High-altitude vision

Small object detection

Urban datasets

Model drift

Domain transfer

Open-vocabulary detection

Edge benchmarking

Privacy-preserving sensing

References

2026

Constellation Dataset: Benchmarking High-Altitude Object Detection for an Urban Intersection

Mehmet Kerem Turkcan, Chengbo Zang, Sanjeev Narasimhan, Gyung Hyun Je, Bo Yu, Mahshid Ghasemi, Javad Ghaderi, Gil Zussman, and Zoran Kostic

International Journal of Computer Vision, Oct 2026

arXiv DOI HTML PDF Code Dataset

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 27.5 ms inference time using TensorRT.

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