Suggestion to distinguish digital digit 2 and 5 in images.

[dsbyprateekg]

Hi,

I need your suggestion to improve my custom-trained model based on YOLOv7W. In the following type of images, my model is getting confused between the digits 2 and 5-

The model is creating 2 bbox around digit 2. The confidence score for digit 2 is 82% and for digit 5 it is 87% although digit 5 is not present here. In my training, I only enabled the rotation hyperparameter to 90.0, the rest are default settings.

Please suggest to me how to fix this issue?

[faizan1234567]

When you rotate 2, it will be like 5. It is recommended to use data augmentation options that capture your real-world scenario. Further, add more images of digit 2 in the training dataset to make it diverse. Hope it helps

[faizan1234567]

did you enable horizontal flip? This might cause making making 5 from 2.

[dsbyprateekg]

@faizan1234567 I had enabled the rotation option and disabled the fliplr option but issue still exists.

[faizan1234567]

Could you please tell what's your performance metric on training, val, and test. And are you testing on a video? Do you have enough images that differentiate 5 from 2. Could you reduce your rotation range? I think adding more data for 2 and 5 in different settings will improve.

[dsbyprateekg]

I am testing on images only. And this was the metric result when I trained with no rotation (0.0) and no fliplr option:

[faizan1234567]

is it on test set?

[dsbyprateekg]

No, training and val accuracy.

[faizan1234567]

Your model has good performance on validation set on 2 and 5 classes. Did you measure it's accuracy on test set? Where does it make errors?

[dsbyprateekg]

Yes, I tested 185 images and among them, in 27 images it is failing.

[damvantai]

@faizan1234567 I had enabled the rotation option and disabled the fliplr option but issue still exists.

You disable in file hyp.scratch.p5 or in code with option --evolve? you should access /run/train/name_train{}/trainbatch{}.jpg? to watch image have flipup or flipleft right?