advances state-of-the-art accuracy for latency-optimized networks on the Cityscapes [179] semantic segmentation dataset via a search space similar to that of MobileNet. An overview of the SqueezeNAS architecture search path is shown in Figure 13. BiX-NAS [175] BiX-NAS is based on a multiscale upgrade of a bidirectional skip-connected network (see Figure 14). It uses a twophase search algorithm, a differentiable search in Phase 1 and an evolutionary search in Phase 2. It reduces network computational costs by sifting out ineffective multiscale features at different levels and iterations. MRF-UNets [177] MRF-UNets extends and improves the recent adaptive and optimal network width search [180] method with a more general MRF framework, a diverse M-best loopy inference [181], and a differentiable parameter learning. This provides the necessary NAS framework to efficiently explore the architecture search space of a U-Net backbone (see Figure 15). 3 × 3 Con Input: ARM ARM Context Info Stage 1 and 2 Stage 3 Stage 4 Stage 5 FFM 8× Output: (a) Spatial Info Detail Loss :Detail GT BN Relu Detail Head 1 × 1 Con 2× 1 × 1 Con 4× -1 -1 -1 Seg Head Seg Loss (b) (c) FIGURE 11. An architectural overview of the STDC network. (a) Network architecture. (b) Train loss. (c) Detail ground-truth generation. (d) Segmentation/detail head. (e) Laplacian kernel. The dashed blue box indicates the detail aggregation module. The dashed red box is the network proposed in STDC. ARM: attention refine module; BN: batch normalization; FFM: feature fusion module; Seg: segmentation; ReLU: rectified linear unit. (Adapted from [152].) floc(∗) Stride = 1 Stride = 2Stride = 4 Laplacian Con :Seg GT 2× 4× 8× (d) -1 8 -1 (e) -1 -1 -1 : 2× Upsample : 4× Upsample : 8× Upsample : Train Only : Train and Test fglo(∗) fjoi(∗) fsur(∗) FIGURE 12. An architectural overview of the context-guided block in CGNet. The structure consists of a local feature extractor f (),) a surrounding context extractor f (), loc sur ) a joint feature extractor f (),) and a global context extractor ().) Concat: concatenate; joi FC: fully-connected layer; GAP: global average pooling. (Adapted from [173].) DECEMBER 2023 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE 107 fglo 1 × 1 Conv 3 × 3 DConv 3 × 3 Conv Concat BN + PReLU GAP FC FC