prepare_roidb() fails on very small overlaps

[Seanmatthews]

Some background, in case I imposed the problem on myself through some improper setup: I'm training a very small dataset of 11 classes over 62 images. I used labelImg to annotate several objects per image (but not every object in every image). I follow the instructions for training here, using the same data for training and testing. Instead of the suggested training command, I altered faster_rcnn_alt_opt.sh to accommodate my dataset. The training progresses through "Stage 1 RPN, init from ImageNet model" and "Stage 1 RPN, generate proposals" to "Stage 1 Fast R-CNN using RPN proposals, init from ImageNet model", where it throws the following error:

Process Process-3: Traceback (most recent call last): File "/usr/lib/python2.7/multiprocessing/process.py", line 258, in _bootstrap self.run() File "/usr/lib/python2.7/multiprocessing/process.py", line 114, in run self._target(*self._args, **self._kwargs) File "./tools/train_faster_rcnn_alt_opt.py", line 189, in train_fast_rcnn roidb, imdb = get_roidb(imdb_name, rpn_file=rpn_file) File "./tools/train_faster_rcnn_alt_opt.py", line 67, in get_roidb roidb = get_training_roidb(imdb) File "/home/dev/src/ariel/deep/py-faster-rcnn/tools/../lib/fast_rcnn/train.py", line 122, in get_training_roidb rdl_roidb.prepare_roidb(imdb) File "/home/dev/src/ariel/deep/py-faster-rcnn/tools/../lib/roi_data_layer/roidb.py", line 51, in prepare_roidb assert all(max_classes[nonzero_inds] != 0) AssertionError

Looking into the function, I noticed that at this step through prepare_roidb, my max_overlaps array contained a number of close-to-zero values that numpy groups into the nonzero_inds array (as it should). However, those values in the max_classes array are indeed zero.

Is this an error on my part? If so, how did I cause it? If not, the solution is clearly to set negligible numbers to zero, but I'm reluctant to make that assumption without your input.

[Seanmatthews]

As an example, he are the list of max_overlaps which coincide with a max_classes value of 0:

[  6.77344190e-07   4.42691453e-06   4.88126204e-07   4.04364187e-07
   2.71648588e-07   9.20962691e-07   6.82705547e-07   9.97144866e-07
   7.92740821e-07   1.30538729e-06   2.19987305e-06   7.94905247e-07
   1.43210104e-06   5.10457312e-07   3.52238385e-06   1.53103429e-06
   2.77859908e-06   1.86146201e-06   5.83908729e-07   4.26315722e-07
   1.38307896e-05   4.33050900e-06   9.02547413e-07   9.53889298e-07
   5.21114907e-06   1.65058395e-06   7.25464179e-06   6.62295361e-06
   1.75827870e-06   4.18398413e-05   2.70081387e-06   4.92181562e-06
   8.22064612e-06   1.44625517e-06   2.23241477e-06   1.47799483e-06
   3.38878613e-06   8.17282944e-06   3.55347788e-06   1.38796704e-05
   8.31836860e-06   2.11551614e-05   4.84880138e-06   2.27672831e-06
   1.75890018e-05   6.81285201e-06   8.14911346e-06   1.18152111e-05]

[micros-uav]

@Seanmatthews I also have this problem. I just train the 'person' class of the voc2007 dataset.

[Neppord]

I'm also interested in this question. What happens if you have regions in your training set that overlaps completely or partially, is this valid input and training.

[alchu8]

@Seanmatthews What did you end up doing?

[FelixZhang7]

@Seanmatthews hi，i am training my data with faster-rcnn but got the same error assert all(max_classes[nonzero_inds] != 0) AssertionError,Please let me know if you have any solution to this problem.

[PoonamRajput]

I am also having the same issue. Does anyone solve it?