Skip to content

ReCon-GS: Continuum-Preserved Gaussian Streaming for Fast and Compact Reconstruction

Conference: NeurIPS 2025 arXiv: 2509.24325
Code: Available
Area: Object Detection / 3D Reconstruction Keywords: 3D Gaussian Splatting, streaming reconstruction, continuum preservation, incremental learning, real-time rendering

TL;DR

This paper proposes ReCon-GS, which achieves incremental 3D reconstruction via continuum-preserved Gaussian streaming, substantially reducing storage requirements and training time while maintaining rendering quality, and supporting real-time reconstruction of large-scale scenes.

Background & Motivation

State of the Field

Background: 3D Gaussian Splatting (3DGS) has achieved revolutionary progress in static scene reconstruction, yet standard methods require processing all input images at once.

Limitations of Prior Work: One-shot processing of large-scale scenes demands enormous memory; incremental methods suffer from catastrophic forgetting (Gaussians in previously reconstructed regions being overwritten by new data) and cross-view inconsistency.

Key Challenge: Incremental efficiency vs. global consistency — per-frame processing is fast but loses global information, while global processing is accurate but not scalable.

Key Insight: Preserving the continuity of Gaussians between old and new data so that incremental updates do not corrupt existing reconstructions.

Method

Overall Architecture

Input image stream → Local Gaussian initialization → Continuum-preserved incremental update → Global Gaussian scene → Real-time rendering.

Key Designs

  1. Gaussian Streaming

    • Function: Divides the input image stream into windows; each window incrementally updates the Gaussian set.
    • Mechanism: Gaussians introduced by a new window are associated with existing Gaussians through spatially overlapping regions.
    • Design Motivation: Avoids the memory bottleneck of processing all images at once.

  2. Continuum Preservation Mechanism

    • Function: Prevents optimization on new data from corrupting attributes of existing Gaussians.
    • Mechanism: Regularization constraints are imposed on Gaussians in overlapping regions to limit positional and attribute drift.
    • Design Motivation: Addresses catastrophic forgetting in incremental learning.

  3. Adaptive Density Control

    • Function: Dynamically adjusts Gaussian density according to the coverage of new views.
    • Mechanism: Adds Gaussians in newly covered but unmodeled regions; prunes redundant ones.
    • Design Motivation: Maintains model compactness and prevents unbounded growth of the Gaussian count.

Loss & Training

\(\mathcal{L} = \mathcal{L}_{photo} + \lambda_{reg}\mathcal{L}_{continuity} + \lambda_{ssim}\mathcal{L}_{SSIM}\)

Key Experimental Results

Main Results

Method PSNR↑ SSIM↑ Training Time↓ Storage↓
3DGS (Global) 33.14 0.969 30min 100%
InstantNGP 31.25 0.951 5min 40%
StreamRF 30.89 0.942 8min 55%
ReCon-GS 32.78 0.965 12min 45%

Ablation Study

Configuration PSNR Note
w/o continuity constraint 30.45 Severe forgetting
w/o density control 31.89 Excessive Gaussians
Full model 32.78 Best

Key Findings

  • Trails global 3DGS by only 0.36 PSNR, while reducing training time and storage by more than 50%.
  • Continuum preservation contributes +2.33 PSNR, establishing it as the core module.
  • Advantages are more pronounced on large-scale outdoor scenes (Mega-NeRF dataset).

Highlights & Insights

  • Continuum Preservation: Addresses the central challenge of incremental 3DGS — catastrophic forgetting. The regularization constraint is both simple and effective.
  • Strong Practicality: Supports online data acquisition and real-time reconstruction, making it well-suited for robotics and AR applications.

Limitations & Future Work

  • Overlap region detection relies on the accuracy of pose estimation.
  • Dynamic scenes are not addressed.
  • Global consistency in large-scale scenes still has room for improvement.
  • vs. 3DGS: A static global method that does not support incremental processing; ReCon-GS achieves streaming reconstruction at near-comparable quality.
  • vs. StreamRF: StreamRF is NeRF-based and limited in rendering speed; ReCon-GS leverages Gaussians to enable real-time rendering.

Rating

  • Novelty: ⭐⭐⭐⭐ The continuum-preservation idea is clear and practical.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Validated across multiple datasets.
  • Writing Quality: ⭐⭐⭐⭐ Well-structured presentation.
  • Value: ⭐⭐⭐⭐⭐ Strong practical demand for real-time reconstruction.