CVPR 2026 · Workshop 3D4S

R3DC: Reliability-Guided
Reveal-to-Revise Depth Completion
for Cross-Domain Sparse Perception

An end-to-end framework that jointly predicts dense metric depth, per-pixel reliability, and aleatoric uncertainty — with 10–67× fewer parameters than competing baselines.

Noor Islam S. Mohammad  ·  Uluğ Bayazıt
Istanbul Technical University  ·  Department of Computer Science & Computer Engineering
📧 {islam23, ulugbayazit}@itu.edu.tr
📄 Read Paper 🔬 Method 📊 Results
0 m RMSE · KITTI
0 δ₁ · NYU Depth V2
1.95M Parameters
4 Domains
≈20 FPS · Tesla T4
SCROLL

Depth completion is fundamental to 3D perception, yet practical deployment is hindered by three challenges: an absence of calibrated per-pixel confidence, poor cross-domain generalization, and benchmarks that evaluate accuracy while ignoring the trustworthiness of uncertainty estimates.

We introduce R³DC, an end-to-end Reveal-to-Revise framework that jointly predicts dense metric depth, per-pixel reliability, and aleatoric uncertainty. The architecture integrates a dual-stream encoder with geometry-adaptive deformable convolutions, hierarchical cross-modal attention, and CSPN++ refinement explicitly gated by learned reliability. Driven by a seven-term composite objective, R³DC stabilizes training across a highly diverse range of depths.

To rigorously assess confidence estimates, we propose RADI (Reliability-Aware Depth Index), a novel evaluation framework measuring reliability-error correlation (REC), revision benefit score (RBS), and calibration error (CAL). Across four heterogeneous benchmarks (KITTI, VisDrone, Drone-Videos, and NYU Depth V2), R³DC achieves 0.24 m RMSE on KITTI and δ₁ = 0.927 on NYU Depth V2, while its core architecture requires 6–67× fewer parameters than existing baselines.

Three Open Challenges
in Depth Completion

Despite impressive progress in depth estimation, these fundamental problems remain unaddressed by prior work.

01
🎯

No Per-Pixel Confidence

Downstream modules like motion planning and obstacle avoidance must know where to trust depth estimates. Textureless walls, reflective surfaces, occlusion boundaries, and distant objects receive identical treatment — yet have vastly different geometric difficulty.

02
🌍

Cross-Domain Fragility

Ground-level LiDAR models (KITTI: d ≤ 80 m, near-horizontal) generalize poorly to aerial cameras (VisDrone: bird's-eye), drone video (d ∈ [0,50] m), or indoor RGB-D (NYU V2: d ≤ 10 m). No prior work validated a single architecture across all four modalities.

03
📏

Reliability Evaluation Gap

Standard metrics like RMSE and δ₁ quantify average accuracy but not trustworthiness. They can heavily favor models that perform well on easy pixels while remaining dangerously overconfident in hard, ambiguous regions.


The Reveal → Reliability → Revise Pipeline

R³DC decomposes depth completion into three ordered stages, each with a specialized mechanism. The key insight: a single decoder cannot simultaneously infer global structure and respect sharp boundaries.

RGB Image I ∈ ℝ³ˣᴴˣᵂ Sparse Depth dₛ, Mₛ DUAL-STREAM ENCODER RGB Stream ResBlocks + DropPath Depth Stream (DCN) Deformable Conv v2 Cross-Modal Attn (CMA) 3 scales, Nmax=512 REVEAL Stage TRANSFORMER BOTTLENECK 8 heads CBAM + SA 1/16 resolution Global context FPN DECODER 4× EfficientUpBlock Dec4: ConvT+DCN+CBAM+CMA Dec3 → Dec2 → Dec1 OUTPUT HEADS D₀ · R̂ · σ̂ RELIABILITY Stage CSPN++ REVISE Reliability-Gated w = 0.8·softmax( AffNet([u₁; R̂])) T=6 iterations Dirichlet anchors REVISE Stage FINAL Dense D₁ + EMA Inference ↑ Metric Depth ① REVEAL ② RELIABILITY ③ REVISE

Architecture Components

Each block is carefully designed to address a specific challenge in cross-domain depth completion with sparse inputs.

🌀

Deformable Conv v2 (DCNv2) in Depth Encoder

Unlike prior work that uses DCN only in propagation, R³DC embeds it directly in the depth encoder. Learned offsets Δpₖ and modulation mₖ align the receptive field with valid sparse points, mitigating artifacts at depth discontinuities.

🔗

Cross-Modal Attention (CMA) at 3 Scales

Depth features query RGB key/value via multi-head attention at stages ℓ ∈ {1,2,3}. GroupNorm stabilizes the query branch; logits clamped to [−8,8]. Memory bounded to O(Nmax²) via pooling to 512 tokens.

Transformer Bottleneck + CBAM

At 1/16 resolution: spatial transformer (8 heads, GELU MLP), followed by CBAM channel-and-spatial attention. Captures long-range dependencies while keeping attention tractable.

🔒

Reliability-Gated CSPN++ Propagation

AffNet takes [u₁; R̂] as input to predict 8 affinity weights per pixel. High-confidence pixels resist perturbation; low-confidence regions absorb spatial context. T=6 iterations with hard Dirichlet boundary conditions at sparse anchors.

🏗️

FPN Decoder with EfficientUpBlocks

4 blocks: transposed Conv4×4 upsample → DCN lateral skip → pre-activation ResBlock → CBAM → CMA with RGB skip. Preserves high-frequency depth boundaries lost in attention pooling.

🧭

Indoor Calibration Head (ICH)

For NYU Depth V2: frozen DA-V2 ViT-S backbone + 3-layer MLP ICH (16,642 params). Predicts dataset-specific shift & scale to convert unscaled relative priors to metric range [0,10] m.


Seven-Term Composite Loss

Multi-task depth learning faces inherent gradient conflicts. A balanced composite loss allows each objective to contribute its signal without domination.

ℒ = 1.00·ℒSILog + 0.60·ℒFocalBerHu + 0.20·ℒSSIM + 0.15·ℒAnchor + 0.10·ℒVNL + 0.05·(ℒDNC + ℒGrad + ℒUNC) + 0.10·ℒAux
1.00 SILog

Scale-Invariant Log Loss

Removes global depth bias via variance − scale terms. Critical for cross-domain transfer: KITTI (80 m) ↔ NYU (10 m).

0.60 FocalBerHu

Focal-BerHu (Hard Example Mining)

Focal weight (1−e−|e|)² continuously up-weights occlusion boundaries. γ=2 provides quadratic emphasis while preserving differentiability.

0.20 SSIM

Structural Similarity

Window size wₛ=7. Preserves edge consistency and local contrast structure beyond simple pixel-wise losses.

0.15 Anchor

Sparse Anchor Loss

Forces reproduction of known sensor measurements. Prevents metric-scale drift, critical at extreme sparsity (0.5–2%).

0.10 VNL

Virtual Normal Loss

Penalizes back-facing surface normals via random triplet sampling. Domain-agnostic: no camera intrinsics required.

0.05 UNC-NLL

Laplace Uncertainty NLL

Self-supervised uncertainty: |d̂−d|/σ̂ + ln σ̂. Penalizes confident-wrong predictions; prevents σ̂→0 collapse. No GT uncertainty labels needed.


RADI: Reliability-Aware Depth Index

Standard metrics answer "How accurate is the model on average?" RADI answers three complementary questions about trustworthiness. These questions are orthogonal — a model can pass one and fail the others.

Reliability–Error Correlation

Question: Does the model know where it is wrong?


Spearman rank correlation between predicted reliability R̂ and the negative absolute error. A positive ρ means high confidence predicts low error — the desired semantic behavior.

ρRECr = Spearman(R̂↾r, −|d̂−d|↾r) ∈ [−1, 1] ρ > 0 → high reliability ⟹ low error ✓ ρ = 0 → reliability is noise ρ < 0 → dangerous overconfidence ✗

Evaluated across 4 spatial regions: All · Edge (Sobel>0.05) · Textureless (luma σ<8) · Far-depth (d>0.75 dmax). Why Spearman? Invariant to absolute scale; robust to heavy-tailed depth error distributions.

Revision Benefit Score

Question: Does the architecture actually use its confidence during refinement?


High REC alone proves the model knows where it's wrong. RBS independently validates that reliability-gated CSPN++ actually improves depth by measuring normalized RMSE reduction from D₀ to D₁.

RBSr = (RMSE(D₀ʳ) − RMSE(D₁ʳ)) / RMSE(D₀ʳ) × 100% > 0 → refinement helps ✓ ≈ 0 → propagation redundant < 0 → miscalibrated affinities ✗

R³DC achieves RBS > 41% in all spatial regions on NYU Depth V2 (corrected metric-space values), confirming that confidence-gated propagation adds genuine value.

Calibration Error (ECE-style)

Question: Are the confidence values numerically calibrated?


Reliability values are grouped into B=15 equal-width bins. For bin b, the empirical accuracy āb is the fraction of pixels where relative error < τ=0.10. Perfect calibration → ECE=0.

ECE = Σb=1..15 (|b|/N) · |r̄b − āb| ECE = 0.000 → perfect calibration ✓ ECE ≈ 0.250 → random predictor ✗ R³DC ECE = 0.041 (NYU Depth V2)

A random uniform predictor yields ECE≈0.25. R³DC achieves ECE = 0.041, 6× better than the MC-Dropout baseline (0.118) and 83% better than random.

🎯 Why RADI is Necessary: A Motivating Example

Model A — R³DC
RMSE: 0.24 m
Well-calibrated confidence · ECE = 0.041
✓ Preferred for safety-critical deployment
Model B — Hypothetical
RMSE: 0.20 m
Overconfident in hard regions · ECE = 0.25
✗ Dangerous: wrong in reflective regions but claims high confidence

Standard RMSE selects Model B. RADI correctly surfaces that Model A is preferable for motion planning and obstacle avoidance: its reliability map reveals precisely where it cannot be trusted.


Competitive Accuracy, Minimal Parameters

R³DC is evaluated across four heterogeneous benchmarks with identical loss weights and optimizer settings — rigorously testing generalization, not benchmark overfitting.

KITTI · Depth Completion
RMSE · 1.95M params
NYU Depth V2 · Indoor
δ₁ accuracy
VisDrone · Aerial
RMSE · first public baseline
Drone-Videos · UAV
RMSE · 1.47M params

Table 1 · Unified Cross-Domain Results

MethodParams KITTI NYU Depth V2 VisDrone ††
RMSE↓δ₁↑ δ₁↑RMSE↓ RMSE↓δ₁↑
Depth Completion Methods
NLSPN26.8M0.762N/EN/EN/EN/EN/E
CSPN++17.4M0.744N/EN/EN/EN/EN/E
PENet131M0.730N/EN/EN/EN/EN/E
GuideFormer27.3M0.625N/EN/EN/EN/EN/E
CompletionFormer12.7M0.217N/EN/EN/EN/EN/E
BP-Net30.4M0.213N/EN/EN/EN/EN/E
Monocular Depth Estimation
AdaBins78.2M0.9020.288
DepthFormer34.3M0.9210.270
NeWCRFs270M0.9220.264
ZoeDepth345M0.9510.270
R³DC / R³DC+ — One Model Family, Four Domains
R³DC KITTI (ours) 1.95M 0.240 0.947
R³DC+ICH NYU (ours) 94.6M 0.927 0.353
R³DC+ v3 VisDrone (ours) 11.22M 2.33 0.928

†† Novel benchmarks: no prior depth completion methods report results on VisDrone or Drone-Videos. N/E = not evaluated in prior work. Bold = best; underline = second best.

Table 2 · RADI vs. Reliability Baselines (NYU Depth V2)

Depth Model Reliability Source REC ρ (All)↑ REC ρ (Edge)↑ REC ρ (Textureless)↑ RBS (%)↑ CAL ECE↓
Baselines (no revision stage — RBS=0 by construction)
CompletionFormerUniform (≡0.5)0.0000.0000.0000.0%0.248
CompletionFormerInv-gradient0.1470.1310.1690.0%0.183
CompletionFormerDepth-error proxy0.2140.1980.2310.0%0.142
CompletionFormerMC-Dropout0.2710.2490.2920.0%0.118
CompletionFormerEnsemble (3 seeds)0.3030.2810.3190.0%0.097
R³DC+ICH (ours) Learned R̂ 0.371 0.358 0.389 41.3% 0.041

R³DC's jointly trained reliability head achieves REC 43% higher than the depth-error proxy heuristic, confirming non-trivial calibration value beyond simple confidence estimates.


Every Component Earns its Place

All ablations run on KITTI (epoch 8, Val RMSE), modifying one factor at a time from the full R³DC configuration (B=64, 1.95M params).

Architecture Components (ΔRMSE from baseline)

Full R³DC (baseline)
0.240m
w/o CSPN++
+0.071m
w/o Cross-Modal Attn
+0.053m
w/o Deformable Conv
+0.033m
w/o Transformer Bottleneck
+0.021m
w/o EMA
+0.018m
Single-stream (depth only)
+0.101m
Single-stream (RGB only)
+0.138m

Loss Function Ablation (ΔRMSE)

Full composite loss
0.240m
SILog only
+0.071m
+Focal-BerHu
0.276m
+VNL
0.261m
+Aux supervision
0.252m
w/o VNL (from full)
+0.021m
w/o Sparse Anchor
+0.020m
Replace Focal→BerHu
+0.010m

Sparsity Robustness (Drone-Videos)

From 50% to 0.5% input density, RMSE increases by only 4.1× while density decreases by 100×, demonstrating sub-linear degradation.

0.5% density
2.31m
5% (train)
0.92m
50% density
0.56m

Reliability Maps Match Ground-Truth Error

The reliability maps consistently assign high confidence to textured surfaces and low confidence to reflective objects, occlusion boundaries, and far-range regions — validated by per-region RADI scores.

RGB Input
Sparse Depth
Predicted D₁
Reliability R̂
+0.371
REC ρ · All regions (p<0.001)
41.3%
RBS · D₀→D₁ RMSE improvement
0.041
ECE · vs. 0.25 random baseline

Honest Assessment

Synthetic Ground Truth for Aerial

VisDrone and Drone-Videos evaluations rely on physics-motivated synthetic depth prior. Proxy validation on DublinCity LiDAR yields Pearson r=0.81, but real aerial LiDAR is needed for full validation. RADI conclusions restricted to KITTI and NYU.

No Temporal Consistency

Current approach treats each frame independently. Online reliability adaptation or temporal smoothing could improve predictions for dynamic video sequences. Semi/self-supervised training on unlabeled frames is a natural extension.

Official KITTI Protocol Gap

Under the official benchmark (structured Velodyne input, mm units), R³DC (1.95M) scores 786.4 mm RMSE vs. CompletionFormer's 708.2 mm — not state-of-the-art. The principal contribution is 6–67× parameter reduction at comparable accuracy, plus per-pixel reliability absent in all baselines.

Implicit Reliability Supervision

The reliability head R̂ receives no direct GT supervision — learned implicitly through CSPN++ gating loss and shared features with σ̂. Adding a direct supervision signal (reliability cross-entropy) could accelerate calibration.


BibTeX

If you find R³DC or the RADI metric useful, please consider citing:

@inproceedings{mohammad2026r3dc, title = {R³DC: Reliability-Guided Reveal-to-Revise Depth Completion for Cross-Domain Sparse Perception}, author = {Mohammad, Noor Islam S. and Bayaz{\i}t, Ulu{\u{g}}}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)}, year = {2026}, note = {Workshop on 3D Scene Understanding from Incomplete/Ambiguous Data (3D4S)}, institution = {Istanbul Technical University}, }