SSS3D: Fast Neural Architecture Search For Efficient Three-Dimensional Semantic Segmentation
This work addresses the need for efficient neural architecture search in 3D semantic segmentation, which is incremental as it builds on existing methods like RandLA-Net.
The paper tackles the problem of finding computationally efficient 3D semantic segmentation networks by introducing SSS3D, a fast multi-objective neural architecture search framework that reduces search time by 99.67% for single-stage searches and by 54% in a two-stage approach.
We present SSS3D, a fast multi-objective NAS framework designed to find computationally efficient 3D semantic scene segmentation networks. It uses RandLA-Net, an off-the-shelf point-based network, as a super-network to enable weight sharing and reduce search time by 99.67% for single-stage searches. SSS3D has a complex search space composed of sampling and architectural parameters that can form 2.88 * 10^17 possible networks. To further reduce search time, SSS3D splits the complete search space and introduces a two-stage search that finds optimal subnetworks in 54% of the time required by single-stage searches.