search

microsoft / Deep3DFaceReconstruction

Accurate 3D Face Reconstruction with Weakly-Supervised Learning: From Single Image to Image Set (CVPRW 2019)
MIT License
2.22k stars 445 forks source link

Question about the values of the losses #166

Open YokkaBear opened 3 years ago

YokkaBear commented 3 years ago

Could the values of the losses partly reflect the performance of the 3D reconstruction model? Since I added some modifications to the original network, I wonder if I could judge the improvement out of the loss values intuitively. i.e. If I observe that at the same training epoch (e.g. Iter: 60000), Deep3DFace-v2 attains lower loss value than Deep3DFace-v1, Can I assert that Deep3DFace-v2 is a better reconstructioin model than Deep3DFace-v1? Looking forward to your opinions, thanks a lot.

YuDeng commented 3 years ago

Hi, the losses in 2d image domain can reflect the 3d reconstruction performance to some extent. However, it is better to evaluate the 3d reconstruction error directly using certain datasets such as facewarehouse or MICC florence, which can give you a clearer understanding of the model performance.