Scholarlink

Monocular Depth Estimation (MDE) plays an important role in many fields. However, current MDE requires complex models that are computationally intensive and unsuitable for real-time applications. Some MDE methods also have a high memory access frequency (MAF), leading to substantial data transfer demands. This paper presents EffiNet, a novel network structure designed for MDE. We emphasize high performance and lightweight computing. EffiNet significantly reduces computational complexity while maintaining accurate depth estimation by utilizing MobileNetV2 as the backbone for the encoder. The newly designed decoding algorithm optimizes the model by implementing block processing strategies, efficient upsampling, and point convolution operations. The decoder enhances performance under various scene conditions while keeping computational costs low. This approach also minimizes MAF, thereby reducing data transmission requirements and improving execution speed and energy efficiency. Experimental results demonstrate that EffiNet achieves a reduction in inference time by at least 18.18% on Jetson Nano and 24.39% on Jetson TX2, and a maximum 96.07% reduction in computational complexity (FLOPs), compared to existing methods. Additionally, the accuracy index \(\delta _1\) shows an improvement of 0.44% to 65.70%. These results validate the superior performance and suitability of resource-constrained mobile and edge computing devices. These advancements position EffiNet as an innovative solution in the field of MDE, capable of delivering high efficiency and reducing resource consumption.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

EffiNet: An Efficient, Low-Complexity Neural Network for Monocular Depth Estimation on Embedded Devices

  • 会议论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年

查看全文
  • Xiaohan Tu,
  • Chuanhao Zhang,
  • Haiyan Zhuang,
  • Zhen Sang,
  • Mengran Liu

摘要

Monocular Depth Estimation (MDE) plays an important role in many fields. However, current MDE requires complex models that are computationally intensive and unsuitable for real-time applications. Some MDE methods also have a high memory access frequency (MAF), leading to substantial data transfer demands. This paper presents EffiNet, a novel network structure designed for MDE. We emphasize high performance and lightweight computing. EffiNet significantly reduces computational complexity while maintaining accurate depth estimation by utilizing MobileNetV2 as the backbone for the encoder. The newly designed decoding algorithm optimizes the model by implementing block processing strategies, efficient upsampling, and point convolution operations. The decoder enhances performance under various scene conditions while keeping computational costs low. This approach also minimizes MAF, thereby reducing data transmission requirements and improving execution speed and energy efficiency. Experimental results demonstrate that EffiNet achieves a reduction in inference time by at least 18.18% on Jetson Nano and 24.39% on Jetson TX2, and a maximum 96.07% reduction in computational complexity (FLOPs), compared to existing methods. Additionally, the accuracy index \(\delta _1\) shows an improvement of 0.44% to 65.70%. These results validate the superior performance and suitability of resource-constrained mobile and edge computing devices. These advancements position EffiNet as an innovative solution in the field of MDE, capable of delivering high efficiency and reducing resource consumption.