$ show QitaoZhao/E-RayZer
E-RayZer is a self-supervised 3D reconstruction system that takes unordered multi-view images as input and outputs estimated camera poses, target-view renders, Gaussian point clouds, and turntable videos. The system uses a transformer-based architecture to estimate camera poses and scene representations, then renders novel views using differentiable Gaussian splatting. It includes a Gradio demo interface for easy inference, along with training and evaluation scripts for the DL3DV dataset. The model is trained without ground truth camera poses or depth information, learning to reconstruct 3D scenes purely from multi-view image consistency.
The LLM identified 3 files worth highlighting for reuse:
Provides self-contained functions for computing camera pose errors including rotation angle, translation angle, and Relative Pose Accuracy (RPA) at various thresholds. Only depends on torch and numpy, making it easily extractable for any 3D vision project needing pose evaluation.
Contains reusable camera transformation utilities including quaternion conversions, pose interpolation for smooth trajectories, focus of attention computation, and pose normalization. Functions like get_interpolated_poses_many and focus_of_attention demonstrate non-obvious skill in 3D geometry processing and could be extracted for use in other projects.
Implements VGG19-based perceptual loss used in the training setup. Self-contained with clear weight loading logic from MatConvNet format. Only depends on PyTorch and SciPy, making it easily reusable in other projects needing perceptual loss computation.