DEDUCE-Net: Efficient Semantic Segmentation on Embedded Systems for Autonomous Driving

Semantic segmentation for autonomous driving is a time-critical application and thus requires an efficient solution. Consequently, designing segmentation models with low computational complexity and low power consumption is crucial for their practical deployment in resource-constrained edge devices. To achieve this, we present an ultra-lightweight network that strikes a decent balance between accuracy and efficiency. Specifically, we introduce the Details-Embedded Depthwise Unified Context Encoding (DEDUCE) module to jointly extract finer details and multi-scale short-range context from nearby regions to improve local understanding. To enable a holistic understanding, we propose a cascaded structure of dilated convolution and pooling-based pyramids. The resulting Pyramid over Pyramid (PoP) module introduces a new pyramid arrangement to maximize the effectiveness of individual pyramids due to their strategic positioning. Based on the DEDUCE and PoP modules, we propose DEDUCE-Net, which achieves extreme parameter efficiency while simultaneously achieving impressive accuracy. With only 0.16 million parameters, it achieves 70.32% mIoU on the cityscapes test set. Moreover, with resolution 512×1024, it achieves 238.5 and 25.76 FPS on RTX 3090 and Jetson Orin Nano (an embedded GPU), respectively. Furthermore, to achieve a power-efficient solution, we deployed the DEDUCE-Net on a ZCU102 board (Zynq UltraScale+ MPSoC), where the power consumption is significantly lower compared to that on the Jetson Orin Nano (embedded GPU). Specifically, the power consumption on ZCU102 is 980 mW, whereas on Jetson Orin Nano, it is 3.7 W; more than 4× higher. The source code and trained models will be released to facilitate further research and applications.