--- license: apache-2.0 pipeline_tag: object-detection library_name: ultralytics datasets: - Voxel51/VisDrone2019-DET tags: - object-detection - aerial-imagery - drone - visdrone - ultralytics - pytorch - computer-vision metrics: - map50 - map50-95 - precision - recall - f1 base_model: "Ultralytics/YOLOv9" --- # YOLOv9c Finetuned on VisDrone Fine-tuned YOLOv9c object detector for aerial imagery using the VisDrone benchmark dataset. This model is part of the **VisDrone Detection Model Zoo**, a collection of YOLO models trained and evaluated under a common pipeline for aerial object detection. ## Detection Showcase

VisDrone Detection Demo

--- ## Performance | Metric | Score (%) | | ---------- | --------------- | | mAP@50 | 37.22 | | mAP@50-95 | 21.73 | | Precision | 51.99 | | Recall | 39.77 | | F1 Score | 45.07 | | Parameters | - | | FLOPs | - | --- ## Evaluation Protocol Metrics reported in this model card are computed on the VisDrone test set with ground-truth annotations available for evaluation. --- ## VisDrone Model Zoo | Rank | Model | mAP@50 | mAP@50-95 | Precision | Recall | | -------------------------- | -------------------- | ------------- | --------------- | ----------------- | -------------- | | | | | | | | | 1 | YOLOv9e | 40.02 | 23.73 | 54.78 | 42.42 | | | | | | | | | 2 | YOLOv11x | 38.44 | 22.6 | 52.41 | 41.43 | | | | | | | | | 3 | YOLOv26x | 38.33 | 22.48 | 52.91 | 41.06 | | | | | | | | | 4 | YOLOv11l | 37.14 | 21.85 | 51.87 | 40.33 | | | | | | | | | 5 | YOLOv10x | 37.24 | 21.81 | 52.59 | 39.84 | | | | | | | | | 6 | YOLOv26l | 37.65 | 21.75 | 51.6 | 40.42 | | | | | | | | | 7 | YOLOv9c | 37.22 | 21.73 | 51.99 | 39.77 | | | | | | | | | 8 | YOLOv8x | 36.81 | 21.52 | 51.91 | 39.78 | | | | | | | | | 9 | YOLOv10l | 35.95 | 21.09 | 52.13 | 38.48 | | | | | | | | | 10 | YOLOv9m | 36.19 | 20.95 | 51.05 | 39.12 | | | | | | | | | 11 | YOLOv8m | 34.39 | 19.95 | 48.18 | 38.2 | | | | | | | | | 12 | YOLOv8s | 31.95 | 18.24 | 45.99 | 35.49 | | | | | | | | | 13 | YOLOv8n | 28.18 | 15.77 | 40.86 | 31.81 | | | | | | | | | 14 | YOLOv11n | 27.59 | 15.46 | 39.58 | 31.74 | | | | | | | | | 15 | YOLOv10n | 27.65 | 15.32 | 41.02 | 31.68 | | | | | | | | | 16 | YOLOv26n | 26.73 | 14.64 | 38.6 | 31.14 | | | | | | | | --- ## Per-Class Performance | Class | mAP@50 | mAP@50-95 | | -------------------------- | --------------- | ----------------- | | | | | | pedestrian | 32.83 | 13.19 | | | | | | people | 17.86 | 6.15 | | | | | | bicycle | 14.65 | 6.23 | | | | | | car | 75.59 | 48.51 | | | | | | van | 40.93 | 27.65 | | | | | | truck | 48.38 | 32.16 | | | | | | tricycle | 25.69 | 13.77 | | | | | | awning-tricycle | 20.42 | 12.05 | | | | | | bus | 58.92 | 42.3 | | | | | | motor | 36.94 | 15.34 | | | | | --- ## Evaluation Visualizations ### Precision-Recall Curve ![PR Curve](BoxPR_curve.png) ### F1 Curve ![F1 Curve](BoxF1_curve.png) ### Confusion Matrix ![Confusion Matrix](confusion_matrix.png) --- ## Dataset VisDrone is a large-scale benchmark for object detection in aerial imagery captured from unmanned aerial vehicles (UAVs). The dataset contains diverse scenes including: * Urban environments * Residential areas * Traffic intersections * Crowded pedestrian regions ### Classes * pedestrian * people * bicycle * car * van * truck * tricycle * awning-tricycle * bus * motor --- ## Usage ### Install Dependencies ```bash pip install ultralytics huggingface_hub ``` ### Load Model from Hugging Face ```python from huggingface_hub import hf_hub_download from ultralytics import YOLO weights = hf_hub_download( repo_id="dronefreak/yolov9c-visdrone", filename="best.pt" ) model = YOLO(weights) ``` ### Run Inference ```python results = model.predict( source="image.jpg", conf=0.25 ) results[0].show() ``` --- ## Training Configuration | Setting | Value | | ---------------- | ------------------------------- | | Epochs | 300 | | Dataset | VisDrone2019-DET | | Framework | Ultralytics YOLO | | Training Toolkit | VisDrone Dataset Python Toolkit | --- ## Repository Contents ```text best.pt results.csv args.yaml BoxPR_curve.png BoxF1_curve.png confusion_matrix.png assets/visdrone_showcase.gif README.md ``` --- ## Related Resources * VisDrone Detection Model Zoo (Hugging Face Collection) * VisDrone Dataset Python Toolkit: https://github.com/dronefreak/VisDrone-dataset-python-toolkit * VisDrone Dataset: https://github.com/VisDrone/VisDrone-Dataset --- ## Training Framework These models were trained using the VisDrone Dataset Python Toolkit, an open-source framework for aerial object detection research and benchmarking on the VisDrone dataset. Features include: * Dataset preparation and conversion utilities * Training and evaluation pipelines * Detection benchmarking * Visualization tools * Support for multiple YOLO model families Repository: https://github.com/dronefreak/VisDrone-dataset-python-toolkit If you find these models useful, please consider starring the repository. --- ## Known Limitations Performance may degrade in: * Extremely dense crowds * Heavy occlusions * Severe motion blur * Very small objects occupying only a few pixels * Night-time or low-light aerial imagery --- ## Citation If you use this model in your research, please consider citing: 1. The VisDrone dataset 2. The original YOLO architecture 3. The VisDrone Detection Toolkit ```bibtex @article{visdrone2019, title={Vision Meets Drones: A Challenge}, author={Zhu, Pengfei and Wen, Longyin and Du, Dawei and Bian, Xiao and Ling, Haibin and Hu, Qinghua}, journal={International Journal of Computer Vision}, year={2021} } @software{Saksena_VisDrone_Detection_Toolkit_2025, author = {Saksena, Saumya Kumaar}, title = {VisDrone Detection Toolkit: Modern PyTorch Implementation for Aerial Object Detection}, url = {https://github.com/dronefreak/VisDrone-dataset-python-toolkit}, version = {2.0.0}, year = {2025} } ```