OASIS-DC: Generalizable Depth Completion via Output-level Alignment of Sparse-Integrated Monocular Pseudo Depth

TL;DR

The paper tackles metric depth completion under severe label scarcity by decoupling from feature-level foundation model coupling. It constructs a pseudo-depth prior by fusing a frozen monocular foundation estimator with sparse LiDAR anchors via a Poisson-based, gradient-guided fusion, then refines this prior with a lightweight residual network that preserves global scale while correcting local misalignments. An affinity-based refinement module using hyperbolic distances and center-tethered propagation further enforces geometry-aware, edge-preserving propagation, with learnable sensor anchoring to balance LiDAR observations and occlusions. Across KITTI-DC and NYUv2, the approach yields stable, sharp depth maps in 1-, 10-, and 100-shot regimes and in a deployment-oriented one-sequence setting, demonstrating strong generalization and practicality for real-world robotics and autonomous driving applications.

Abstract

Recent monocular foundation models excel at zero-shot depth estimation, yet their outputs are inherently relative rather than metric, limiting direct use in robotics and autonomous driving. We leverage the fact that relative depth preserves global layout and boundaries: by calibrating it with sparse range measurements, we transform it into a pseudo metric depth prior. Building on this prior, we design a refinement network that follows the prior where reliable and deviates where necessary, enabling accurate metric predictions from very few labeled samples. The resulting system is particularly effective when curated validation data are unavailable, sustaining stable scale and sharp edges across few-shot regimes. These findings suggest that coupling foundation priors with sparse anchors is a practical route to robust, deployment-ready depth completion under real-world label scarcity.

Figures (6)

• Figure 1: Pipeline overview. RGB, sparse depth, and relative (monocular) depth are concatenated and encoded. A residual decoder predicts an initial depth anchored to sparse points, while a refinement branch estimates an affinity map for edge/structure–aware propagation. The two streams are fused and passed through a hyperbolic refinement module (Hyp-Refine) to yield the final dense map. Symbols: $\oplus$ element-wise add, $\otimes$ channel-wise concat.
• Figure 2: Gradient-guided densification. The aligned prior induces a smooth gradient field (background), while sparse metric anchors (blue) and the image boundary are kept fixed. Depth values are propagated along the prior’s gradient directions to fill unknown cells, preserving structure and avoiding cross-edge bleeding; anchors are not moved—only the missing pixels are completed.
• Figure 3: KITTI-DC qualitative comparison under few-shot supervision. Top: 10-shot; bottom: 100-shot. Columns: RGB, CompletionFormer, BPNet, DepthPrompting, and OASIS-DC (Ours). Few-shot settings sample only from the training split; visual examples are evaluated against the official 1,000-frame validation set. Red boxes highlight challenging regions (thin structures, far-field). Our results preserve road–wall boundaries and fine details with reduced cross-edge bleeding.
• Figure 4: NYUv2 qualitative comparison. Top row: 1-shot (left: RGB); bottom row: 100-shot (left: Ground Truth). Remaining columns show DP (DepthPrompting), BPNet, UniDC, and Ours. The proposed method produces cleaner planar surfaces and sharper discontinuities (e.g., cabinet edges), while suppressing noise and texture copying across shots.
• Figure 5: Imperfect NYUv2 ground truth. Arrows mark GT artifacts (holes/smoothing); metrics use valid GT masks.