基于转置注意力的多尺度深度融合单目深度估计
Multi-scale Depth Fusion Monocular Depth Estimation Based on Transposed Attention
程亚子 1雷亮 2陈瀚 1赵毅然1
作者信息
- 1. 广东工业大学物理与光电工程学院,广东 广州 510006
- 2. 广东工业大学物理与光电工程学院,广东 广州 510006;广东省信息光子技术重点实验室,广东 广州 510006
- 折叠
摘要
单目深度估计是计算机视觉领域中一项基础任务,其目标是通过单张图像预测深度图,并获取每个像素位置的深度信息.本文提出一种新的单目深度估计网络结构,旨在进一步提高网络的预测准确性.转置注意力机制在降低参数量和计算量的同时引入了自注意力机制,以关注图像中的特定区域,并结合不同通道之间的信息.这种机制能够有效地关注到图像中的细小区域和边缘信息,并进行学习.本文还提出一种改进的转置注意力机制,以更少的参数量保留语义信息.多尺度深度融合根据不同通道提取不同深度特征的特点,计算每个通道的平均深度,以增强模型的深度感知能力.此外,它能够建模垂直距离的长距离关系,有效地分离物体之间的边缘,有助于减少细粒度信息的损失.最后,本文在NYU Depth V2数据集和KITTI数据集上进行实验,验证了所提出模块的有效性,并取得了出色的性能表现.
Abstract
Monocular depth estimation is a fundamental task in computer vision,aiming to predict depth maps from single im-ages and retrieve depth information for corresponding pixel positions.This paper proposes a novel network architecture for mon-ocular depth estimation to further enhance the predictive accuracy of the network.Transposed attention introduces a self-attention mechanism,enabling it to focus on specific regions within the image while reducing the parameter and computation re-quirements.By incorporating information across different channels,it effectively captures fine-grained regions and edge details for learning.The paper presents an improved version of transposed attention that retains semantic information with fewer param-eters.Multi-scale depth fusion leverages the characteristic of extracting features with different depths from distinct channels.It computes the average depth for each channel,enhancing the model's depth perception capability.Furthermore,it models long-range dependencies for vertical distances,effectively separating edges between objects and mitigating the loss of fine-grained in-formation.Finally,the proposed modules'effectiveness is validated through experiments conducted on the NYU Depth V2 data-set and the KITTI dataset,demonstrating exceptional performance.
关键词
深度学习/单目深度估计/转置注意力/多尺度深度融合/通道平均深度Key words
deep learning/monocular depth estimation/transposed attention/multi-scale deep fusion/channel average depth引用本文复制引用
基金项目
国家自然科学基金资助项目(62006046)
出版年
2024
国家科技期刊平台
NSTL
万方数据