Golden Cudgel Network for Real-Time Semantic Segmentation

2025-03-05

AI Summary

Real-time semantic segmentation suffers from high inference latency due to multi-path architectures, while prevailing methods rely on external teacher models for knowledge distillation. Method: This paper proposes a training-inference dynamic reconfigurable architecture: during training, it employs jointly expanded vertical multi-convolutional and horizontal multi-path structures to enhance representational capacity; at inference time, structural reparameterization equivalently compresses the model into a single-path, single-convolution architecture—enabling zero-overhead acceleration. Crucially, this self-distillation mechanism operates without any external teacher model. Contribution/Results: The method achieves state-of-the-art performance on Cityscapes, CamVid, and Pascal VOC 2012, simultaneously advancing both accuracy and inference speed along the Pareto frontier. It establishes a new benchmark for efficient real-time semantic segmentation by unifying architectural flexibility, training-time expressiveness, and inference-time efficiency.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Deep Neural Architectures and Foundation Models
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Recent real-time semantic segmentation models, whether single-branch or multi-branch, achieve good performance and speed. However, their speed is limited by multi-path blocks, and some depend on high-performance teacher models for training. To overcome these issues, we propose Golden Cudgel Network (GCNet). Specifically, GCNet uses vertical multi-convolutions and horizontal multi-paths for training, which are reparameterized into a single convolution for inference, optimizing both performance and speed. This design allows GCNet to self-enlarge during training and self-contract during inference, effectively becoming a"teacher model"without needing external ones. Experimental results show that GCNet outperforms existing state-of-the-art models in terms of performance and speed on the Cityscapes, CamVid, and Pascal VOC 2012 datasets. The code is available at https://github.com/gyyang23/GCNet.

Problem

• Overcomes speed limitations in real-time semantic segmentation models
• Eliminates dependency on high-performance teacher models for training
• Optimizes performance and speed using reparameterized convolution techniques

Innovation

• Vertical multi-convolutions for training efficiency
• Horizontal multi-paths reparameterized for inference speed
• Self-enlarging and self-contracting without external teacher models