DECoDeNet: Achieving Impressive Accuracy–Efficiency Trade-Off for Semantic Segmentation in Embedded Applications

Semantic segmentation is a vital task in autonomous driving, providing pixel-level understanding of road scenes. However, deploying such models on embedded platforms requires a careful balance between accuracy and efficiency due to limited computational resources and the safety-critical nature of the application. To address this, we propose DECoDeNet, a lightweight segmentation network featuring a novel non-residual block that employs Densely Embedded Contextual Descriptors (DECoDe) to aggregate contextual information densely without creating large receptive field gaps. Unlike traditional squeeze-expand designs that compress and then expand channel dimensions, our approach maintains a uniform channel space, thereby preserving critical information along the channel dimension. Additionally, we introduce the Hierarchically Extracted Low-dimensional Pyramid (HELP) module, which complements the DECoDe module by enabling holistic scene understanding through multi-scale pooling kernels. With only 0.27M parameters, DECoDeNet achieves an impressive 75.4% mIoU on the Cityscapes test set, making it the smallest model known to exceed the 75% mIoU mark. Moreover, it achieves 221.4 FPS on an RTX 3090 (a high-end GPU) with resolution 512×1024. We also deployed our model on Jetson Orin Nano, an embedded platform with constrained resources, where it delivers a notable 15.9 FPS while maintaining decent segmentation accuracy, demonstrating its suitability for real-world edge applications.