Research map / Architectures and efficiency / CNN design
CompletionFormer: Depth Completion With Convolutions and Vision Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine convolutional attention and Transformer features in parallel rather than cascaded streams (ablation) | CompletionFormer JCAT · NYUv2, 500 samples | RMSE (mm): 91.5 -> 90.0 | Helped |
| add spatial and channel attention to the JCAT convolutional path (ablation) | CompletionFormer JCAT · NYUv2, 500 samples | RMSE (mm): 91.1 -> 90.0 | Helped |
| embed RGB and depth early in a single branch instead of separate encoder branches (ablation) | CompletionFormer · NYUv2, 500 samples | RMSE (mm): 94.0 -> 90.0 | Helped |
| add spatial and channel attention to multiscale decoder feature fusion (ablation) | CompletionFormer decoder · NYUv2 | RMSE (mm): 92.3 -> 91.4 | Helped |
| use 6 rather than 18 SPN refinement iterations (ablation) | CompletionFormer · NYUv2 | RMSE (mm): 90.1 -> 90.0 | No clear effect |
| replace non-local NLSPN refinement with fixed-local CSPN++ (ablation) | CompletionFormer · NYUv2 | RMSE (mm): 90.0 -> 90.3 | No clear effect |
| use only Transformer layers rather than the complete convolution-Transformer block (ablation) | CompletionFormer · KITTI DC, 1 scanning line | RMSE (mm): 3507.2 -> 3250.2 | Hurt |
| combine convolutional and Transformer features rather than use a pure Transformer variant (ablation) | CompletionFormer · NYUv2, 500 samples | RMSE (m): 0.091 -> 0.090 | Helped |
| train with L1 rather than combined L1 and L2 loss (ablation) | CompletionFormer · KITTI DC | MAE (mm): 203.45 -> 183.88 | Helped |
| train with combined L1 and L2 rather than L1 loss alone (ablation) | CompletionFormer · KITTI DC | RMSE (mm): 764.87 -> 708.87 | Helped |
Closest papers by meaning
- Bilateral Propagation Network for Depth Completion (CVPR 2024 · similarity 0.89)
- DeCoTR: Enhancing Depth Completion with 2D and 3D Attentions (CVPR 2024 · similarity 0.89)
- Efficient Indoor Depth Completion Network Using Mask-adaptive Gated Convolution (AAAI 2025 · similarity 0.88)
- Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion (CVPR 2025 · similarity 0.88)
- Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior (AAAI 2025 · similarity 0.85)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.85)
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.