Ming233/YOLOX-Nano-P2-UAV-Small-Detection

YOLOX-Nano+-P2 Checkpoints

This repository contains YOLOX-Nano+-P2 checkpoints from the UAV-QIEA Edge Detection project.

Paper: Edge-Constrained UAV Small-Object Detection with P2 Enhancement and Quantum-Inspired Lightweight Structure Search
Code: Ming23233/UAV-QIEA-Edge-Detection

Model Description

YOLOX-Nano+-P2 is a YOLOX-Nano detector variant with an added high-resolution P2 detection branch for dense UAV small-object scenes. The P2 branch adds a stride-4 detection path, complementing the original stride-8/16/32 detection heads and improving small-object detection on VisDrone-style aerial imagery.

These checkpoints are provided to make the paper easier to reproduce and inspect. The full research code, dataset conversion utilities, and evaluation scripts are maintained in the GitHub repository linked above.

Files

Path	Description
checkpoints/yolox_nano_p2_seed42_best_ckpt.pth	YOLOX-Nano+-P2 best checkpoint, seed 42
checkpoints/yolox_nano_p2_seed43_best_ckpt.pth	YOLOX-Nano+-P2 best checkpoint, seed 43
checkpoints/yolox_nano_p2_seed44_best_ckpt.pth	YOLOX-Nano+-P2 best checkpoint, seed 44
experiments/yolox_nano_visdrone_det_640_p2.py	Main YOLOX experiment file for the P2 variant
experiments/	Minimal experiment configuration chain copied from the release code
configs/uav_tdmnet.yaml	Project-level experiment metadata
results/	Lightweight summary tables from the manuscript
metadata/checkpoints_manifest.json	Checkpoint provenance, metrics, size, and SHA256 hashes

Checkpoint Summary

The three released checkpoints correspond to 100-epoch VisDrone-DET training runs with input size 640 x 640.

Checkpoint	Seed	Best epoch	AP50:95	AP50	AP-small	SHA256
yolox_nano_p2_seed42_best_ckpt.pth	42	95	0.0667	0.1365	0.0386	8520A42760A34D6DCAA58292C4659572BC2B0761133D22071C7BACBDA0F171B0
yolox_nano_p2_seed43_best_ckpt.pth	43	86	0.0699	0.1419	0.0400	08619886486339D8D9FAF5A70486BDA38BA81430A935CCDA8104FC313E990347
yolox_nano_p2_seed44_best_ckpt.pth	44	95	0.0688	0.1384	0.0374	BF8AEF009D2B18EFEA609CEE538B6D65AF1C66F4446DF8A527A1BF9314DD0D64

The manuscript-level three-seed summary reports:

Model	Seeds	AP50:95 mean	AP-small mean	Recall50 mean	Small-object gain
Baseline	3	0.0635	0.0295	0.2936	0.00%
+P2	3	0.0684	0.0387	0.2964	31.10%
QIEA-Final	3	0.0530	0.0328	0.3056	11.08%

Usage

Clone and install the code repository first:

git clone https://github.com/Ming23233/UAV-QIEA-Edge-Detection.git
cd UAV-QIEA-Edge-Detection
pip install -r requirements.txt
pip install -r third_party/ByteTrack/requirements.txt
pip install -e third_party/ByteTrack

Download a checkpoint from this Hugging Face repository:

pip install -U "huggingface_hub[cli]"
hf download Ming233/YOLOX-Nano-P2-UAV-Small-Detection checkpoints/yolox_nano_p2_seed43_best_ckpt.pth --local-dir ./hf_checkpoints

Then evaluate or run inference with the corresponding experiment file from the GitHub repository:

python third_party/ByteTrack/tools/eval.py \
  -f third_party/ByteTrack/exps/example/uav/yolox_nano_visdrone_det_640_p2.py \
  -c ./hf_checkpoints/checkpoints/yolox_nano_p2_seed43_best_ckpt.pth \
  -b 1 \
  -d 1

Adjust dataset paths according to your local VisDrone-DET COCO-format layout. See the GitHub repository documentation for dataset conversion and reproducibility details.

Training Data

The released checkpoints were trained on a COCO-format conversion of VisDrone-DET. This repository does not redistribute VisDrone images or annotations. Please obtain datasets from their official sources and follow their licenses.

Limitations

These checkpoints are research artifacts for UAV small-object detection. They are not a general-purpose object detector and may not generalize reliably to non-UAV imagery, different camera altitudes, unusual sensor types, or deployment settings outside the evaluation protocol.

Citation

If you use these checkpoints or the associated code, please cite the manuscript and the original YOLOX/ByteTrack work.

@misc{lei2026edgeconstraineduavsmallobject,
  title={Edge-Constrained UAV Small-Object Detection with P2 Enhancement and Quantum-Inspired Lightweight Structure Search},
  author={Wuming Lei and Yanbin Gao and Mingyan Sun and Xiaobin Li and Xuechen Liang},
  year={2026},
  eprint={2606.09081},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2606.09081}
}

License

The newly added research code is released under the MIT license. Third-party components retain their original licenses. Dataset usage remains governed by the dataset owners.