Several issues:
1) In training process, the training loss rate does not decrease, sometimes, even oscillation.
2) In inference process, it may take a long time to do the prediction (between 1 mins to 3mins)
3) In inference process, the predict result does not look good. If we compare the predicted bbox and labeled bbox, we can not find many overlap between them. The expected result will be: the iou between predicted bbox and labeled bbox should be large.
luhuicode
commented
6 years ago
Several issues: 1) In training process, the training loss rate does not decrease, sometimes, even oscillation. 2) In inference process, it may take a long time to do the prediction (between 1 mins to 3mins) 3) In inference process, the predict result does not look good. If we compare the predicted bbox and labeled bbox, we can not find many overlap between them. The expected result will be: the iou between predicted bbox and labeled bbox should be large.