train error about keras-fasterrcnn HOT 4 OPEN

I just had this issue recently. Any suggestions how to solve this?

from keras-fasterrcnn.

Diagnosing dimension out of bounds errors is pretty hard since it could be a lot of things. Try using my pull request since I fixed a lot of issues with the repository there. If that doesn't fix your issue I can try diagnosing with you.

from keras-fasterrcnn.

@collinarnett there was a try/except block which was commented. What I did was, before you mentioned your PR, I removed the comments so that whole training process is inside a try/except block. It works but from time to time exception is caught. Would you be able to explain what is the main reason of this exception? How does it affect the training performance and as a result testing performance?

from keras-fasterrcnn.

As far as I know ROI Pooling is responsible for finding the location of the desired object by taking the best same from a region as outlined in more detail here. For this error, I would guess that the sliding window is overlapping out of bounds and if you're putting the entire training loop in a try except block, it would skip the training step.

I think the easiest way to diagnose this problem would be to test the ROI width as indicated by @wudi00 's error:

File "/media/disk/wudi/Keras-FasterRCNN-master/keras_frcnn/RoiPoolingConv.py", line 67, in call
w = rois[0, roi_idx, 2]

from keras-fasterrcnn.