NAS-SpatialFlow : neural architecture search for panoptic segmentation

• Main Author: Cao, Liu
• Other Authors: Chen Change Loy
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/138013
• Institution: Nanyang Technological University

Neural architecture search (NAS) has achieved success in various deep learning tasks. NAS can automatically find an efficient neural network architecture for a certain task on a dataset, which can outperform human-designed neural architectures. NAS has proved its efficiency in classification task (ImageNet). However, no previous work re- lated to NAS has been done on panoptic segmentation, a challenging task which unifies instance segmentation and semantic segmentation. In this final year project, we apply a gradient-based NAS method called DARTS to search for a network structure to bet- ter fuse features between sub-networks based on SpatiaFlow, a unified neural network with four sub-networks for panoptic segmentation. We name the final searched model NAS-SpatialFlow. Experimental results on the MS-COCO dataset show that NAS- SpatialFlow achieves 39.7 PQ, which is inferior to the human-designed model Spa- tialFlow (40.5 PQ). Our results demonstrate the ineffectiveness of the current gradient- based NAS methods in more complicated tasks like panoptic segmentation. The main contributions of this final year project are three-fold (1) We reimplement SpatialFlow as our baseline model and prove its effectiveness on panoptic segmenta- tion. (2) We apply NAS methods to search for network structures for feature fusion between sub-networks (3) We design attention modules to further enhance the recog- nition of the background stuff of SpatialFlow.