Yolov3: output of the accuracy/quality of the network

[Alex1101a]

Hi everybody,

I have a question about the Yolov3 network: is there a way to determine the accuracy like with a traditional classifier for the training process? I already watched several tutorials and did some research on the Tensorboard webpage, but couldn't find an answer because it seems like everbody just uses the loss function and so on...
Or is there only the possibility to determine the accuracy/ the quality of the net via Precision, recall etc?
Thank you very much for tips and explanations :)

[monocongo]

The darknet model (which I assume this one is based upon) provides periodic mAP scores as it's training if you use the -map option. See section 8.1 in the README. Something like that may be possible for this model as well?

[glenn-jocher]

@Alex1101a @monocongo this repo provides you with Precision, Recall, mAP and F1 on your validation set after each training epoch automatically. See https://github.com/ultralytics/yolov3/wiki/Train-Custom-Data to get started.

[monocongo]

@glenn-jocher excellent work, very helpful, thanks!

[Alex1101a]

@glenn-jocher : okay then these are all details about the accuracy. i looked again at what the individual terms mean and think that's what i want.
Since I've only worked with classifiers so far, I thought YOLO also has an accuracy for the whole network, which is output during the training process.
But I've now discovered that YOLO has to take more into account when it comes to calculating the accuracy. Did I understand correctly that the mAP indicates the quality of the net, so to speak?

[glenn-jocher]

@Alex1101a accuracy is a classification metric, this is object detection. The metrics supplied all measure the 'quality' of the output so to speak, and are the most commonly used metrics.