SasWOT: Real-Time Semantic Segmentation Architecture Search WithOut Training
Published in AAAI, 2024
SasWOT introduces a training-free approach to semantic segmentation architecture search by automatically discovering effective zero-cost proxies. By optimizing proxy functions through evolutionary search, it eliminates expensive training during architecture exploration, significantly accelerating the search process while maintaining strong accuracy–efficiency trade-offs.
Recommended citation: Zhu, C., Li, L., Wu, Y., & Sun, Z. (2024, March). Saswot: Real-time semantic segmentation architecture search without training. In Proceedings of the AAAI conference on artificial intelligence (Vol. 38, No. 7, pp. 7722-7730). http://chendi23.github.io/files/SasWOT.pdf