DOC-Depth: Multi-Domain Depth Modeling

• DOC-Depth is a multi-domain framework that employs depth-aware techniques, from dense ground truth generation using LiDAR-camera calibration and dynamic object segmentation to address sensor sparsity.
• It leverages advanced methods such as MAP inversion with SF-CRF in SD-OCT and DCT-based progressive prediction in monocular setups to ensure high metric accuracy and reliable depth estimation.
• The approach further extends to reliability-driven 3D reconstruction and output-level alignment for depth completion, while also inspiring hierarchical planning in text generation with controlled outline depth.

DOC-Depth refers to a family of methods, frameworks, and modules in computer vision and related fields that employ “depth” as a core structural, algorithmic, or supervisory concept. Across the literature, “DOC-Depth” appears both as the explicit name of methods for dense depth ground truth generation, as in "DOC-Depth: A novel approach for dense depth ground truth generation" (Moreau et al., 4 Feb 2025), and as a defining element in modules like the depth-compensated processing pipeline for SD-OCT (Boroomand et al., 2015), monocular depth estimation in the DCT domain (Wang et al., 2024), sparse-view 3D Gaussian Splatting reconstruction (Li et al., 8 Apr 2026), and output-level alignment for generalizable depth completion (Cho et al., 1 Feb 2026). The term is also found in linguistic domains for hierarchical outline depth control (Yang et al., 2022), but the dominant thread is depth estimation, compensation, or supervision in visual scene understanding. The following sections summarize the principal methodologies, architectures, and evaluation protocols for “DOC-Depth” across its variants.

1. Dense Depth Ground Truth Generation with DOC-Depth

DOC-Depth for dense ground-truth depth map generation (Moreau et al., 4 Feb 2025) addresses a core problem in computer vision datasets: available LiDAR-camera datasets are typically semi-dense due to occlusions, surface reflectance, and sensor sparsity. Standard compression or completion strategies are often heuristic, require image priors, or hallucinate geometry.

The DOC-Depth pipeline integrates four calibrated submodules:

• Sensor Calibration via Zhang’s checkerboard method to determine camera intrinsics/extrinsics, with explicit computation of rotation $R^{c\to l}$ and translation $u^{c\to l}$ by fitting correspondence planes in the two modalities.
• LiDAR Odometry & Aggregation using a low-drift SLAM (e.g., Exwayz LiDAR SLAM) to produce pose-registered frames; aggregation in the world frame allows dense point coverage.
• Dynamic Object Classification (DOC) which segments point clouds into static and dynamic points using a voting scheme. Spherical range images are used to compare per-point ranges across frames, with special rules to avoid ground occlusion false negatives.
• Composite Depth Rendering: For each camera timestamp, (i) static points are projected from all relevant aggregated frames, (ii) dynamic points are projected only from the frame closest in time, and (iii) a GPU z-buffer pass fuses splats, ensuring correct occlusion and minimizing “ghost” trails.

This unified approach improves depth map density from 16.1% to 71.2% on KITTI, with demonstrated generalizability across LiDAR sensors and environments (no retraining required for new hardware or domains).

Component	Purpose	Implementation Highlight
Sensor Calibration	Align LiDAR and RGB frames in a common coordinate system	Checkerboard-based, analytic extrinsics
LiDAR Odometry & Aggregation	Produce spatially consistent dense 3D maps	SLAM with low-drift
DOC Module	Segmentation of static and dynamic scene points for compositing	Spherical voting, parameter-free DOC
Composite Rendering	Pixel-wise dense depth assignment with correct occlusion and object placement	GPU z-buffer pass with dynamic/static

2. Depth Compensation in Spectral Domain OCT Imaging

The DOC-Depth concept in the context of Spectral Domain Optical Coherence Tomography (SD-OCT) (Boroomand et al., 2015) models and corrects depth-varying imaging degradations. SD-OCT suffers from four main depth-dependent artifacts: axial resolution loss, lateral resolution loss, SNR roll-off, and sidelobe artifacts. The DOC-Depth approach achieves digital-only compensation using a MAP framework augmented by a stochastically fully-connected conditional random field (SF-CRF):

• Axial/Lateral PSF Estimation: Empirical measurement and modeling of resolution loss as functions of depth, used directly in the forward model.
• MAP Inversion with SF-CRF: The measurement at pixel $i$, $M_i$, is modeled as a convolution of the underlying intensity $V_i$ with the depth-varying PSFs and sidelobe kernel, plus multiplicative speckle noise. The inverse problem is cast as maximizing $P(x|y)$, with unary (data fidelity) and pairwise (edge-preserving, depth-dependent) terms.
• Edge-Preserving Regularization: SF-CRF weights are made depth-adaptive and robust to local noise by incorporating measurement patches and stochastic clique selection.
• Gradient Descent Optimization: The energy function is minimized iteratively over the entire image or volume.

DOC-Depth achieves near-constant axial and lateral resolution and compensates SNR losses, outperforming both baseline and ML-based SD-OCT methods across standard physical phantoms and biological tissues.

3. Frequency-Domain Monocular Depth Estimation (DCDepth)

The “DOC-Depth” framework in DCDepth (Wang et al., 2024) reformulates monocular depth estimation as progressive prediction of discrete cosine transform (DCT) coefficients for small spatial patches. Depth patches of size $S×S$ are projected onto DCT bases $B^{i,j}_{u,v}$; coefficients near $(0,0)$ represent low-frequency (global) scene structure, whereas high-frequency entries encode sharp local variations.

The pipeline comprises:

• Patch-wise DCT: Transforming spatial depth into frequency coefficients.
• Coarse-to-Fine Prediction: GRU-based progressive head predicts DCT coefficients in sub-diagonal frequency groups, beginning with low-frequencies and refining with higher bands. Each step produces an intermediate spatial depth map via inverse DCT, which is supervised by a scale-invariant log loss.
• Feature Extraction and Fusion: Swin-Large transformer for multi-scale encoding; DCT-based feature fusion via Pyramid Feature Fusion (PFF), which preserves frequency content during downsampling.
• Regularization: Explicit penalties on high-frequency coefficient magnitude ($L_f$) and edge-aware smoothness ($u^{c\to l}$0).

DCDepth demonstrates improved metric accuracy (e.g., lower AbsRel, RMSE) and parameter efficiency across NYU-Depth-V2, TOFDC, and KITTI compared to prior approaches. Progressive intermediate predictions yield smoother planar regions and sharper edges.

4. Dual-Domain Depth Reliability in 3D Gaussian Splatting

DOC-Depth as per "DOC-GS" (Li et al., 8 Apr 2026) is an explicit module for ensuring reliability in sparse-view 3D Gaussian Splatting (3DGS) scene reconstruction where geometric supervision is insufficient. The core mechanism comprises:

• Continuous Depth-Guided Dropout (CDGD): For each Gaussian primitive $u^{c\to l}$1, a continuous probability $u^{c\to l}$2 (function of camera-space depth $u^{c\to l}$3 and normalized “importance” $u^{c\to l}$4) governs stochastic dropout. Deeper Gaussians, typically weakly supervised, are dropped more often, enforcing a smooth, depth-aware inductive regularization directly in the optimization domain.
• Observation-Domain Calibration via Dark Channel Prior (DCP): Rendered views are analyzed with the DCP, flagging haze-like artifacts. Anomaly scores are cross-view accumulated for each Gaussian.
• Reliability-Driven Pruning: Gaussians with high DCP anomaly scores and low opacity are pruned at regular intervals, effectively suppressing “floater” artifacts without harming well-constrained inliers.

This dual-domain, end-to-end differentiable approach improves reconstruction fidelity under sparse-view regimes, providing both a smooth regularizing bias and adaptive artifact detection.

5. Output-Level Alignment for Generalizable Depth Completion

“DOC-Depth” in OASIS-DC (Cho et al., 1 Feb 2026) addresses depth completion under extreme label scarcity. The method couples a nonlearned pseudo-metric prior, generated by fusing a monocular foundation model’s relative depth with sparse LiDAR anchors using a Poisson boundary-value problem, to a lightweight refinement network:

• Pseudo-metric Prior Construction: Given dense relative depth $u^{c\to l}$5 and sparse LiDAR depth $u^{c\to l}$6, Dirichlet boundary values are imposed on pixels with LiDAR readings and at image edges. A Poisson fusion then solves for a globally metrically anchored, but edge-preserving, prior $u^{c\to l}$7.
• Refinement Network: A minimal encoder-decoder network produces residuals, and a hyperbolic affinity propagation module sharpens structure and boundaries. Local updates are “tethered” to the prior and recalibrated to LiDAR at every step via sensor-anchoring.
• Output-Level Alignment: Only the prediction head (not the encoder) is trained, ensuring adaptation with zero validation curation, sharp edges, and robust scale correction, even in 1-shot or 10-shot training.

Empirical results demonstrate top-tier accuracy (RMSE, MAE) with only $u^{c\to l}$8M parameters trained, outperforming more complex methods on KITTI and NYUv2 in 1/10/100-shot regimes.

Subsystem	Role	Notable Feature
Poisson Prior	Edge-preserving, metrically anchored depth from LiDAR+MDE	Linear system, no learned parameters
Residual Decoder	Local correction of prior	Small capacity, output-level only
Hyperbolic Affinity	Propagates fine detail, prevents bleeding across structures	Tethered, multi-scale

6. Outline Depth in Hierarchical Long-Form Generation

Outside visual domains, "DOC-Depth" denotes maximum outline depth $u^{c\to l}$9 for hierarchical story planning in the Detailed Outline Control (DOC) framework (Yang et al., 2022). Here, outline depth controls planning granularity in a tree, with deeper $i$0 supporting finer event decomposition. More granular outlines yield improved coherence, relevance, and control in generated stories, with $i$1 achieving performance gains over $i$2 (shallow planning).

This use-case, while lexically aligned, does not interact with geometric or metric imaging depth, instead referring to planning tree depth in structured content generation.

7. Summary and Impact

Across its visual, signal processing, and linguistic instantiations, “DOC-Depth” connotes structured depth modeling—either geometric depth for dense annotation, reliability-aware 3D modeling, progressive frequency decomposition, or hierarchical planning. The recurring methodological innovation is to enforce depth-aware constraints at the level of modeling (e.g., via regularization, dropout, or prior fusion), algorithmic supervision (e.g., via data-anchored priors or hierarchical expansion), or artifact correction (e.g., via DCP driven pruning). These frameworks prioritize robustness to sensor sparsity, occlusion, dynamic environments, and minimal supervision, and they characterize a set of best practices underpinning state-of-the-art performance for depth estimation, completion, and reliability correction in computer vision and beyond (Boroomand et al., 2015, Wang et al., 2024, Moreau et al., 4 Feb 2025, Li et al., 8 Apr 2026, Cho et al., 1 Feb 2026, Yang et al., 2022).