ProDA: Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation

[JisuHann]

ProDA: Prototypical Pseudo Label Denoising and Target Structure Learning for Domain Adaptive Semantic Segmentation

Problem of self-training

1. Pseudo labels are noisy
2. The target features are dispersed due to the discrepancy between source and target domains.

   Idea

   • focus on Representative prototypes, the feature centroids of classes
   • Exploit the feature distances from prototypes that provide richer information than mere prototypes
   • Estimate the likelihood of pseudo labels to facilitate online correction in the course of training
   • rectify the pseudo labels by estimating the class-wise likelihoods according to its relative feature dis- tances to all class prototypes
   • propose to align soft pro- totypical assignments for different views of the same target, which produces a more compact target feature space.
   • Distilling the already learned knowledge to a self-supervised pertained model further boosts the performance