← Back to Projects

Arena Battle Robot — Sensor Fusion & Embedded Control

2025 · University of Pennsylvania · Coursework Project

Python PyTorch 3D Vision Bundle Adjustment Multi-View Geometry LoFTR COLMAP

Gallery

cloud COLMAP training

Overview

  • Multi-View Reconstruction Pipeline: Built an end-to-end 3D reconstruction system from multi-view images, including feature extraction, matching, camera pose estimation, and point cloud generation.
  • Bundle Adjustment Optimization: Implemented bundle adjustment in PyTorch to jointly optimize camera poses and 3D structure, minimizing reprojection error across all views.
  • Learned Feature Matching: Integrated LoFTR for detector-free feature matching and compared against traditional SIFT features, improving robustness in low-texture and wide-baseline scenarios.
  • System Validation & Visualization: Developed reprojection and loss visualization tools to analyze convergence behavior and reconstruction quality.

Highlights

  • Implemented full multi-view geometry pipeline from feature matching to 3D reconstruction
  • Designed joint optimization (BA) over camera poses and 3D landmarks
  • Compared classical vs learned feature matching (SIFT vs LoFTR)
  • Achieved stable convergence with reprojection error minimization
  • Reconstructed dense and consistent 3D point clouds across multiple views