Efficient Semantic Segmentation via Liquid Time-constant Networks with Adaptive Dynamics

Engineering notes

Semantic segmentation is a vital and foundational task in computer vision, yet recent deep-learning-based methods impose prohibitively high computational costs and model complexity.In this paper, we introduce a novel segmentation architecture that integrates Liquid Time-Constant (LTC) networks with conventional convolutional neural networks, achieving competitive performance while dramatically reducing model size.Inspired by biological neural circuits of Caenorhabditis elegans, LTC networks achieve adaptive processing dynamics through inputdependent time constants.Our architecture combines a U-Net style encoder-decoder with a Closed-form Continuousdepth (CfC) feature modulation head that learns to process spatial features using biologically-plausible temporal dynamics.We evaluate our method on the DDOS (Drone Depth and Obstacle Segmentation) dataset and the CamVid urban driving dataset, demonstrating that our approach achieves 0.53 mIoU with only 2.4M parameters representing a 13× reduction compared to U-Net (31M parameters) while retaining 96% of its accuracy.Furthermore, our model requires only 1.8 GFLOPs compared to U-Net's 54.8 GFLOPs, a 30× computational reduction.To isolate the contribution of CfC dynamics from generic channel gating, we compare against matched-parameter baselines (SEblock, Linear Gate) and demonstrate that liquid neural network dynamics provide measurable benefits beyond simple channel attention.The experimental results suggest that liquid neural networks offer a promising paradigm for building efficient yet flexible computer vision systems deployable in resource-constrained environments such as drones, mobile devices, and embedded systems.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Semantic segmentation is a vital and foundational task in computer vision, yet recent deep-learning-based methods impose prohibitively high computational costs and model complexity.In this paper, we introduce a novel segmentation architecture that integrates Liquid Time-Constant (LTC) networks with conventional convolutional neural networks, achieving competitive performance while dramatically reducing model size.Inspired by biological neural circuits of Caenorhabditis elegans, LTC networks achieve adaptive processing dynamics through inputdependent time constants.Our architecture combines a U-Net style encoder-decoder with a Closed-form Continuousdepth (CfC) feature modulation head that learns to process spatial features using biologically-plausible temporal dynamics.We evaluate our method on the DDOS (Drone Depth and Obstacle Segmentation) dataset and the CamVid urban driving dataset, demonstrating that our approach achieves 0.53 mIoU with only 2.4M parameters representing a 13× reduction compared to U-Net (31M parameters) while retaining 96% of its accuracy.Furthermore, our model requires only 1.8 GFLOPs compared to U-Net's 54.8 GFLOPs, a 30× computational reduction.To isolate the contribution of CfC dynamics from generic channel gating, we compare against matched-parameter baselines (SEblock, Linear Gate) and demonstrate that liquid neural network dynamics provide measurable benefits beyond simple channel attention.The experimental results suggest that liquid neural networks offer a promising paradigm for building efficient yet flexible computer vision systems deployable in resource-constrained environments such as drones, mobile devices, and embedded systems.

Links and sources

• PDF from original source

Comments