Dual Dense Upsampling Convolution for road scene Semantic Segmentation

There are many upsampling methods have been proposed such as interpolation, deconvolution network, and upsampling convolutional during the development of semantic segmentation model. These upsampling methods have contributed to the rapid improvement of the segmentation effects of varIoUs networks. This paper also starts with optimizing the upsampling operation and proposes a new semantic segmentation model. Firstly, we propose Dual Dense Upsampling Convolution (DDUC) to obtain pixel-level prediction and achieve a learnable full convolution network. DDUC which achieves upsampling convolution while integrating the underlying feature maps of the encoding path to utilize multi-scale features can be approximated as a Dense Upsampling Convolution (DUC) with an upsampling factor of 2 concatenated with another DUC with an upsampling factor of 4. Secondly, we introduce a simplified Channel Self-attention Mechanism(CAM-s) to model the similarity between feature channels. Traditional self-attention requires multiple high-dimensional feature transformations to obtain attention feature maps, which is computationally expensive. The simplified attention module adds global average pooling to aggregate feature information and reduce the subsequent computational burden. We train our model on the Cityscapes dataset only using fine-annotated images and achieve an mIoU of 80.7%. In particular, our model demonstrates sensitivity to small objects such as traffic signs and traffic lights in urban road scenes.