Open mychina75 opened 4 years ago
@mychina75 Sorry, the V100 GPU is not available for me now, so I can't give the speed for the required model using V100. If possible, you may test the speed yourself. And the speed for the present code should be higher than those listed in the arxiv paper, because we have optimized the code for the inference process recently.
Thank you for your reply. I am interesting in try your model on mobile device, like Android phone. Any suggestion about this? Thank you~
I am not familiar with the implementation in mobile device. But to enable the running in mobile device, decreasing the computational cost is really important. Here are some naive suggestions: 1) decrease the input size. 2) decrease the network parameters. Maybe using smaller backbone, reducing the parameters of heads, or using more fancy ways like networks slimming. 3) decrease the computational cost in post processing. For example, reducing the number of candidate proposals before NMS, and constrict the proposal numbers after NMS. 4) code implementation is also important. Maybe figuring out how YOLACT speed up their method is a good way.
Hi, Thank you for your work. The results are very good. I have a question about the comparison with state-of-the-art methods, in your paper Table1. All models are comparing using V100 GPU, and images size for EmbedMask defaults as 800 to 1333. so, for EmbedMask R-50-FPN, fps: 16.7 EmbedMask-600(size as 600 to 800) R-101-FPN, fps: 21.7
How about fps of EmbedMask-600, R-50-FPN? I'd like to compare with YOLACT-520. Thank you.