Title: MCPDepth: Achieving Omnidirectional Depth Estimation Through Stereo Matching Across Multi-Cylindrical Panoramas

Abstract:

Estimating depth in omnidirectional scenes is notoriously difficult due to the severe distortions inherent in panoramic imagery. While the field has seen considerable progress, the specific influence of various projection techniques has not been thoroughly investigated. To address this gap, we present MCPDepth, a new two-stage framework that boosts omnidirectional depth accuracy by employing stereo matching across multiple cylindrical panoramic views. The approach first conducts stereo matching using cylindrical panoramas and then robustly fuses the depth maps generated from these different perspectives.

In contrast to prior methods that depend on specialized kernels to manage image distortions, MCPDepth leverages standard neural network components. This design choice enables efficient deployment on resource-constrained embedded devices without compromising performance. To specifically mitigate vertical distortions common in cylindrical projections, the framework integrates a circular attention module, which effectively enlarges the receptive field well beyond the limits of conventional convolutional operations.

We offer a detailed theoretical and experimental evaluation of standard panoramic projection types—specifically spherical, cylindrical, and cubic—highlighting the superior effectiveness of the cylindrical projection method. Our results demonstrate significant improvements in mean absolute error (MAE), achieving an 18.8% reduction on the outdoor Deep360 dataset and a 19.9% reduction on the real-world 3D60 dataset. This study provides valuable practical guidance for related tasks and real-world applications, setting a fresh standard in omnidirectional depth estimation. The source code is accessible at https://github.com/Qjizhi/MCPDepth.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC