Research map / Architectures and efficiency / CNN design
Improving Depth Completion via Depth Feature Upsampling
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.
Visualizes encoder-decoder depth completion networks and proposes DFU, a plug-and-play depth feature upsampling module guided by dense ED-Net features via confidence-aware guidance with adaptive receptive fields, improving single-branch, dual-branch, and SPN-based depth completion on KITTI and NYUv2.
Problem
Encoder-decoder networks for depth completion are treated as black boxes; decoder features become sparse after fusing dense features, causing information loss during the "dense-to-sparse" process.
Approach
Instead of using the sparse decoder feature, upsample a low-resolution depth feature to high resolution under guidance from the dense fused encoder-decoder features, using a confidence map to filter unreliable values and guidance with multi-scale Fourier-Bessel adaptive receptive fields.
Key results
- Improved LRRU-Base with three-layer DFU achieves 686.46 mm RMSE on the KITTI online benchmark, ranked Top-1 (vs 696.51 mm for LRRU-Base).
- One-layer DFU reduces single-branch ED-Net KITTI validation RMSE from 772.87 mm to 750.58 mm (Table 1 reports 750.58 for one-layer DFU) at +1.67M params and +43.79G FLOPs.
- DFU yields larger gains on sparser depth (16/32-line LiDAR) than on 64-line depth, indicating improved generalization.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add three-layer DFU to LRRU-Base | LRRU-Base · KITTI online benchmark | RMSE [mm]: 696.51 -> 686.46 | Helped |
| Add three-layer DFU | Our Single-branch ED-Net · KITTI online benchmark | RMSE [mm]: 745.16 -> 719.65 | Helped |
| Add three-layer DFU | Our Dual-branch ED-Net · KITTI online benchmark | RMSE [mm]: 720.96 -> 706.23 | Helped |
| Add three-layer DFU (ablation) | Our single-branch ED-Net · KITTI Validate Dataset | RMSE [mm]: 772.87 -> 746.98 | Helped |
| Increase DFU depth from one to two layers (ablation) | Our single-branch ED-Net · KITTI Validate Dataset | RMSE [mm]: 750.58 -> 745.27 | Helped |
| Increase DFU depth from two to three layers (ablation) | Our single-branch ED-Net · KITTI Validate Dataset | RMSE [mm]: 745.27 -> 746.98 | Hurt |
| Add one-layer DFU (ablation) | Our Dual-branch ED-Net · KITTI Validate Dataset | RMSE [mm]: 753.81 -> 741.39 | Helped |
| Add three-layer DFU (ablation) | LRRU-Mini · KITTI Validate Dataset | RMSE [mm]: 806.27 -> 767.67 | Helped |
| Add the confidence-aware module and GARF (ablation) | One-layer DFU on Our single-branch ED-Net · KITTI validation | RMSE [mm]: 761.24 -> 751.76 | Helped |
| Add the confidence-aware module (ablation) | One-layer DFU on Our single-branch ED-Net · KITTI validation | RMSE [mm]: 758.67 -> 751.76 | Helped |
| Add GARF (ablation) | One-layer DFU on Our single-branch ED-Net · KITTI validation | RMSE [mm]: 756.31 -> 751.76 | Helped |
| Add three-layer DFU to LRRU-Base | LRRU-Base · NYUv2 | RMSE [m]: 0.091 -> 0.091 | No clear effect |
Surprising findings
- Improvement on NYUv2 for the LRRU-Base model is negligible, attributed to overfitting on the small dataset.
- Performance gains saturate beyond two or three DFU layers.
Limitations
- Gains on LRRU-Base are negligible on NYUv2 due to overfitting with limited samples.
- Performance improvement saturates as the number of DFU layers increases.
Experimental setup
Datasets: KITTI depth completion NYUv2
Benchmarks: KITTI depth completion online leaderboard KITTI validation NYUv2
Metrics: RMSE MAE iRMSE iMAE REL delta1.25
Baselines: S2D TWISE FCFRNet PENet ACMNet PointFusion GFormer DySPN GraphCSPN RigNet CompletionFormer BEV@DC
Builds on: CSPN GuideNet LRRU S2D CompletionFormer
Closest papers by meaning
- DeCoTR: Enhancing Depth Completion with 2D and 3D Attentions (CVPR 2024 · similarity 0.88)
- Decoupling Fine Detail and Global Geometry for Compressed Depth Map Super-Resolution (CVPR 2025 · similarity 0.87)
- Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior (AAAI 2025 · similarity 0.86)
- Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion (CVPR 2025 · similarity 0.86)
- DegBins: Degradation-Driven Binning for Depth Super-Resolution (NeurIPS 2026 · similarity 0.86)
- Masked Spatial Propagation Network for Sparsity-Adaptive Depth Refinement (CVPR 2024 · similarity 0.86)
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.