LEGO-SCD learns to detect changes in 3D scenes at 17.2 FPS without manual labeling

A team of Changrui Zhu and Simon Julier has released LEGO-SCD, an online scene change detection method. It trains without manual change annotations, using automatic edits to a reference 3D Gaussian Splatting model. On the PASLCD benchmark, the method reaches 67.64% mIoU, and on CCS3D it reaches 75.82%. The full pipeline processes query sequences on PASLCD at 17.2 frames per second, faster than all the online methods the authors compared it against.

The task: figuring out what changed while new frames arrive

Online change detection works like this. The system has a reference 3D scene, such as a room, a warehouse, or a factory floor. A camera moves through it and sends new images. The system has to determine on the fly which regions differ from the reference: whether a box has appeared, a chair has been removed, or equipment has been moved.

A single frame is not enough for this. Lighting, viewpoint, and rendering artifacts create false discrepancies, so confidence in a change has to build up as observations arrive. The authors do this through generalized Bayesian evidence fusion. That raises two problems. Neither the prior probability of change nor the evidence models are known in advance. And numerical Bayesian inference is expensive, which is critical in an online setting.

How to train a detector without a single labeled edit

The authors found a shortcut in the 3DGS scene itself. The reference model is edited automatically, and each edit yields training views with known labels. The model knows where and what it changed, so no annotators are needed. Unchanged observations of the reference are added to these views. Together they provide a supervised signal for three things:

• scene-specific evidence models;
• an efficient pixel-level change prior;
• response statistics, that is, how the signal behaves on changed and unchanged regions.

Online, the system combines the current evidence with the accumulated history. Expensive numerical inference is replaced with local Bayes factors in closed form. Their parameters come from the learned response statistics. The result updates the accumulated 3D Gaussian change field. In the final step, object-aware refinement turns this field into the final change masks.

The numbers and what's in the release

• Method: LEGO-SCD, online change detection built on 3D Gaussian Splatting.
• Authors: Changrui Zhu and Simon Julier, arXiv preprint 2610.28490.
• PASLCD: 67.64% mIoU, the best result according to the authors.
• CCS3D: 75.82% mIoU, also the best result.
• Speed: 17.2 FPS on PASLCD for the full pipeline on entire query sequences, the fastest result among the online methods tested.
• Labeling: manual change annotations are not used.
• Availability: code on GitHub, a project page, and learned scene parameters on Hugging Face in the Teleeeeeeeeeee/LEGO-SCD repository.

An important detail is that the 17.2 FPS figure was measured on the full pipeline, including mask refinement, and not just on the Bayesian fusion core. Speed was measured on entire query sequences, not on selected best-case frames.

Where this is useful and what to check

The most obvious use cases involve robots and inspection. A mobile robot benefits from knowing that something has changed in a familiar space. Facility monitoring, warehouses, and construction sites need the same. 3DGS reconstructions are becoming more common, and a change detector that builds on top of them fits naturally into existing pipelines.

The approach also has a limitation, built into the phrase "scene-specific models." The statistics have to be learned for each scene, which means an offline stage on the reference reconstruction is required before launch. For static sites that are surveyed regularly, this is an acceptable cost. For constantly changing spaces, it is higher.

In practice, a team that wants to try LEGO-SCD should do three things. First, take the published scene parameters and run the code on PASLCD or CCS3D to reproduce 17.2 FPS on its own hardware. Second, check how closely the synthetic edits to the reference resemble real changes in its own data, since that determines how well quality transfers. Third, weigh the cost of offline training for each new scene against the benefit of dropping manual labeling.