zju3dv/Scal3R

What it solves

Scal3R addresses the challenge of large-scale 3D reconstruction, specifically focusing on making the process scalable through a technique called test-time training.

How it works

The system takes a folder of images as input and performs inference to reconstruct 3D scenes. It utilizes a model configuration system and supports chunking for long-sequence inference via block and overlap sizes. To manage memory and hardware constraints during large-scale processing, it includes features to offload batches and outputs to disk.

Who it’s for

It is designed for researchers and developers working in computer vision and 3D scene reconstruction who need to process large-scale image datasets to generate camera poses, depth maps, and point clouds.

Highlights

• Scalable Reconstruction: Implements test-time training to handle large-scale 3D environments.
• Detailed Outputs: Generates predicted camera poses (camera-to-world transform matrices), depth maps, and point clouds.
• Memory Management: Includes offloading capabilities for batches and outputs to disk to support larger workloads.
• Inference Acceleration: Includes built-in support for accelerated inference.