Dynamic scene reconstruction is critical for robotic perception, where robots must understand changing environments in a temporally consistent way. Deformable 3D Gaussian Splatting (3DGS) is a recent technique that models dynamic scenes, but it faces challenges when only monocular input is available. A new paper on arXiv proposes DispFlow-GS, a method that addresses these challenges by adding displacement flow supervision and motion disentangling.
The core idea is to supervise the displacement of 3D Gaussians using optical flow, which helps maintain consistency across frames. Additionally, the method disentangles motion into separate components, allowing the model to better handle complex scenes with multiple moving objects or different types of motion. This is particularly useful for monocular setups, where depth and motion cues are ambiguous.
While the abstract does not provide quantitative results, the approach suggests a promising direction for improving the accuracy and robustness of dynamic scene reconstruction. By explicitly modeling flow and separating motion, DispFlow-GS could enable more reliable perception for robots operating in dynamic environments. The paper is available on arXiv under the identifier 2609.36940.