Bias-Free Semi-Supervised 3D Reconstruction via Occlusion Sensitivity-Guided Semantic Disentanglement

3D reconstruction faces challenges such as geometric warping and structural ambiguity, particularly in intricate topologies, heavy occlusions and complex backgrounds. These problems are partly attributed to excessive feature entanglement, which induces semantic confusion and spatial ambiguity. To address these limitations, we propose an occlusion sensitivity-guided semantic-disentangled Mamba-CNN network that enables human-controllable disentanglement of multi-attribute information within a bias-free semi-supervised framework. Specifically, we create various occlusion conditions and assign pseudo-labels to the augmented data within a semi-supervised framework, which enables the exploration of occlusion sensitivity of different semantic attributes for human-controllable semantic disentanglement. To reduce the bias between the augmented samples and their assigned pseudo-labels, we use linear PIoU and nonlinear MS-SSIM algorithms to calculate the confidence of pseudo-labels, which minimizes the error propagation caused by the bias. Then, we develop a disentangled multi-depth Mamba-CNN block that combines CNN’s local feature extraction capability with Mamba’s ability to capture long-range dependencies. This allows our model to effectively capture disentangled multi-attribute spatial features and semantic representations. However, critical cross-level semantic attribute connections could be lost in the disentanglement process. To tackle this, we propose a multi-attribute semantic query block to dynamically re-establish these connections and minimize cross-attribute information loss. Extensive quantitative and qualitative evaluations on object and face reconstruction demonstrate that our method outperforms existing state-of-the-art approaches. Codes and all resources will be publicly available on GitHub.