Improving performance and understanding some detect.py params

[schubackindustries]

Hey all, great library here, and thankfully I was able to get it up and running on Windows! Not always a guarantee these days...

Anyways I was wondering about performance and some of the parameters I have available for detect.py. Specifically:

1. What are the main parameters for improving real-time detection speed?
   An obvious one is "--imgsz" which I have tried reducing.

2. What is "--max-det" aka "maximum detections per image"?

3. Can I improve performance by only trying to detect a given subset of object classes? For example I'm mainly interested in "person" and "car".

4. What is the difference between yolov5s.pt and yolov5s.ps ? Does one run faster than the other?

Thanks for any help!

[glenn-jocher]

@schubackindustries you can speed up detection with smaller models, smaller image sizes, and batched inference using PyTorch Hub (see PyTorch Hub tutorial):

YOLOv5 Tutorials

• Train Custom Data  🚀 RECOMMENDED
• Tips for Best Training Results  ☘️ RECOMMENDED
• Weights & Biases Logging  🌟 NEW
• Supervisely Ecosystem  🌟 NEW
• Multi-GPU Training
• PyTorch Hub  ⭐ NEW
• TorchScript, ONNX, CoreML Export 🚀
• Test-Time Augmentation (TTA)
• Model Ensembling
• Model Pruning/Sparsity
• Hyperparameter Evolution
• Transfer Learning with Frozen Layers  ⭐ NEW
• TensorRT Deployment