voc 2017 training error

[jwnsu]

Thanks a lot for a few recent fixes, now training is underway.

During training, it would stop on random itreation (it may continue for hundreds of images, then encounter the error ):
" labels, regression_targets = anchor_targets(image, boxes_batch[0],len(self.classes))
File "/home/ai/krn/keras_retinanet/preprocessing/anchors.py", line 29, in anchor_targets
argmax_overlaps_inds = np.argmax(overlaps, axis=1)
File "/home/tf/local/lib/python2.7/site-packages/numpy/core/fromnumeric.py", line 963, in argmax
return _wrapfunc(a, 'argmax', axis=axis, out=out)
File "/home/tf/local/lib/python2.7/site-packages/numpy/core/fromnumeric.py", line 57, in _wrapfunc
return getattr(obj, method)(*args, **kwds)
ValueError: attempt to get argmax of an empty sequence".

Seems to be some type of batch data generation issue somewhere.

[hgaiser]

Ah seems you have an image where none of the anchors have a significant overlap with any of the ground truth bounding boxes. Could you find out for which image this error occurs? Then I can debug it better.

Regardless, processing images with no ground truth boxes or where no box has significant overlap should be fixed.