this repo is good ONLY for INFERENCE with PROVIDED WEIGHTS.

[ScottHoang]

like the title said, this implementation suffers serious issue when trained from scratch with mAP stalling at around 20%. If you are planning to train this on custom data, I suggest go looking at a different PyTorch implementation of yolo.

[tjiagoM]

Hi @voodoopotato do you have a suggestion of a good pytorch implementation? I've found different ones around so was curious to see what you think.

[ScottHoang]

@tjiagoM yes I did.
https://github.com/ultralytics/yolov3. I achieved much much better results using their model.

[Khalifa1997]

@voodoopotato @tjiagoM do you think I can achieve results here training on only 100 images?

[ScottHoang]

No.

[Khalifa1997]

@voodoopotato thanks for the quick reply, I am abit lost here though, what approachs do you think I should look into with dataset of that size? thanks!

[ScottHoang]

@Khalifa1997 maybe I was a bit in haste. One hundred images are very small sample population but can be artificially multiplied with proper data-augmentation. It depends on how many cls you have, how common they are in your samples, etc...
But as I said in my title, this repo is only suitable for inference.

[aningineer]

@tjiagoM yes I did.
https://github.com/ultralytics/yolov3. I achieved much much better results using their model.

Are there any minimal Yolov3 implementations that achieve the original results from the authors? The Ultralytics repo seems to be quite heavy to develop on.

[XZLeo]

This issue may not be true. I trained on my own dataset from scratch, the mAP reaches 50+%.