Title: Mitigating Depth Ambiguity: Employing Mixture-Density Representations for Robust Depth Estimation Without Flying Points

Abstract: Although significant progress has been made in depth estimation, "flying points" continue to plague the field as a recurring error. These artifacts manifest as spurious 3D points appearing in the voids between foreground and background objects, particularly near object edges. This issue stems from the conventional approach of assigning a single depth value to each pixel. At boundaries, a pixel often overlaps both foreground and background surfaces, creating genuine depth ambiguity. Because a standard single-hypothesis model cannot simultaneously represent both distinct depths, the training process forces the prediction toward an intermediate value that corresponds to neither surface, resulting in erroneous geometry. To resolve this, we introduce MDA, a method utilizing a mixture-density representation that enables the network to output multiple depth hypotheses alongside their respective probabilities for every pixel. This approach allows distinct hypotheses to align with separate surfaces near boundaries, ensuring the final decoded depth is chosen from valid hypotheses rather than falling into the empty space between them. Experiments across various backbones demonstrate that MDA significantly enhances boundary reconstruction and effectively eliminates flying-point artifacts, even when inputs are heavily blurred, all while incurring minimal computational cost. Furthermore, this mixture-density framework is versatile: it naturally accommodates transparent objects by predicting multiple depth layers for transparent pixels and addresses sky regions through a specialized component that distinguishes the unbounded sky from finite-depth areas, thereby generating clean, flying-point-free skylines. Project Page: https://biansy000.github.io/mda-site/.

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