Deformable convolution can handle the geometric transform appearing in the front camera image of the autonomous driving system. However, this method increases the size of the intermediate feature maps because it finds samples referenced by the kernel for each location. Moreover, the feature maps are interpolated to obtain accurate location data. Thus, it is difficult to configure modules due to increased GPU memory usage. In this paper, we propose squeezed deformable convolution robust to geometric changes while optimizing resource usage and a deformable spatial pyramid pooling (DSPP) module. The proposed convolution estimates the same data location as deformable convolution, but it uses squeeze convolution to compress the number of channels in feature processing. Therefore, memory usage and computations are decreased during sampling processing. Furthermore, memory usage can be more optimized for edge networks by removing the interpolation process using integer offsets additionally. The proposed DSPP module reinforces the scale-invariant characteristics by placing squeezed deformable convolutions in parallel and supplying inputs of different sizes. By sharing weights except for the squeeze layer, constraints are added to the sampling process, and the number of effective training samples is increased. The proposed methods are applied to the representative semantic segmentation network, ESPNet V2 and DeepLab V3+ with ResNet-50 to verify the effect. The mean intersection over union (mIoU) of DeepLab V3+ with the DSPP module is increased 1.5% on Cityscapes validation set. Also, ESPNet V2 with it provides 1.4% higher mIoU on Camvid test set. Therefore, the proposed method not only optimizes memory usage and computations but also improves recognition performance with enough margin.
Deformable Spatial Pyramid Pooling for Road Scene Semantic Segmentation
2021-09-19
1288254 byte
Conference paper
Electronic Resource
English