MRF-UNets: Searching UNet with Markov Random Fields

UNet [27] is widely used in semantic segmentation due to its simplicity and effectiveness. However, manually-designed architecture applied a large number of problem settings, either with no optimizations, or manual tuning, which time consuming can be sub-optimal. In this work, firstly, we propose Markov Random Field Neural Architecture Search (MRF-NAS) that extends improves the recent Adaptive Optimal Network Width (AOWS) method [4] (i) more general MRF framework (ii) diverse M-best loopy inference (iii) differentiable parameter learning. This provides necessary NAS efficiently explore network architectures induce graphs, including loops arise from skip connections. With as backbone, find an architecture, MRF-UNet, shows several interesting characteristics. Secondly, through lens these characteristics, identify sub-optimality original further improve our results MRF-UNetV2. Experiments show MRF-UNets significantly outperform benchmarks on three aerial image datasets two medical while maintaining low computational costs. The code available at: https://github.com/zifuwanggg/MRF-UNets .