search

hujiecpp / ISTR

ISTR: End-to-End Instance Segmentation with Transformers (https://arxiv.org/abs/2105.00637)
203 stars 28 forks source link

why SOLOv2 speed slower than MaskRCNN in paper? #5

Open lucasjinreal opened 3 years ago

lucasjinreal commented 3 years ago

SOLOv2 becomes slowest model in your papers, is that really? But as SOLOv2 official paper claims it should faster than most instance segmentation methods, at least compare with MaskRCNN.

hujiecpp commented 3 years ago

SOLOv2 becomes slowest model in your papers, is that really? But as SOLOv2 official paper claims it should faster than most instance segmentation methods, at least compare with MaskRCNN.

Hi @jinfagang , We are sorry that the running time of SOLOv2 in our paper is WRONG ... as we list the whole testing time of SOLOv2 (based on mmdet) including images-reading, post-processing, etc.

The pure running time of SOLOv2 on a single V100 GPU is (as shown in this page): 54ms (Res50) / 66ms (Res101).

The pure running time of SOLOv2 on a single 1080Ti GPU is (we re-run the code in this page): ～63ms (Res50) / ～84ms(Res101).

Thanks for indicting this issue and we will update the paper as soon as possible.

lucasjinreal commented 3 years ago

@hujiecpp that make more sense. Actually I hope ISTR to be more faster, if it's accurate and e2e easy to deploy. What's the fastest model currently ISTR can achieve?

hujiecpp commented 3 years ago

Currently, the fastest model is still ISTR-Res50 as I am now lacking of GPUs to train new models with ISTR. (Orz I plan to implement ISTR with mobilenet / vovnet backbone like centermask2.

lucasjinreal commented 3 years ago

@hujiecpp Does LSTR also requires RoiAlign layer?

hujiecpp commented 3 years ago

@jinfagang yes