Research map / Architectures and efficiency / CNN design
Efficient Indoor Depth Completion Network Using Mask-adaptive Gated Convolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an efficient encoder-decoder indoor depth completion network built on Mask-adaptive Gated Convolution (MagaConv) and a Bi-directional Aligning Projection (Bid-AP) fusion module for RGB-guided depth completion.
Problem
Indoor depth completion with large missing regions, where standard and partial convolutions treat varying degrees of invalidity uniformly and naive RGB-depth fusion introduces depth-irrelevant features.
Approach
MagaConv uses a shared kernel applied to |W| and the iteratively updated mask to produce a per-position gating signal (via a Reverse-and-Cut activation) that selects between vanilla and partial-convolution-like behavior; Bid-AP performs bi-directional, channel-shared-MLP spatial-adaptive normalization between depth and color features.
Key results
- On NYU-Depth V2, the full model reaches RMSE 0.083 and Rel 0.011 with 30.1M parameters, beating TPVD (0.086) and AGG-Net (0.092).
- On DIML, RMSE 0.060 vs. 0.086 for AGG-Net, reported as about 20% improvement in RMSE.
- Ablation: replacing MagaConv with partial convolution raises RMSE from 0.109 to 0.134 on NYU-Depth V2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the proposed depth completion model | indoor depth completion · NYU-Depth V2 | RMSE: 0.086 -> 0.083 | Helped |
| use the proposed depth completion model | indoor depth completion · DIML | RMSE: 0.086 -> 0.060 | Helped |
| use the proposed depth completion model | indoor depth completion · SUN RGB-D | RMSE: 0.202 -> 0.197 | Helped |
| use MagaConv with three parallel operations per M-Layer rather than one 5 × 5 MagaConv (ablation) | MagaConv depth encoder · NYU-Depth V2 | RMSE: 0.114 -> 0.109 | Helped |
| replace Partial Convolution with MagaConv (ablation) | depth encoder · NYU-Depth V2 | RMSE: 0.134 -> 0.109 | Helped |
| replace Gated Convolution with MagaConv (ablation) | depth encoder · NYU-Depth V2 | RMSE: 0.127 -> 0.109 | Helped |
| use Bid-AP instead of the baseline's direct feature concatenation (ablation) | depth completion network · NYU-Depth V2 | RMSE: 0.188 -> 0.113 | Helped |
| use bidirectional alignment instead of unidirectional C → D (ablation) | Bid-AP · NYU-Depth V2 | RMSE: 0.139 -> 0.113 | Helped |
| include the channel-shared MLP in CMAP (ablation) | Bid-AP · NYU-Depth V2 | RMSE: 0.125 -> 0.113 | Helped |
| use CMAP rather than concatenation and convolution (ablation) | Bid-AP · NYU-Depth V2 | RMSE: 0.135 -> 0.113 | Helped |
| add Structure-Consistency loss to MSE loss (ablation) | depth completion network · NYU-Depth V2 | RMSE: 0.087 -> 0.083 | Helped |
| use MagaConv kernel sizes 3, 5, and 7 with 3 down-sampling layers (ablation) | MagaConv depth encoder · NYU-Depth V2 | RMSE | Helped |
Limitations
- Ablation and comparisons are limited to NYU-Depth V2 for most components; SUN RGB-D results are described as only competitive.
- Mask-update and gating rules are hand-designed heuristics with no theoretical analysis.
- Training was reported only for 100 epochs on a GTX 2080Ti, with limited hyperparameter details.
Experimental setup
Datasets: NYU-Depth V2 DIML SUN RGB-D
Benchmarks: NYU-Depth V2 DIML SUN RGB-D
Metrics: RMSE Rel threshold accuracy δ1.25 δ1.25^2 δ1.25^3
Baselines: CSPN++ NLSPN RDF-GAN GraphCSPN AGG-Net CFormer TPVD DfuseNet DM-LRN
Builds on: Partial Convolution Gated Convolution SPADE U-Net
Closest papers by meaning
- CompletionFormer: Depth Completion With Convolutions and Vision Transformers (CVPR 2023 · similarity 0.88)
- Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion (CVPR 2025 · similarity 0.86)
- DeCoTR: Enhancing Depth Completion with 2D and 3D Attentions (CVPR 2024 · similarity 0.85)
- Bilateral Propagation Network for Depth Completion (CVPR 2024 · similarity 0.85)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.84)
- Improving Depth Completion via Depth Feature Upsampling (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.