search

mks0601 / I2L-MeshNet_RELEASE

Official PyTorch implementation of "I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image", ECCV 2020
MIT License
710 stars 130 forks source link

about my own dataset #38

Closed lll-gen closed 3 years ago

lll-gen commented 3 years ago

Greatest job! I would like to ask why the SMPL parameter is needed in the training process. The loss function in the paper does not use the SMPL parameter. If I use my own data set, what should I do

mks0601 commented 3 years ago

The mesh vertices are obtained by forwarding SMPL parameters to SMPL layers. For a new dataset, all you have to do is just making another data/$DB_NAME/$DB_NAME.py by refering another data/$DB_NAME/$DB_NAME.py

lll-gen commented 3 years ago

thanks for you applyment! there is another question,How is joint_cam obtained and what does it do

pablovela5620 commented 3 years ago

joint_cam would be the camera coordinate system 3d joints given from whatever dataset you're using (Human3.6m, PW3d, etc..) so they're the ground truth labels. They can also be obtained via the forward pass of the network

lll-gen commented 3 years ago

hello,I want to ask why the camera parameters don't have R and T?In other word ,it's only in the camera coordinate not in the world coordinate?

mks0601 commented 3 years ago

Could you let me know which dataset are you talking about?

lll-gen commented 3 years ago

firsr of all,thanks for your patient response!I have 3 question: Question1 :oh at present ,I use my own dataset ,it only contains SMPL parameters,rendered imgs,camera parameters.does it work？ Question2: But when I download your MUCO datasets,I found the camera parameters only contains focal and principal,it doesn't need Rotation and translation matrices to Transform to the World coordinate? Question3: Farther, Can I abandon the SMPL expression?in other words,I obtain the coordinate by non-rigid Registration,not use the smpl parameters?

mks0601 commented 3 years ago

Q1. Yes. It may work. Q2. The 3D coordinates and SMPL parameters of the MuCo dataset is camera-centered ones. Q3. I can't get your question :(

lll-gen commented 3 years ago

Q3:I mean, I don't get the Mesh coordinates from the SMPL parameters,I get the vertex coordinates directly the other way around

mks0601 commented 3 years ago

Q3. That would be no problem.

lll-gen commented 3 years ago

Hi，I'm back again！ Do I need to rerun the ROOTNET code to get the root node depth to obtain the final mesh?

mks0601 commented 3 years ago

There are two kinds of meshes: 1) lixel-based and 2) param-based as you can see in here. For the 1) lixel-based mesh, yes, you need to add the depth from the RootNet. for the 2) param-based mesh, no, you don't need to add the depth from the RootNet.