search

microsoft / Deep3DFaceReconstruction

Accurate 3D Face Reconstruction with Weakly-Supervised Learning: From Single Image to Image Set (CVPRW 2019)
MIT License
2.18k stars 443 forks source link

Question about the scale of the training dataset #160

Closed YokkaBear closed 3 years ago

YokkaBear commented 3 years ago

Hi @YuDeng , thank you for your excellent work and open-source code. I intend to train R-Net from scratch using a new dataset composed of ~200k images, and currently I am under the data preprocessing step. Noticing that the script preprocessing_img.py takes a relatively long time to run over my dataset, I wonder if decreasing the scale of the training dataset, e.g. down to ~50k, would work and achieve an acceptable experimental result or not. Looking forward to your reply, thanks a lot.

YuDeng commented 3 years ago

Hi, for image pre-processing, most of the time is cost by data IO. It is possible to reduce its time if it is combined into the dataloader part to avoid saving processed images to the hard disk.

We have tried to use dataset with around 70K images to train the model, e.g FFHQ. The performance of the model may drop slightly, but it still give reasonable result.

YokkaBear commented 3 years ago

Clearer sight now, thank you for your quick response.