Evaluating Monocular Depth Estimation Methods on Industrial Objects

Abstract. Monocular Depth Estimation (MDE) has become a valuable tool in 3D reconstruction, especially when traditional methods such as photogrammetry are impractical. This study evaluates the performance of three state-of-the-art MDE algorithms, namely Depth Pro, Depth Anything V2 and Metric3Dv2, in estimating depth for challenging industrial objects with complex properties like reflectivity, transparency or lack of texture. Using both synthetic and real-world objects, algorithms' abilities to accurately estimate depth and generate 3D models are evaluated. Our findings show that Depth Pro outperforms Metric3Dv2 in handling these difficult scenarios, with significantly lower error rates and better handling of details such as edges and surfaces. The results demonstrate the potential of MDE in industrial applications, particularly where multi-camera systems or additional sensors are not feasible. However, while MDE offers promising solutions, further improvements are needed to fully address the unique challenges posed by industrial environments.