Graph suggestion - confusion matrices and Precision vs Recall curves side by side for each model architecture

[valentinitnelav]

Hi @stark-t , I just saw this pre-print today and got inspired by their figure 3. They present confusion matrices and Precision vs Recall curves side by side and we could do the same for each model architecture: YOLOv5 nano, small and YOLOv7 tiny.

What do you think? Can we easily get these Precision vs Recall curves for the test datasets? Perhaps we can easily get them if I run val.py on the test dataset, but I didn't try and not sure if it makes sense, because I saw that detect.py might even load data differently than val.py and therefore might give different results - for example see this issue Why result on val.py and detect.py is different? with this comment https://github.com/ultralytics/yolov5/issues/8133#issuecomment-1149237054 :

thanks for asking about the differences between train.py, detect.py and val.py in YOLOv5 rocket. These 3 files are designed for different purposes and utilize different dataloaders with different settings. train.py dataloaders are designed for a speed-accuracy compromise, val.py is designed to obtain the best mAP on a validation dataset, and detect.py is designed for best real-world inference results.

If you can give me precision and recall per class for each architecture derived from the results given by detect.py, I can make these graphs myself.

[stark-t]

provide all heatmaps in appendix