search

princeton-vl / CornerNet-Lite

BSD 3-Clause "New" or "Revised" License
1.78k stars 431 forks source link

The detection effect is much worse than yolov3. #94

Open tuyaliang opened 5 years ago

tuyaliang commented 5 years ago

The CornerNet Saccade detection effect is much worse than yolov3. The first picture is CornerNet Saccade,and the second is yolov3. demo_out

predictions

nanyoullm commented 5 years ago

how about the CornerNet-Squeeze?

atlascollege commented 5 years ago

how about the CornerNet-Squeeze?

even worth

mochechan commented 5 years ago

My result also shows that Yolov3 has better accuracy than Cornernet_Lite. I have no idea why the paper shows that Cornernet_Lite has higher mAP.

heilaw commented 5 years ago

@mochechan Can you provide more details? Did you evaluate it on COCO?

mochechan commented 5 years ago

@heilaw Sure. I never evaluated Cornernet_Lite on COCO. I just train my own dataset to detect license plates in order to evaluate the possibility to replace my current object detection by using Yolov3. Cornernet_Saccade detects objects successfully in most cases for me, but I notice that Cornernet_Saccade makes more mistakes than Yolov3 do. First, Cornernet_Saccade makes more false negative and false positive. Second, Cornernet_Saccade generates redundant bounding boxes for a single object. Currently, I have no idea how to solve those problems. Could you please propose some advices? Thank you. Therefore, Yolov3 has better accuracy than Cornernet_Lite for the same given dataset in my evaluation.

By the way, Cornernet_Saccade has better stability than Yolov3. In my current object detection for video, the bounding boxes have the syndrome, jitter. For this point, Cornernet_Saccade is better than Yolov3.