search

omasaht / headpose-fsanet-pytorch

Pytorch implementation of FSA-Net: Learning Fine-Grained Structure Aggregation for Head Pose Estimation from a Single Image
MIT License
113 stars 31 forks source link

Time cost for each frame #7

Open sunwell1994 opened 3 years ago

sunwell1994 commented 3 years ago

I use the demo.py for a 3-hour video. I find the time cost for each frame increases by the timestamp of the video. At the first half-hour, each frame costs ~20ms, then it costs ~50ms. It costs ~110ms for the final frames of the video. I have checked memory and GPU memory, they did not change a lot. Do you have any idea why it slows down?

omasaht commented 3 years ago

Hey,

Sorry, I have not tested it on such a long video so I can't comment on that. I will, however, refer you to try the original author's implementation and its demo: https://github.com/shamangary/FSA-Net . It's in Keras. Please do let me know if you find whether it works there or not. I would also ask you try using PyTorch model directly for inferencing and don't use ONNX runtime. Maybe there's a bug with ONNX runtime.