BiSeNet with Depthwise Attention Spatial Path for Semantic Segmentation

Abstract

This paper proposes a new structure to obtain similar results while reducing the computational amount of BiSeNet for Real-Time Semantic Segmentation. Among the Spatial Path and Context Path of BiSeNet, the study was conducted focusing on the large size kernel of the Spatial Path. Spatial Path has rich spatial information by creating a feature map 1/8 times the size of the original image through three convolution operations. The convolution operation used at this time is performed in the order of 7×7, 3×3, and 3×3. When a general convolution is used for a kernel of such a large size, the calculated cost increases due to a large number of parameters. To solve this problem, this paper uses Depthwise Separable Convolution. At this time, in Depthwise Separable Convolution, loss occurs in Spatial Information. To solve this information loss, an attention mechanism [1] was applied by elementwise summing between the input and output feature maps of depthwise separable convolution. To solve the dimensional difference between input and output, PPM: Pooling Pointwise Module is used. PPM uses Maxpooling to change the Spatial Dimension of input features and Channel Dimension through Pointwise Convolution (lx1 Convolution) [2]. This paper propose to use Depthwise Attention Spatial Path for BiSeNet using these methods. Through our proposed methods, mIoU in SS, SSC, MSF, and MSCF were 72.7%, 74.1 %, 74.3%, and 76.1 %. Proposed network can segment the part that the original one can't when using our Depthwise Attention Spatial Path.