prediction boxes is wrong.

[DeepTiming]

First Image is CornerNet-Saccade result and second Image is CornerNet-squeeze result.

[DeepTiming]

Is there a problem in NMS process?

[heilaw]

CornerNet-Saccade estimates object locations via attention maps and detects object in a small region around each location. If the predicted attention maps are not sufficiently good, CornerNet-Saccade may miss an object or only see part of it. It seems to me that CornerNet-Saccade is not able to generate sufficiently good attention maps for this image.

More details about how the quality of the attention maps affects the performance of CornerNet-Saccade can be found in the experiment section of our paper.

[DeepTiming]

Get it! Thank you very much! Good job!