Core Concepts
BlindNet proposes a novel approach for semantic segmentation, utilizing covariance alignment and semantic consistency contrastive learning to address style variations and enhance generalization capabilities.
Abstract
BlindNet introduces a novel method for semantic segmentation, focusing on addressing style variations that affect model performance in unseen domains. The approach includes covariance alignment to handle style variations in the encoder and semantic consistency contrastive learning in the decoder. By aligning covariance matrices and disentangling features vulnerable to misclassification, BlindNet outperforms existing methods in robustness and performance across various datasets.
Stats
DeepLabV3+ with ResNet50 backbone: 40K iterations, batch size of 8, SGD optimizer with momentum of 0.9 and weight decay of 5e-4.
Weighting parameters: ω1=0.2, ω2=0.2, ω3=0.3, ω4=0.3.
Synthetic datasets: GTAV (G) and SYNTHIA (S), Real-world datasets: Cityscapes (C), BDD-100K (B), Mapillary (M).
Quotes
"Our proposed BlindNet consists of two components: covariance alignment and semantic consistency contrastive learning."
"To achieve consistent feature representation in DGSS across various styles, we introduce semantic consistency contrastive learning."
"Our method operates comparably to baseline models by learning features intrinsically without adopting a separate module."