Neural Field Conditioning Strategies for 2D Semantic Segmentation

Abstract Neural fields are neural networks which map coordinates to a desired signal. When field should jointly model multiple signals, and not memorize only one, it needs be conditioned on latent code describes the signal at hand. Despite being an important aspect, there has been little research conditioning strategies for fields. In this work, we explore use of as decoders 2D semantic segmentation. For task, compare three methods, simple concatenation code, Feature-wise Linear Modulation (FiLM), Cross-Attention, in conjunction with codes either describe full image or local region image. Our results show considerable difference performance between examined strategies. Furthermore, that via Cross-Attention achieves best is competitive CNN-based decoder