LightSeg: Local Spatial Perception Convolution for Real-Time Semantic Segmentation

Semantic segmentation is increasingly being applied on mobile devices due to advancements in chipsets, particularly low-power consumption scenarios. However, the lightweight design of poses limitations receptive field, which crucial for dense prediction problems. Existing approaches have attempted balance designs and high accuracy by downsampling features backbone. this may result loss local details at each network stage. To address challenge, paper presents a novel solution form compact efficient convolutional neural (CNN) real-time applications: our proposed model, spatial perception convolution (LSPConv). Furthermore, effectiveness architecture demonstrated Cityscapes dataset. The results show that model achieves an impressive between inference speed. Specifically, LightSeg, does not rely ImageNet pretraining, mIoU 76.1 speed 61 FPS validation set, utilizing RTX 2080Ti GPU with mixed precision. Additionally, it 115.7 Jetson NX int8