LonGaussian: Tracking Foundation Priors for Long-Context Feedforward Gaussian Reconstruction

Key points

• Reconstructing high-quality, renderable 3D scenes from long videos is essential for robotic perception and embodied intelligence.1
• Many feed-forward Gaussian reconstruction methods rapidly predict 3D scenes from a small set of images, but extending them to longer sequences presents two bottlenecks: jointly processing all views increases GPU memory requirements, while accumulating pixel-aligned predictions redundantly represents the same surfaces and increases storage and rendering costs.1
• Built on a pretrained streaming 3D reconstruction backbone, LonGaussian decodes long-context geometric features into dense Gaussians.1
• Rendered-depth consistency filtering further suppresses Gaussians from other chunks that incorrectly occlude locally reconstructed surfaces.1

Sources (1)

• [1]LonGaussian: Tracking Foundation Priors for Long-Context Feedforward Gaussian Reconstruction
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 10:24 AM

  We present LonGaussian, a streaming feed-forward framework that reconstructs compact, globally renderable Gaussian scenes from long image sequences with known intrinsics and unknown camera poses.

  Reconstructing high-quality, renderable 3D scenes from long videos is essential for robotic perception and embodied intelligence.

Extractive summary: sentences quoted from the sources.