This repository contains the code for our project for the course 'Machine Learning for 3D Geometry' at TUM [IN2392]. In this work, we had improved the performance of the model in terms of Intersection over Union (IoU) compared to the baseline Occupancy Networks - Learning 3D Reconstruction in Function Space. Additionally, we had reduced the number of parameters to make the training of the model more efficient.
Authors
- Yujun Lin
- Joong-Won Seo
- Yunan Li
For our experiments, we use Pix3D as our new dataset which contains images, masks, meshes and camera positions.
- Replace the backbone ResNet with ConvNeXt
- Integrate feature pyramid to enable multi-scale inputs
- Incorporate camera pose information
To see the experiment results, please check the final paper [ML3D_Report.pdf] in the repository: