Optimizing 3D Semantic Scene Completion on Embedded Systems

Autonomous agents need an accurate perception of their 3D environment from sparse inputs to navigate effectively. One way to achieve this is to use semantic scene completion (SSC), which recreate the structure and meaning of the surroundings from sparse LiDAR points. However, current SSC techniques rely on computationally expensive 3D convolutions. Other works have leveraged self-attention and cross-attention as alternatives. However, they still suffer from quadratic complexity. All of these elements make it challenging to run on embedded systems without on-device optimization. For our project, we propose a simple solution targeted in optimizing LiDAR-based SSC models on embedded systems using structural pruning. This method reduces computation by identifying and removing redundant layers in the model without rewriting custom kernels. We conducted experiments on the SemanticKITTI dataset using both RTX 3090 and Jetson Xavier NX GPUs. Our results demonstrated significant performance improvements and faster inference speeds on both platforms, closing the gap between research and deployment.