서브-모달리티 주의 기반 효율적인 의미론적 분할 기술 = Efficient Semantic Segmentation Based on Sub-Modality Attention

다국어 초록 (Multilingual Abstract)

Semantic segmentation plays a crucial role in autonomous driving by assigning pixel-wise labels to images. Traditional convolutional neural networks (CNNs) based semantic segmentation approaches incorporate conventional down-sampling layers in the initial stage to enhance computational efficiency. The feature maps in the initial layers of a CNN are more effective when individual channels capture diverse and complementary information. In this work, we introduce a sub-modality attention network that explicitly separates high-frequency and low frequency component, and attention is employed to integrate separated pieces of information, allowing them to complement each other's deficiencies at the early feature extraction stage. Our results demonstrate that deepening CNNs is not the only path to performance improvement—incorporating handcrafted priors, such as Wavelet transform, can also yield significant gains.