search

neufieldrobotics / NeuFlow_v2

Apache License 2.0
92 stars 6 forks source link

Hello, I have calculated the FLOPs of NeuFlowV2 and SEA-RAFT and found that the FLOPs of NeuFlowV2 seem to be higher, but why is it so much faster than SEA-RAFT? #3

Open yuefanhao opened 2 months ago

Study-is-happy commented 2 months ago

Very good question. Our assumption is that, depth-wise conv has way much lower FLOPs but the low-level library (cudnn) does not optimize the computation time. We use many 3x3 convolution layers which is highly optimized by low-level library.

yuefanhao commented 2 months ago

@Study-is-happy Thank you very much. Another question is, have you done any optimization for cost volume? Because you have done one iteration at 1/16 resolution and eight iterations at 1/8 resolution, which is an operation that consumes a lot of computing and memory. I need to estimate optical flow at a larger resolution, so the cost volume will be very large, and it's very time-consuming to look up.