Reconstructing 3D scenes from long videos is important for robotics and embodied AI, but existing feed-forward Gaussian methods struggle when extended beyond a few images. Processing every view jointly blows up GPU memory, while per-frame predictions pile up redundant Gaussians that represent the same surfaces, inflating storage and rendering costs.
LonGaussian, introduced by researchers at Huazhong University of Science and Technology, takes a streaming approach. It builds on a pretrained streaming 3D reconstruction backbone and decodes long-context geometric features into dense Gaussians chunk by chunk. Within each chunk, multi-frame point tracks identify repeated observations of the same surface and feed a learned merger that consolidates the corresponding Gaussian predictions. A rendered-depth consistency filter then removes Gaussians from other chunks that would incorrectly occlude locally reconstructed surfaces.
The authors evaluate LonGaussian on the DL3DV and RealEstate10K benchmarks, reporting improved novel-view rendering quality while using substantially fewer Gaussians than the feed-forward baselines they compared against. The method assumes known camera intrinsics but unknown poses, and the paper is available on arXiv as preprint 2610.23250.