Research map / Label-efficient and robust learning / Test-time adaptation
RobIA: Robust Instance-aware Continual Test-time Adaptation for Deep Stereo
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 RobIA, a continual test-time adaptation framework for stereo depth estimation combining a row-wise self-attention mixture-of-experts (AttEx-MoE) on frozen CoEx with an AdaptBN teacher that densifies sparse SGM pseudo-labels.
Problem
Stereo depth estimation under continually shifting target domains suffers from catastrophic forgetting, error accumulation, and sparse, unreliable handcrafted pseudo-labels, with input-invariant PEFT adapters limiting adaptation.
Approach
An instance-aware MoE adapter for CNN stereo networks whose channel-expert gates are produced by row-wise (epipolar-line) self-attention with sigmoid soft gating, combined with a dense pseudo-label scheme where an AdaptBN-tuned teacher fills in regions where SGM proxy labels are filtered out.
Key results
- On the DrivingStereo CTTA benchmark, AttEx-MoE + AT reaches 2.77% mean D1-all and 0.91 px EPE, versus 3.08% for AdaptBN and 3.04% for full fine-tuning on CoEx.
- On DSEC CTTA, the FT + AT baseline reaches 4.38% mean D1-all, while AttEx-MoE alone reaches 4.47%, and the AttEx-MoE + AT variant reaches 4.46%.
- AttEx-MoE uses about 1.19M trainable parameters versus 2.73M for full fine-tuning on CoEx, with runtime of 97 ms versus 255 ms for FT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the AdaptBN teacher for dense pseudo-supervision (ablation) | AttEx-MoE on DrivingStereo CTTA · DrivingStereo CTTA 10 rounds | D1-all: 2.98 -> 2.77 | Helped |
| add the AdaptBN teacher for dense pseudo-supervision | AttEx-MoE on DSEC CTTA · DSEC CTTA 10 rounds | D1-all: 4.47 -> 4.46 | Mixed |
| add the AdaptBN teacher for dense pseudo-supervision | AttEx-MoE on KITTI RAW CTTA · KITTI RAW CTTA 10 rounds | D1-all: 1.06 -> 1.09 | Hurt |
| replace row-wise self-attention with a shallow embedding router (ablation) | AttEx-MoE on DrivingStereo · DrivingStereo CTTA 10 rounds | D1-all: 2.98 -> 3.16 | Hurt |
| replace row-wise self-attention with GAP routing (ablation) | AttEx-MoE on DrivingStereo · DrivingStereo CTTA 10 rounds | D1-all: 2.98 -> 4.06 | Hurt |
| replace row-wise self-attention with column-wise attention (ablation) | AttEx-MoE on DrivingStereo · DrivingStereo CTTA 10 rounds | D1-all: 2.98 -> 4.11 | Hurt |
| replace Sigmoid gating with ReLU gating (ablation) | AttEx-MoE with row-wise attention on DrivingStereo · DrivingStereo CTTA 10 rounds | D1-all: 2.98 -> 3.77 | Hurt |
| set the teacher loss weight λ to 0.05 instead of 0.1 (ablation) | AttEx-MoE + AT on DrivingStereo · DrivingStereo CTTA 10 rounds | D1-all: 2.77 -> 2.97 | Hurt |
| increase the teacher loss weight λ to 0.2 instead of 0.1 (ablation) | AttEx-MoE + AT on DrivingStereo · DrivingStereo CTTA 10 rounds | D1-all: 2.77 -> 2.84 | Hurt |
| increase the teacher loss weight λ to 0.3 instead of 0.1 (ablation) | AttEx-MoE + AT on DrivingStereo · DrivingStereo CTTA 10 rounds | D1-all: 2.77 -> 2.89 | Hurt |
| replace proxy-only supervision with photometric-only supervision (ablation) | AttEx-MoE on DrivingStereo · DrivingStereo CTTA 10 rounds | D1-all: 2.98 -> 3.55 | Hurt |
| replace proxy-only supervision with AdaptBN-only supervision (ablation) | AttEx-MoE on DrivingStereo · DrivingStereo CTTA 10 rounds | D1-all: 2.98 -> 4.16 | Hurt |
Surprising findings
- Full fine-tuning (FT) achieves strong early-round gains but degrades over later rounds due to error accumulation and forgetting, while PEFT methods remain more stable.
- ReLU gating underperforms sigmoid gating across all router types in the frozen-backbone PEFT setting.
Limitations
- Reliance on predefined expert structures may limit flexibility in highly heterogeneous environments.
- Benchmark uses short sequences of 500 frames per domain with 10 rounds, which differs from prior long-sequence protocols.
Experimental setup
Datasets: SceneFlow FlyingThings3D KITTI RAW DrivingStereo DSEC
Benchmarks: DrivingStereo DSEC KITTI RAW
Metrics: D1-all EPE
Baselines: MADNet 2 MAD++ CoEx no adapt. AdaptBN FT EcoTTA (MetaNet) EMA teacher
Builds on: AdaptBN DeepMoE EcoTTA CoEx MADNet 2
Closest papers by meaning
- PPMStereo: Pick-and-Play Memory Construction for Consistent Dynamic Stereo Matching (NeurIPS 2025 · similarity 0.87)
- Rectification-specific Supervision and Constrained Estimator for Online Stereo Rectification (CVPR 2025 · similarity 0.86)
- Stereo Anywhere: Robust Zero-Shot Deep Stereo Matching Even Where Either Stereo or Mono Fail (CVPR 2025 · similarity 0.86)
- PPEA-Depth: Progressive Parameter-Efficient Adaptation for Self-Supervised Monocular Depth Estimation (AAAI 2024 · similarity 0.85)
- DEFOM-Stereo: Depth Foundation Model Based Stereo Matching (CVPR 2025 · similarity 0.85)
- Stereo Risk: A Continuous Modeling Approach to Stereo Matching (ICML 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.