Results comparison

[longsc2603]

Hi, I am currently trying to modify the structure of FSANet for a scientific research project in my uni. I wonder if you have any idea as to why your result (average MAE) is somewhat different to the original result? I try to train the model using your ipynb file and got protocol 2's average MAE of about 4.608, which is almost 1 larger than the original average MAE (3.60) as presented in Table 3 in the paper.

The new structure of FSANet that I am trying to use is potential for achieving a lower average MAE than yours, but as low as in the paper is really hard. Can you please give some insights? Big thanks!!!

[omasaht]

Hey Long, this is most likely due to a bug in model implementation. I didn't update the repository as I have been quite busy with other projects but you can find more detail and fix on this in this issue: #4 (comment)

The code replacement is applied here: https://github.com/omasaht/headpose-fsanet-pytorch/blob/master/src/model.py#L235

[longsc2603]

Thanks for replying, but I am not quite clear about what you say yet. In the issue #4, the posed problem is not very relevant to why your MAE result is much higher than the result in the original repo. If I'm wrong and it is, the code replacement you mentioned is what I already used in the first place @@
If you could give a more clear explanation, that would be a big help. But anyways, thanks again for replying.