search

Juyong / DHNN_BodyRepresentation

Source code for the paper "Disentangled Human Body Embedding Based on Deep Hierarchical Neural Network" (TVCG2020).
40 stars 5 forks source link

How to achieve the progress from the Eq.(5) to the Eq.(6) in your paper #6

Open sunjc0306 opened 3 years ago

sunjc0306 commented 3 years ago

Excuse me, I want to know how to derive the Eq.(6) from the Eq.(5) in yout paper. Thanks for you very much.

jby1993 commented 3 years ago

The minimum of Eq.(5) is achieved on its stationary point, which is the zero gradient point of Eq.(5) on pi.

sunjc0306 commented 3 years ago

There is not the variable named Tj in the Eq.(5) , but it appears at the Eq.(6). In addition, I have a question what the function named batch_knn_gpu in your library. Thanks for you much.

jby1993 commented 3 years ago

The index i or j is vertex index. Tj means the transformation matrix of j'th vertex. You should treat Eq(5) of all vertices as a whole, and compute the gradient for vertices' positions. The batch_knn_gpu is used to compute k nearest points from a source pointclound to target pointclound. Except that this function can compute for multiple groups of pointclounds data in parallel.

sunjc0306 commented 3 years ago

OK,thanks. Did you release your Encodeing code. Where can I get your code about Encodeing part?

jby1993 commented 3 years ago

The architecture of the encoder has been clarified in the paper, it's just MLP(multi-layer perceptron) structure.

moustach87 commented 3 years ago

Hi, thank you for sharing this work.

I followed this conversation and got back to the paper to see the architecture of the encoder. Two questions arises:

  1. what do you mean by "shared MLP"? Are the parameters shared or the dimensions or something else?
  2. what output dimension did you chose for the first unit of T?

Thanks!

moustach87 commented 3 years ago

Okay I think I got it. It's analogue to keeping the same number of points but changing their features right? Concerning the dimension, 400 is the output of T while my second question was about the dimension of the new feature vector (output of the first layer). I really appreciate your collaboration! Thanks again!

jby1993 commented 3 years ago

Sorry, I think I made a wrong statement. In this work, I treat the 9vnum length feature as a 1-dimentional feature, and use a fully connected layer to transform the 9vnum length feature to 400 length. In this work, we do not utilize the connectivity of the mesh .