Long-LRM: Long-sequence Large Reconstruction Model for Wide-coverage Gaussian Splats

2024-10-16

arXiv.org

AI Summary

To address the challenges of long-sequence modeling and slow inference in 360° high-resolution 3D Gaussian reconstruction for large-scale scenes, this paper proposes the first end-to-end feed-forward reconstruction framework integrating Mamba2 and Transformer architectures. Our method enhances sequence processing capacity via dynamic token merging and adaptive Gaussian density pruning, and achieves single-pass reconstruction by coupling with differentiable 3D Gaussian rendering. On 32-frame 960×540 spherical inputs, it completes full-scene reconstruction in just 1.3 seconds on an A100 GPU—two orders of magnitude faster than state-of-the-art optimization-based methods. Quantitative and qualitative evaluations on DL3DV-140 and Tanks and Temples demonstrate reconstruction quality on par with SOTA optimization approaches. To our knowledge, this is the first work to enable high-fidelity, real-time, large-scale 3D Gaussian splatting reconstruction.

Technology Category

• Computer Vision: 3D Computer Vision
• Machine Learning: Large Multimodal Models (LMMs)
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

We propose Long-LRM, a generalizable 3D Gaussian reconstruction model that is capable of reconstructing a large scene from a long sequence of input images. Specifically, our model can process 32 source images at 960x540 resolution within only 1.3 seconds on a single A100 80G GPU. Our architecture features a mixture of the recent Mamba2 blocks and the classical transformer blocks which allowed many more tokens to be processed than prior work, enhanced by efficient token merging and Gaussian pruning steps that balance between quality and efficiency. Unlike previous feed-forward models that are limited to processing 1~4 input images and can only reconstruct a small portion of a large scene, Long-LRM reconstructs the entire scene in a single feed-forward step. On large-scale scene datasets such as DL3DV-140 and Tanks and Temples, our method achieves performance comparable to optimization-based approaches while being two orders of magnitude more efficient. Project page: https://arthurhero.github.io/projects/llrm

Problem

• Instant high-resolution 360° 3D Gaussian reconstruction
• Handling long sequences of 250K tokens efficiently
• Achieving 800x speedup over optimization-based methods

Innovation

• Mamba2 and transformer blocks mixture
• Light-weight token merging module
• Gaussian pruning for quality efficiency