Jianzhu Guo
Jianzhu Guo
I have found a similar issue: https://github.com/eragonruan/text-detection-ctpn/issues/380. BTW, good luck : )
If this fix works across platforms, welcome for your PR : )
Faster on macOS (intel).
**The texture is sampled from the original image**, so invisible (self-occluded) regions will be filled with pixels where the vertices are projected. It is not the bug, or the projector...
FaceBoxes may not recall that face. I suggest you may use a stronger face detector to recall the right profile face.
The benefit of meta-joint optimization is limited, more augmentation produces better results. If possible, using externel data (the paper only uses 300W-LP) may promote the generalization ability. @zcy8123878
Sorry for the delayed reply. I think it may be caused by the version of `scipy`. You can check the md5 of BFM_UV.mat: d58359aa4678abc576fe77c87feb1dbe
Any face detector will be fine with small modifications of the bounding box scale and offset.
The transformation of the face. 3DDFA_V2 adopts the scale orthographic projection.
@abcliguanxi The same.