Puzzles: Unbounded Video-Depth Augmentation for Scalable End-to-End 3D Reconstruction

TL;DR

Puzzles is a novel data augmentation technique that synthesizes diverse, high-quality video-depth data from limited sources, significantly improving 3D reconstruction accuracy without changing existing models.

Contribution

Introduces Puzzles, a data augmentation method that generates unbounded, realistic video-depth data from minimal input to enhance 3D reconstruction performance.

Findings

Boosts reconstruction accuracy across various pipelines

Achieves comparable results with only 10% of original data

Enhances data diversity without modifying network architecture

Abstract

Multi-view 3D reconstruction remains a core challenge in computer vision. Recent methods, such as DUST3R and its successors, directly regress pointmaps from image pairs without relying on known scene geometry or camera parameters. However, the performance of these models is constrained by the diversity and scale of available training data. In this work, we introduce Puzzles, a data augmentation strategy that synthesizes an unbounded volume of high-quality posed video-depth data from a single image or video clip. By simulating diverse camera trajectories and realistic scene geometry through targeted image transformations, Puzzles significantly enhances data variety. Extensive experiments show that integrating Puzzles into existing video-based 3D reconstruction pipelines consistently boosts performance without modifying the underlying network architecture. Notably, models trained on only ten percent of the original data augmented with Puzzles still achieve accuracy comparable to those trained on the full dataset. Code is available at https://jiahao-ma.github.io/puzzles/.