Learning for 3D Vision
● Trained an encoder-decoder architecture for regressing voxels, point clouds and mesh representation from a single view RGB input.
● Implemented Nerf from scratch with Differentiable Renderer for emission-absorption volumes, uniform ray sampler and view dependent MLP with position encoding.
● Created a code base for sphere tracing for rendering Sign Distance Function and trained an ML architecture for neural SDF on point cloud data.
Learning Outcomes:
- 3D Representations.
- Single-view 3D: Objects.
- Articulated Models - SMPL, SMPLX, Hands and animals.
- Neural Radiance Fields (Nerf) and Volumetric Rendering.
- Neural Surface Rendering.
- Pointcloud - Classification, detection and Segmentation.
Programming Language: Pytorch, Pytorch3D, Python