This repository represents Ultralytics open-source research into future object detection methods, and incorporates our lessons learned and best practices evolved over training thousands of models on custom client datasets with our previous YOLO repository https://github.com/ultralytics/yolov3. All code and models are under active development, and are subject to modification or deletion without notice. Use at your own risk.
** GPU Latency measures end-to-end latency per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 32, and includes image preprocessing, PyTorch FP32 inference, postprocessing and NMS.
- June 9, 2020: CSP updates to all YOLOv5 models. New models are faster, smaller and more accurate. Credit to @WongKinYiu for his excellent work with CSP.
- May 27, 2020: Public release of repo. YOLOv5 models are SOTA among all known YOLO implementations, YOLOv5 family will be undergoing architecture research and development over Q2/Q3 2020 to increase performance. Updates may include CSP bottlenecks, YOLOv4 features, as well as PANet or BiFPN heads.
- April 1, 2020: Begin development of a 100% PyTorch, scaleable YOLOv3/4-based group of future models, in a range of compound-scaled sizes. Models will be defined by new user-friendly
*.yamlfiles. New training methods will be simpler to start, faster to finish, and more robust to training a wider variety of custom dataset.
| Model | APval | APtest | AP50 | LatencyGPU | FPSGPU | params | FLOPs | |
|---|---|---|---|---|---|---|---|---|
| YOLOv5-s (ckpt) | 35.5 | 35.5 | 55.0 | 2.5ms | 400 | 7.1M | 12.6B | |
| YOLOv5-m (ckpt) | 42.7 | 42.7 | 62.4 | 4.4ms | 227 | 22.0M | 39.0B | |
| YOLOv5-l (ckpt) | 45.7 | 45.9 | 65.1 | 6.8ms | 147 | 50.3M | 89.0B | |
| YOLOv5-x (ckpt) | 47.2 | 47.3 | 66.6 | 11.7ms | 85 | 95.9M | 170.3B | |
| YOLOv3-SPP (ckpt) | 45.6 | 45.5 | 65.2 | 7.9ms | 127 | 63.0M | 118.0B |
** APtest denotes COCO test-dev2017 server results, all other AP results in the table denote val2017 accuracy.
** All AP numbers are for single-model single-scale without ensemble or test-time augmentation. Reproduce by python test.py --img 736 --conf 0.001
** LatencyGPU measures end-to-end latency per image averaged over 5000 COCO val2017 images using a GCP n1-standard-16 instance with one V100 GPU, and includes image preprocessing, PyTorch FP32 inference at batch size 32, postprocessing and NMS. Average NMS time included in this chart is 1-2ms/img. Reproduce by python test.py --img 640 --conf 0.1
** All checkpoints are trained to 300 epochs with default settings and hyperparameters (no autoaugmentation).
Python 3.7 or later with all requirements.txt dependencies installed, including torch >= 1.5. To install run:
$ pip install -U -r requirements.txt
Inference can be run on most common media formats. Model checkpoints are downloaded automatically if available. Results are saved to ./inference/output.
$ python detect.py --source file.jpg # image
file.mp4 # video
./dir # directory
0 # webcam
rtsp://170.93.143.139/rtplive/470011e600ef003a004ee33696235daa # rtsp stream
http://112.50.243.8/PLTV/88888888/224/3221225900/1.m3u8 # http streamTo run inference on examples in the ./inference/images folder:
$ python detect.py --source ./inference/images/ --weights yolov5s.pt --conf 0.4
Namespace(agnostic_nms=False, augment=False, classes=None, conf_thres=0.4, device='', fourcc='mp4v', half=False, img_size=640, iou_thres=0.5, output='inference/output', save_txt=False, source='./inference/images/', view_img=False, weights='yolov5s.pt')
Using CUDA device0 _CudaDeviceProperties(name='Tesla P100-PCIE-16GB', total_memory=16280MB)
Downloading https://drive.google.com/uc?export=download&id=1R5T6rIyy3lLwgFXNms8whc-387H0tMQO as yolov5s.pt... Done (2.6s)
image 1/2 inference/images/bus.jpg: 640x512 3 persons, 1 buss, Done. (0.009s)
image 2/2 inference/images/zidane.jpg: 384x640 2 persons, 2 ties, Done. (0.009s)
Results saved to /content/yolov5/inference/output
Download COCO, install Apex and run command below. Training times for YOLOv5s/m/l/x are 2/4/6/8 days on a single V100 (multi-GPU times faster). Use the largest --batch-size your GPU allows (batch sizes shown for 16 GB devices).
$ python train.py --data coco.yaml --cfg yolov5s.yaml --weights '' --batch-size 64
yolov5m 48
yolov5l 32
yolov5x 16
To access an up-to-date working environment (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled), consider a:
- GCP Deep Learning VM with $300 free credit offer: See our GCP Quickstart Guide
- Google Colab Notebook with 12 hours of free GPU time.
- Docker Image https://hub.docker.com/r/ultralytics/yolov5. See Docker Quickstart Guide
Ultralytics is a U.S.-based particle physics and AI startup with over 6 years of expertise supporting government, academic and business clients. We offer a wide range of vision AI services, spanning from simple expert advice up to delivery of fully customized, end-to-end production solutions, including:
- Cloud-based AI surveillance systems operating on hundreds of HD video streams in realtime.
- Edge AI integrated into custom iOS and Android apps for realtime 30 FPS video inference.
- Custom data training, hyperparameter evolution, and model exportation to any destination.
For business inquiries and professional support requests please visit us at https://www.ultralytics.com.
Issues should be raised directly in the repository. For business inquiries or professional support requests please visit https://www.ultralytics.com or email Glenn Jocher at glenn.jocher@ultralytics.com.