[Question] From a single 2d image to 3d mesh (with 3d supervision, but without differentiable rendering?) #463
Labels
call-for-collaboration
User seeks collaborators for their project
how to
How to use PyTorch3D in my project
Dear PyTorch3D team,
Thank you for the development of this wonderful library and your help on setting up this environment. I am a graduate student in Architecture and Urban Design from Politecnico di Milano. I am passionate on applying a mature AI framework in the AEC industry and currently doing research on 3d reconstruction of a 3d printing pre-tension textile. I have a strong feeling that PyTorch3d could be the best library to do this.
I am wondering if you could give me some feedback or ideas on the following description?😎
TASK:
3D Reconstruction, from a single 2d image to 3d mesh
DATASET
I have a dataset of M objects paired with(INPUT: single image, OUTPUT: mesh model, .obj). I enclose a screenshot for a peek. Left is input, right is 3d-view of output.
🔥🔥🔥❗❗BOTTLENECK/DIFFERENCE FROM MeshRCNN or other 3d reconstruction jobs:
From the above image, you might notice there is no 'ground-truth' view on the mesh. Since the objective is to predict the fabric after tension release, the shape/mesh changed from the initial stage. I enclose a .gif to illustrate the deformation.
You can also take a look at the following screenshot. The left is the PLA printed on the pre-tension textile while the right is after cropping from the fabric(tension released).
CONCLUSION
That said, the INPUT is the top-view of the fabric BEFORE deformation. The OUTPUT is a 3d mesh AFTER deformation(tension released), this is run by a Physical simulation(FEA-finite element analysis) and each INPUT is affected by a SAME Force Field.
MY THOUGHT
Since I am not from the field of Computer Vision, correct me if I am wrong😅.
If I understood the paper correctly, the MeshRCNN somehow considers the projection in the framework. For example, the Rol Align refers to the rotation perspective between image and 3d models. The Vert Align refers the vertex between image and 3d models. But in my case, it is a bit difficult to do such alignment since the deformation. Is it possible to decode the image and calculate the Voxel Loss only?
I don't know if I make the description compact. Looking forward to your favorable reply and feedback.
Sincerely thanks to the PyTorch3d team, Amitav Baruah, Steve Branson, Luya Gao, Georgia Gkioxari, Taylor Gordon, Justin Johnson, Patrick Labtut, Christoph Lassner, Wan-Yen Lo, David Novotny, Nikhila Ravi, Jeremy Reizenstein, Dave Schnizlein, Roman Shapovalov, Olivia Wiles.
Xingxin
The text was updated successfully, but these errors were encountered: