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

Abstract

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

Keywords

• Semantic segmentation
• feature scaling
• feature aggregation
• deep learning
• scene understanding
• convolutional neural networks