Closed xuanzhezhao closed 3 years ago
Hi @xuanzhezhao, thanks for your interest in our work!
In the case where you just want to test with our pre-trained weights, you don't need 'segm' or 'trans'. If you check out the 'Usage' part in the readme, you just need to pass in a (1, K, N, 3) tensor where K is the number of point clouds and N is the number of points in each point cloud.
If you want to train using your own data, then 'trans' should be a dictionary containing the following keys: 'cam'
, 1
, 2
, ..., NUM_PARTS
. Each key should be a list of size K, so that trans[i][k]
is the 4x4 matrix denoting the transformation of part i
(or the camera) in view k
. You can also checkout these lines for the actual usage of the transformation matrices.
Best wishes, Jiahui
Thank you very much, It makes sense to me.
Hi, thanks for your work. I notice that the dataset you provided(like data/000000.npz) contains three parts: pc, segm and trans. Could you please explain more details about the generation of "trans"? And how to generate my custom dataset if I want to test multibody-sync on it? Thanks a lot!
Best, Xuanzhe