confusing about the shape of B?

[CuddleSabe]

B = PCA(Ck, p) = R^{mxp}， but the shape of C_bar is R^d, how to compute it?

[CuddleSabe]

maybe the paper is wrong.... paper say u use the pca on the second dim, but the code is the first dim!

[ygtxr1997]

PCA aims to reduce the dimension of a matrix in shape $m\times n$ into $m\times k$ in the 2nd dimension. However, in 3DMM and our method, we DO NOT directly use this reduced $m\times k$ matrix, but the itermediate result P in the shape of $k \times m$. Here P can be considered as a mapping matrix, that is, the relationship between the unreduced dimensions (768) and the reduced dimensions (maybe 512). You may refer to the 3DMM paper to find out more details. Our paper's B is mainly based on the P mapping matrix but not the reduced input embeddings.

[CuddleSabe]

PCA aims to reduce the dimension of a matrix in shape m × n into m × k in the 2nd dimension. However, in 3DMM and our method, we DO NOT directly use this reduced m × k matrix, but the itermediate result P in the shape of k × m . Here P can be considered as a mapping matrix, that is, the relationship between the unreduced dimensions (768) and the reduced dimensions (maybe 512). You may refer to the 3DMM paper to find out more details. Our paper's B is mainly based on the P mapping matrix but not the reduced input embeddings.

thanks! my fault ;)

[ygtxr1997]

🤗