TTT3R: 3D Reconstruction as Test-Time Training
This work addresses the degradation in performance for 3D reconstruction models when applied beyond training context lengths, offering a training-free solution for improved generalization.
The paper tackled the problem of limited length generalization in recurrent neural networks for 3D reconstruction by framing it as a test-time training problem, resulting in a 2x improvement in global pose estimation over baselines while operating at 20 FPS with 6 GB GPU memory.
Modern Recurrent Neural Networks have become a competitive architecture for 3D reconstruction due to their linear-time complexity. However, their performance degrades significantly when applied beyond the training context length, revealing limited length generalization. In this work, we revisit the 3D reconstruction foundation models from a Test-Time Training perspective, framing their designs as an online learning problem. Building on this perspective, we leverage the alignment confidence between the memory state and incoming observations to derive a closed-form learning rate for memory updates, to balance between retaining historical information and adapting to new observations. This training-free intervention, termed TTT3R, substantially improves length generalization, achieving a $2\times$ improvement in global pose estimation over baselines, while operating at 20 FPS with just 6 GB of GPU memory to process thousands of images. Code available in https://rover-xingyu.github.io/TTT3R