uwaisasghar/p2pnet-visdrone-head-counting

P2PNet VisDrone head-point localization

This is an inference-only P2PNet checkpoint fine-tuned on fold 0 of a deterministic, sequence-level VisDrone2020-CC split. It returns one (x, y, confidence) proposal per retained head and uses the number of retained points as the crowd count.

Source and complete training documentation: owaisasghar/uav-crowd-counting.

The generic Hugging Face object-detection widget is disabled because its bounding-box output contract does not match point localization. Use the included inference script or the Colab notebook, which compares this model with UAV-Dot side by side.

Results

Independent test set

These results use the official VisDrone2020-CC test sequences (900 frames from 30 held-out sequences, 137 people per frame), threshold 0.50 and the DroneCrowd localization protocol:

Metric	P2PNet (this model)	UAV-Dot, released weights
Count MAE / RMSE	21.86 / 29.58	25.04 / 33.66
F1 at 16 px	0.725	0.726
L-mAP (DroneCrowd toolkit, 1–25 px)	52.49	51.00
fp16 latency, 1920×1080, RTX 4060 Laptop GPU	114.7 ms	29.3 ms

Validation fold

The values below use the 510-frame validation fold that also selected the checkpoint. They are not independent-test results.

Metric	Value
Count MAE at threshold 0.50	10.0255
Count RMSE at threshold 0.50	13.0832
Precision at 16 px	0.7988
Recall at 16 px	0.8315
F1 at 16 px	0.8148
Ranked AP at 16 px	0.8558
Validation images	510
Validation head points	72,171

Point metrics use one-to-one Hungarian matching and a fixed 16-pixel radius. Ranked AP is fixed-radius AP, not a head-size-normalized nAP metric. Lower MAE and RMSE are better; the count metrics are not classification accuracy.

Files

File	Purpose
model.safetensors	Inference-only model state
p2pnet_runtime.py	Standalone architecture and preprocessing
inference.py	Command-line image inference
config.json	Architecture and preprocessing metadata
export_manifest.json	Export checksum and provenance
LICENSE	Academic-research-only terms

The original .pth training checkpoint also contains optimizer and scheduler state. This model repository intentionally distributes only model tensors in SafeTensors format.

Verified inference export:

File: model.safetensors
Size: 86,362,656 bytes
SHA-256: acb2f7196538fff3187b2e02fd9629802a9f7d5faec3d8bdbe9c47cd6ed003c9
Tensors: 123
Source checkpoint epoch: 49 (zero-based value 48)

Usage

Install the dependencies and run the included script from the model repository directory:

python -m pip install -r requirements.txt
python inference.py \
  --weights model.safetensors \
  --input aerial_image.jpg \
  --output predicted_points.png \
  --json-output predicted_points.json \
  --threshold 0.5 \
  --max-side 1920 \
  --device auto

The JSON output contains the predicted count, inference time, device, input scale, and every retained point with its confidence score.

Training data and split

• Dataset: VisDrone2020-CC, obtained separately from its authorized source.
• Training: 1,950 frames and 290,556 annotated head points.
• Validation: 510 frames and 72,171 annotated head points.
• Split: sequence-level fold 0 of 5, seed 42.
• Best epoch: 49; training stopped at epoch 79 by early stopping.

VisDrone images and annotations are not included in this model repository.

Relationship to the referenced UAV paper

This model is an extension inspired by aerial crowd-counting research, not a reproduction of the model in Ptak et al., Remote Sensing 2022. That work used density regression, optical-flow channels, multiple architectures and input resolutions, embedded-device measurements, and five-fold validation. This checkpoint instead adapts the point-based P2PNet architecture to return individual head locations.

Limitations

• Only one of five planned sequence-level folds has been trained.
• Threshold selection and checkpoint selection used the same validation fold.
• Full-resolution forward inference averaged 243.14 ms on an RTX 4060 Laptop GPU, approximately 4.11 forward passes per second before most I/O.
• Small, occluded, blurred, or low-contrast heads may be missed.
• Background structures may create false positives.
• Performance is not established for new cameras, altitudes, locations, weather, thermal imagery, or the separate task of vehicle detection.

Intended use

Academic research on aerial crowd counting and explicit point localization. The model is not intended for identity recognition, demographic inference, autonomous law-enforcement decisions, or unvalidated safety-critical use.

License

P2PNet redistribution and use are restricted to academic research. Review LICENSE before downloading, using, or redistributing the checkpoint. This is not an MIT, Apache, or commercially permissive model.

Citation

Please cite the upstream P2PNet work:

@inproceedings{song2021p2pnet,
  title={Rethinking Counting and Localization in Crowds: A Purely Point-Based Framework},
  author={Song, Qingyu and Wang, Changan and Jiang, Zhengkai and Wang, Yabiao and Tai, Ying and Wang, Chengjie and Li, Jilin and Huang, Feiyue and Wu, Yang},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2021}
}