PixMatch: Unsupervised Domain Adaptation via Pixelwise Consistency Training

Unsupervised domain adaptation is a promising technique for semantic\nsegmentation and other computer vision tasks for which large-scale data\nannotation is costly and time-consuming. In semantic segmentation, it is\nattractive to train models on annotated images from a simulated (source) domain\nand deploy them on real (target) domains. In this work, we present a novel\nframework for unsupervised domain adaptation based on the notion of\ntarget-domain consistency training. Intuitively, our work is based on the idea\nthat in order to perform well on the target domain, a model's output should be\nconsistent with respect to small perturbations of inputs in the target domain.\nSpecifically, we introduce a new loss term to enforce pixelwise consistency\nbetween the model's predictions on a target image and a perturbed version of\nthe same image. In comparison to popular adversarial adaptation methods, our\napproach is simpler, easier to implement, and more memory-efficient during\ntraining. Experiments and extensive ablation studies demonstrate that our\nsimple approach achieves remarkably strong results on two challenging\nsynthetic-to-real benchmarks, GTA5-to-Cityscapes and SYNTHIA-to-Cityscapes.\n Code is available at: https://github.com/lukemelas/pixmatch\n