Location: Ann Arbor, MI | 01/2026 - 04/2026

Github: https://github.com/sacchinbhg/Gsplat-Object-Reconstruction

 

This project presents an extension to the Gaussian Object framework for object-centric 3D reconstruction under uncertain camera poses. While the original Gaussian Object model relies on accurate pose estimates from Structure-from-Motion (SfM), our approach introduces bundle adjustment into the training loop, enabling joint optimization of both the scene representation and camera poses. We implement this via learnable pose deltas for rotation and translation, optimized alongside Gaussian parameters using a staggered schedule where pose refinement is activated only after sufficient geometric structure has been learned.

Using the MipNeRF360 kitchen scene as a benchmark, we compare our Bundle-Adjusting Gaussian Splatting (BAGS) model against the baseline under varying pose perturbations. Results show that BAGS successfully recovers accurate geometry and camera poses from noisy initialization, maintaining high perceptual quality in novel view synthesis. This extension improves robustness to pose noise and expands the applicability of Gaussian Splatting to low-fidelity settings, such as mobile devices or real-time robotics, where reliable camera poses may not be readily available.