search

WXinlong / SOLO

SOLO and SOLOv2 for instance segmentation, ECCV 2020 & NeurIPS 2020.
Other
1.69k stars 307 forks source link

Question about Matrix NMS time measurement #150

Closed dps42 closed 3 years ago

dps42 commented 3 years ago

Hi. I'm reading your SOLOv2 paper, which says Matrix NMS takes < 1 ms. How was the timing measured?

"For usage, we just need thresholding and selecting top-k scoring masks as the final predictions."

Does the < 1ms measurement include both thresholding and selecting top-k scoring masks too? (I think top-k is max_per_img = 100).

Or does it only include what's in "def matrix_nms()" in https://github.com/WXinlong/SOLO/blob/04c9c42803c3341b8f31b6e9ed0f5d32b7aee72a/mmdet/core/post_processing/matrix_nms.py ? If yes, what's the overall timing for the whole NMS process?

Thanks.

zzzzzz0407 commented 3 years ago

We return None before NMS for simplicity, therefore, I think the time including matrix nms / thresholding and selecting top-k scoring masks.

dps42 commented 3 years ago

thank you for the response. < 1ms measurement for everything seems incredibly fast.