Reproduce “RoDUS: Robust Decomposition of Static and Dynamic Elements in Urban Scenes” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Thang-Anh-Quan Nguyen, Luis Roldão, Nathan Piasco, Moussab Bennehar, Dzmitry Tsishkou
RoDUS (Robust Decomposition of Static and Dynamic Elements in Urban Scenes) presents a self-supervised neural scene representation capable of cleanly disentangling static backgrounds from dynamic foregrounds in complex urban driving footage. The framework achieves state-of-the-art performance in static novel view synthesis and motion segmentation on challenging datasets like KITTI-360 and Pandaset, yielding a PSNR of 23.43 for static NVS and an IoU of 64.84 for dynamic object masks on KITTI-360.
Key insights
- A robust initialization phase for the static NeRF branch, specifically rejecting dynamic elements as outliers, is critical for establishing a clean background prior to full dynamic scene learning.
- Guiding the dynamic branch with 2D semantic information and selectively masking static background classes prevents entanglement, ensuring the dynamic model focuses solely on movable objects.
- Explicitly modeling dynamic object shadows as transient elements, along with applying density regularization to sky and road, addresses specific challenges of urban scene geometry and dynamics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.