Reproduce “LMSCNet: Lightweight Multiscale 3D Semantic Completion” 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
Luis Roldão, Raoul de Charette, Anne Verroust-Blondet
LMSCNet, from Inria and AKKA Technologies, presents a lightweight multiscale architecture for 3D semantic scene completion from sparse LiDAR data. The method achieves competitive semantic completion (2nd in mIoU) and superior occupancy completion (1st in IoU) on SemanticKITTI, operating at 21.28 FPS at full resolution and up to 372.24 FPS at 1:8 scale with only 0.35M parameters.
Key insights
- Hybrid 2D/3D convolutional architectures can achieve efficient 3D semantic completion, significantly reducing computational load compared to purely 3D methods.
- Designing a modular multiscale output pipeline allows for flexible inference speeds, making real-time 3D semantic completion feasible for varying application requirements.
- Enhancing the UNet decoder with cross-level feature concatenation and using ASPP in 3D segmentation heads improves feature richness and contextual understanding for completion tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.