Open liuhuiCNN opened 6 months ago
maybe you should read the paper first
Hi @liuhuiCNN, we evaluate the FPS with V100. For LVIS, we do not use TensorRT or FP16. For COCO, we adopt TensorRT(8) (CUDA=11.3)
Thanks. And is the FPS tested include NMS?
Considering that NMS can be optimized by reducing the pre_nms
(reducing pre_nms
to 3000 does not affect the accuracy) and downstream deployment, we exclude the NMS during FPS evaluation (LVIS).
Under what environment were performance FPS tested? What’s the CUDA cudnn tensorrt version? And what‘s the test methond or test script? Thank you!