High-Quality Textureless Object Reconstruction via Depth-Constrained 3D Gaussian Splatting

High-quality 3D reconstruction of textureless objects is of significant importance in applications such as Augmented Reality （AR）， industrial inspection， and robotic autonomous grasping. However， owing to the lack of distinctive surface textures， traditional methods based on Multi-View Stereo （MVS） and feature matching often fail to establish reliable correspondences， leading to unsuccessful reconstructions. In recent years， 3D Gaussian Splatting （3DGS） has achieved notable advances in rendering quality and computational efficiency. Nevertheless， it continues to face inherent limitations when applied to textureless surfaces， frequently producing incomplete reconstructions characterized by holes and artifacts. To address the lack of standardized evaluation datasets in this domain， we construct a high-quality RGB-D dataset comprising 11 categories of representative textureless objects， thereby supporting further research in this area. Building upon this dataset， we propose a novel reconstruction framework， termed QDE-GS. This framework incorporates the Mask R-CNN instance segmentation network to effectively eliminate background interference and enable focused geometric reconstruction of the target object. To tackle the core challenge of missing geometric information in textureless regions， we further develop a dual-branch network， referred to as GP-FusionNet. The proposed network performs a deep fusion of the input raw depth maps and RGB photometric information， followed by regularization using a 3D U-Net architecture to produce a high-fidelity dense depth map. This output provides robust geometric priors for the initialization and optimization of 3D Gaussians. Furthermore， to suppress rendering artifacts and enhance robustness to low-quality inputs， we introduce a quality-aware optimization strategy along with a distortion network based on No-Reference Image Quality Assessment （NR-IQA）. In addition， to improve overall pipeline efficiency and eliminate redundant camera views， we design a view selection algorithm guided by camera pose information. Experimental results demonstrate that QDE-GS consistently achieves superior performance on the Novel View Synthesis （NVS） task across both the proposed dataset and publicly available benchmarks. Quantitative evaluations further confirm that the proposed method surpasses existing state-of-the-art methods on key image quality metrics， including Peak Signal-to-Noise Ratio （PSNR）， Structural Similarity Index Measure （SSIM）， and Learned Perceptual Image Patch Similarity （LPIPS）. In addition， the method substantially improves the rendering fidelity of textureless objects under challenging and complex viewing conditions.