Research map / Label-efficient and robust learning / Test-time adaptation
Test-Time Adaptation for Depth Completion
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 ProxyTTA, an online single-pass test-time adaptation method for depth completion that learns a source-trained proxy mapping from sparse-depth features to image+depth features and uses it to guide updates of a small adaptation layer.
Problem
Depth completion models degrade when transferred to a target domain with covariate shift, and existing adaptation methods need source data or multiple passes over test data; the goal is single-pass, source-free test-time adaptation for depth completion.
Approach
Sensitivity analysis shows sparse depth shifts less than RGB across domains. A proxy embedding module, trained on source data, maps sparse-depth-only features to image+depth features; at test time these proxy embeddings are used as targets to align target RGB features by updating only a convolutional adaptation layer in the image branch.
Key results
- ProxyTTA improves over the BN Adapt and CoTTA baselines by an average of 21.09% across methods and datasets.
- Updating only the adaptation layer (ProxyTTA-fast) improves over BN Adapt with l_z, l_sm by 34.07% vs 32.77% average improvement over pretrained, and full ProxyTTA with BN update improves the pretrained model by 44.53%.
- Sensitivity study shows sparse-depth-only inference in the target domain often outperforms image+depth inference, indicating depth is less sensitive to domain shift than image.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ProxyTTA instead of BN Adapt, ℓz, ℓsm and CoTTA | depth completion models · all reported indoor and outdoor datasets | average improvement: 21.09% | Helped |
| use ProxyTTA instead of BN Adapt, ℓz, ℓsm | depth completion models · indoor datasets | average improvement: 11.60% | Helped |
| use ProxyTTA instead of BN Adapt, ℓz, ℓsm | depth completion models · outdoor datasets | average improvement: 19.73% | Helped |
| use ProxyTTA-fast without batch norm update instead of BN Adapt | depth completion models · all reported datasets | average improvement: 34.07% | Helped |
| add batch norm updates to ProxyTTA-fast (ablation) | ProxyTTA · all reported datasets | average improvement over pretrained model: 44.53% | Helped |
| remove batch norm updates from ProxyTTA (ablation) | ProxyTTA · all reported datasets | adaptation runtime: 25.32% | Hurt |
| replace proxy loss with L1 consistency loss to pretrained prediction (ablation) | CoTTA · all reported datasets | depth completion error: 25.26% average improvement for our method without proxy over CoTTA | Hurt |
| use sparse depth only instead of image plus sparse depth | pretrained depth completion models · target domains, including NYUv2, ScanNet, Waymo, and nuScenes | error | Mixed |
| use image plus sparse depth instead of sparse depth only | pretrained depth completion models · source domain | error | Helped |
| remove the image input (ablation) | pretrained depth completion models · target domain | RMSE: 1528.98 -> 1046.28 | Helped |
Surprising findings
- In the target domain, sparse depth alone often yields lower error than sparse depth plus image, while the opposite holds in the source domain.
- CoTTA alone could not adapt depth completion models, requiring added sparse-depth and smoothness losses.
Limitations
- Gains are expected to be small when the source dataset shares no objects with the target test set.
- Assumes access to the source dataset before deployment for initialization and proxy training.
- Evaluated only on supervised pretrained depth completion models in controlled experiments.
Experimental setup
Datasets: VOID KITTI NYUv2 SceneNet ScanNet Virtual KITTI nuScenes Waymo Open Dataset
Benchmarks: VOID to NYUv2 VOID to SceneNet VOID to ScanNet KITTI to VKITTI-FOG KITTI to nuScenes KITTI to Waymo
Metrics: MAE RMSE
Baselines: BN Adapt BN Adapt with l_z and l_sm CoTTA Pretrained source model
Builds on: TENT BYOL CoTTA TTT++
Closest papers by meaning
- Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior (AAAI 2025 · similarity 0.88)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.87)
- Flexible Depth Completion for Sparse and Varying Point Densities (CVPR 2024 · similarity 0.85)
- Improving Depth Completion via Depth Feature Upsampling (CVPR 2024 · similarity 0.85)
- Test3R: Learning to Reconstruct 3D at Test Time (NeurIPS 2025 · similarity 0.85)
- ProtoDepth: Unsupervised Continual Depth Completion with Prototypes (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.