omasaht / headpose-fsanet-pytorch

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

Ask for your advice #2

Open zhaojc001 opened 4 years ago

zhaojc001 commented 4 years ago

Thanks your great work.I have the following questions,please help me: how to get the fsanet-var-iter-688590.onnx model in the demo.py code. what is the distinct of between fsanet-1x1-iter-688590.onnx and fsanet-var-iter-688590.onnx.

omasaht commented 4 years ago

Hey, right now I am running both fsanet-var-iter-688590.onnx and fsanet-1x1-iter-688590.onnx models in demo.py and averaging their result. If you only want fsanet-var-iter-688590.onnx , you can simply remove the other model and take only single model's prediction as result.

The difference between two models is related to ScoringFunction (i.e. whether to calculate variance or use 1x1 convolution layer). Both functions get different features so averaging model output gives best result.

If you want to know more about the theory, I suggest you read the original author's paper since it is explained in detail there: https://ieeexplore.ieee.org/document/8954346

zhaojc001 commented 4 years ago

Thanks very much.