A simple idea to solve the error : "rpn_loss_box: nan"

There is an error with bounding boxes’ coordinates,when some of coordinate values become 65535.

So, we can open the file named “minibatch.py” , find the line:

gt_boxes = np.empty((len(gt_inds), 5), dtype=np.float32)

and insert some codes in front of it :

boxes = roidb[0]['boxes'][gt_inds, :]         
for idx in range(len(boxes)):
    boxes[idx][boxes[idx]==65535] = 0

这可能算不上是个好的解决方案 ~

[kinzzoku]

@FMsunyh Did you delete the .pkl files created in in the cache folder of the dataset? For me deleting the -1 didn’t have any effect too at first. But deleting the -1 and after that the .pkl files solved it for me.

[endernewton]

the underlying reason might be your dataset is 0-based rather than 1-based in coordinates?