基于切片自查询的室外自监督单目深度估计

Outdoor self supervised monocular depth estimation based on slice self query

摘要:

针对室外单目深度估计中垂直方向敏感与局部卷积感受野受限导致的全局深度信息缺失问题，开展列方向全局建模研究。在编码阶段引入垂直像素自注意力以提取列级全局深度信息；在解码路径中对U-Net视觉特征执行按列重标定，使含全局先验的列获得更高权重并抑制噪声列；同时设置自查询模块，对编码器与解码器输出特征进行粗粒度查询，构建自生成代价体与深度分布特征并加权聚合以回归深度。在KITTI基准上，AbsRel降至0.087，SqRel降至0.653，RMSE降至4.120，δ<1.25提升至0.921；在Make3D 上，AbsRel由0.306降至0.304，RMSE由6.856降至6.778，RMSElog由0.151降至0.149。结论显示，列方向全局依赖的显式建模与概率融合弥补了局部卷积的全局感知不足，提升远距与纹理稀疏区域的估计稳定性与跨域鲁棒性，能够满足复杂室外场景对单目深度估计精度与泛化能力的需求。

Abstract:

To mitigate the loss of global depth information caused by vertical-direction sensitivity and the limited receptive fields of local convolutions in outdoor monocular depth estimation, a column-wise global modeling approach is adopted. The encoder employs a vertical-pixel self-attention module to extract column-level global cues; the decoder re-calibrates U-Net features column-wise to emphasize globally informed columns and suppress noise. A self-query module performs coarse queries on encoder/decoder features to build a self cost volume and depth-distribution features, which are aggregated for depth regression. On KITTI, AbsRel decreases to 0.087, SqRel to 0.653, RMSE to 4.120, and δ<1.25 increases to 0.921; on Make3D, AbsRel drops from 0.306 to 0.304, RMSE from 6.856 to 6.778, and RMSElog from 0.151 to 0.149. The method compensates for limited global perception in local convolutions, improving far-range/texture-sparse stability and cross-domain robustness.