Abstract:

Monocular depth estimation is a core technology for 3D environmental perception and is of significant value in applications such as autonomous driving and robot navigation. As a representative lightweight monocular depth estimation model in these fields， Lite-Mono achieves an excellent balance between performance and efficiency through dilated convolutions and a local-global attention mechanism. However， there are two major bottlenecks in current Lite-Mono architecture. The first one is that fixed dilation rate in the dilated convolutions leads to mismatch between receptive field and target scale， so that features of small objects are lost due to excessively large detailed rate， while contextual information of large objects is lacked due to insufficient receptive field. The second one is that high computational complexity of O（H2W2C） in local-global attention module hinders processing of high-resolution input， thus limiting real-time application. To address these issues， an enhanced model based on Lite-Mono architecture， DDHL （Dynamic Dilated convolution and Hierarchical Local-global attention）， was proposed. First， a Dynamic Dilated Convolution Module （DDCM） was introduced to adjust the receptive field in real time via a fusion weight rate predictor and generate adaptive weights by combining channel attention. Second， a hierarchical local-global attention module was designed to reduce the computational complexity to O（（HW/M）2C）. Finally， a learnable global token was incorporated to establish cross-window dependencies， so as to assist models in understanding and analyzing visual content more accurately in complex scenes， thereby enhancing perception and modeling capabilities of the overall scene. Experimental results on Make3D dataset demonstrate that DDHL model achieves significant improvements on key metrics： Absolute Relative Error （Abs Rel） is reduced from 0.462 to 0.290 with a decrease of 37.2%， and Squared Relative Error （Sq Rel） is decreased by 49.7%. It can be seen that this model achieves good balance between accuracy and efficiency， demonstrating practical application value.

Key words: monocular depth estimation, self-supervised learning, dynamic dilated convolution, hierarchical attention mechanism, multi-scale feature extraction

摘要：

单目深度估计作为三维环境感知的核心技术，在自动驾驶和机器人导航等领域具有重要的应用价值。在这些领域，Lite-Mono作为轻量级单目深度估计的代表模型，通过空洞卷积和局部-全局注意力机制在保持高效的同时实现了优异的性能。然而，目前的Lite-Mono架构面临两大瓶颈：空洞卷积受限于固定空洞率，导致感受野与目标尺度严重失配，从而使小目标因过大空洞率丢失特征，而大目标因过小感受野缺失上下文；另外，局部-全局注意力模块因O(H2W2C)计算复杂度难以处理高分辨率输入，导致实时应用受到限制。针对上述问题，在Lite-Mono架构的基础上，提出一种改进模型DDHL（Dynamic Dilated convolution and Hierarchical Local-global attention）。首先，提出动态空洞卷积模块（DDCM），以通过融合权重率预测器实时调整感受野，并结合通道注意力生成自适应权重；其次，提出分层局部-全局注意力机制模块将复杂度降至O({(HW/M)}^{\mathrm{2}}C)；最后，引入可学习全局Token构建跨窗口约束，以助力模型在复杂场景中更精准地理解和分析视觉内容，从而提升对整体场景的感知与建模能力。实验结果表明，DDHL模型在Make3D数据集上的关键指标取得了突破：绝对相对误差（Abs Rel）从0.462降至0.290（降幅37.2%），平方相对误差（Sq Rel）降低了49.7%。可见，该模型在精度与效率之间取得了良好平衡，具有实际应用价值。

关键词: 单目深度估计, 自监督学习, 动态空洞卷积, 分层注意力机制, 多尺度特征提取