Rethinking Transparent Object Grasping: Depth Completion with Monocular Depth Estimation and Instance Mask

Accurate depth maps are essential for robotic grasping. However, transparent objects often cause depth cameras to produce missing or distorted depth due to reflection and refraction, making grasping them particularly challenging. Precise depth estimation for transparent objects is therefore crucial. Existing end-to-end methods aim to predict depth directly from RGB-D inputs. However, depth distortion varies significantly with environment, viewpoint, and illumination, making it difficult to predict. Models trained on limited datasets often fail to generalize effectively in real-world applications. In this letter, we propose ReMake, a novel transparent object depth completion framework that injects structured priors to reformulate the depth estimation learning objective. First, we introduce a mask prior to explicitly separate transparent regions from reliable background depth. This separation ensures consistency between training and real-world application, where masked areas correspond to transparent regions and unmasked areas to opaque ones. This improves the model's generalizability in real-world applications. Second, we further incorporate relative depth as an additional prior obtained from monocular depth estimation, which encodes the depth relationship between transparent objects and their background to guide the reconstruction of transparent regions more effectively. Extensive experiments across datasets and real-world scenes demonstrate that ReMake significantly outperforms existing methods, especially in different-view prediction experiments, evaluating its generalizability.