Why not train on the coco dataset about snake HOT 8 CLOSED

We did train our network on COCO.
Currently, Snake with CenterNet gives 30.3 AP on test-dev set, and Snake with RepPoints gives about 35 AP.
We have been searching the good way to trained with the state-of-the-art detectors.

from snake.

We did train our network on COCO.
Currently, Snake with CenterNet gives 30.3 AP on test-dev set, and Snake with RepPoints gives about 35 AP.
We have been searching the good way to trained with the state-of-the-art detectors.

Thanks for your reply, maybe that the coco is not friendly for boundary based methods, maybe it is more focus on the detection. Hope to see your new results

from snake.

Thanks for your encourage!

from snake.

Thanks for your encourage!

Hi, I notice that you said you have tried snake with reppoints and got 35 mAP. I want to further ask with what backbone network you obtain this result? Thanks!

from snake.

This one: https://github.com/open-mmlab/mmdetection/blob/master/configs/reppoints/reppoints_moment_x101_dcn_fpn_2x_mt.py

from snake.

This one: https://github.com/open-mmlab/mmdetection/blob/master/configs/reppoints/reppoints_moment_x101_dcn_fpn_2x_mt.py

Thanks so much!

from snake.

35mAP means mask mAP or box mAP?

from snake.

35mAP means mask mAP.

from snake.