Multi-encoder Context Aggregation Network for Structured and Unstructured Urban Street Scene Analysis

Developing computationally efficient semantic segmentation models that are suitable for resource-constrained mobile devices is an open challenge in computer vision research. To address this challenge, we propose a novel real-time scene model called Multi-encoder Context Aggregation Network (MCANet), which offers the best combination of low complexity and state-of-the-art (SOTA) performance on benchmark datasets. While follow multi-encoder approach, our novelty lies varying number scales to capture both global context local details effectively. We introduce lateral connections between sub-encoders improved feature refinement. also optimize backbone by exploiting residual block MobileNet applications. On decoder side, proposed includes new Local Global (LGCA) module significantly enhances output. Finally, use several known convolution techniques classification make more efficient. provide comprehensive evaluation MCANet multiple datasets containing structured unstructured urban street scenes. Among existing with less than 3 million parameters, competitive as it achieves SOTA without ImageNet pre-trained weights environments while being compact