Research map / Continual learning / Domain and prompt-based
ProtoDepth: Unsupervised Continual Depth Completion with Prototypes
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 ProtoDepth, which freezes a pretrained unsupervised depth completion model and learns per-domain prototype sets (global multiplicative and local additive biases) plus domain descriptors for domain-agnostic selection, reducing forgetting in continual depth completion.
Problem
Unsupervised depth completion models catastrophically forget previously learned domains when continually trained on non-stationary sequences of datasets, especially when test-time domain identity is unknown.
Approach
Freeze the pretrained depth completion model and learn lightweight per-dataset prototype sets that act as global multiplicative (1x1 depthwise conv) and local attention-based additive biases on latent features; learn domain descriptors via a cosine-based objective to select the prototype set at test time when domain identity is withheld.
Key results
- ProtoDepth in the domain-incremental setting has zero forgetting, and ProtoDepth-A (agnostic) reduces average forgetting by 52.22% on indoor and 53.21% on outdoor sequences versus the best baseline, averaged across models and metrics.
- Incremental ProtoDepth improves average performance by 5.15% (indoor) and 6.88% (outdoor) over the best baseline.
- Each task adds under 5% of original model parameters (e.g., 0.25M, 3.6% for KBNet indoor).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ProtoDepth rather than the strongest baseline | VOICED · Indoor datasets (NYUv2 → ScanNet → VOID) | Average performance MAE (mm): 63.352 -> 56.359 | Helped |
| use ProtoDepth rather than the strongest baseline | FusionNet · Indoor datasets (NYUv2 → ScanNet → VOID) | Average performance MAE (mm): 61.204 -> 57.486 | Helped |
| use ProtoDepth rather than the strongest baseline | KBNet · Indoor datasets (NYUv2 → ScanNet → VOID) | Average performance MAE (mm): 56.208 -> 52.497 | Helped |
| use ProtoDepth rather than the strongest baseline | VOICED · Outdoor datasets (KITTI → Waymo → VKITTI) | Average performance MAE (mm): 524.114 -> 445.419 | Helped |
| use ProtoDepth rather than the strongest baseline | FusionNet · Outdoor datasets (KITTI → Waymo → VKITTI) | Average performance MAE (mm): 419.044 -> 400.888 | Helped |
| use ProtoDepth rather than the strongest baseline | KBNet · Outdoor datasets (KITTI → Waymo → VKITTI) | Average performance MAE (mm): 454.896 -> 401.075 | Helped |
| add global prototypes (ablation) | ProtoDepth with KBNet · ScanNet | MAE: 18.12 -> 14.59 | Helped |
| add projection matrix W (ablation) | ProtoDepth with KBNet · ScanNet | MAE: 17.59 -> 14.59 | Helped |
| use decoupled optimization of K and P (ablation) | ProtoDepth with KBNet · ScanNet | MAE: 16.36 -> 14.59 | Helped |
| add global prototypes (ablation) | ProtoDepth with KBNet · Waymo | MAE: 505.01 -> 486.95 | Helped |
| add projection matrix W (ablation) | ProtoDepth with KBNet · Waymo | MAE: 495.05 -> 486.95 | Helped |
| use decoupled optimization of K and P (ablation) | ProtoDepth with KBNet · Waymo | MAE: 491.59 -> 486.95 | Helped |
Surprising findings
- ProtoDepth-A sometimes outperforms the zero-forgetting incremental ProtoDepth on indoor metrics, which the authors attribute to appropriate prototype selection under domain overlap.
Limitations
- Requires known dataset boundaries to instantiate new prototype sets, which is infeasible for online training without defined domain boundaries.
- Does not consider cases where domain gaps are small or where there are significant shifts within a dataset.
- Domain-agnostic setting incurs some forgetting due to descriptor selection errors and domain overlap.
Experimental setup
Datasets: NYUv2 ScanNet VOID KITTI Waymo VKITTI
Benchmarks: NYUv2 ScanNet VOID KITTI Waymo VKITTI
Metrics: MAE RMSE iMAE iRMSE average forgetting average performance stability-plasticity trade-off (SPTO)
Baselines: Finetuned EWC LwF Experience Replay UnCLe
Builds on: UnCLe benchmark scaled dot-product attention EWC Learning without Forgetting Experience Replay
Closest papers by meaning
- SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation (CVPR 2025 · similarity 0.86)
- Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior (AAAI 2025 · similarity 0.85)
- DinoComplete: 3D Shape Completion with Distilled Semantic Priors and State Space Models (NeurIPS 2026 · similarity 0.85)
- GenPC: Zero-shot Point Cloud Completion via 3D Generative Priors (CVPR 2025 · similarity 0.85)
- Test-Time Adaptation for Depth Completion (CVPR 2024 · similarity 0.85)
- Distilling Monocular Foundation Model for Fine-grained Depth Completion (CVPR 2025 · 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.