Research map / Label-efficient and robust learning / Test-time adaptation
Backpropagation-free Network for 3D Test-time Adaptation
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 BFTT3D, a backpropagation-free two-stream test-time adaptation method for 3D point cloud classification that combines a frozen source model with a non-parametric prototype-matching branch aligned by TCA and fused via entropy-based weights.
Problem
Test-time adaptation for 3D point cloud classification under distribution shift, where existing backpropagation-based methods suffer from forgetting, error accumulation from pseudo-labels, and high compute/memory cost.
Approach
Keeps the source model frozen and adds a training-free target-specific branch built on a non-parametric point cloud network (Point-NN) that matches test features to a static source prototype memory in a TCA-learned shared subspace; the two logits are fused with an entropy-derived adaptive weight, so no parameters are updated by backpropagation at test time.
Key results
- On ScanObjectNN-C with CurveNet, BFTT3D reduces mean error from 28.47% (source-only) to 19.23%, the best among compared methods including SHOT (19.79%).
- On ModelNet-40C with PointNet, BFTT3D achieves mean error 33.78% vs 35.88% source-only and 34.25% for SHOT.
- On ScanObjectNN-C with PointNet, BFTT3D lowers mean error from 59.35% to 54.46%, with the largest gains on occlusion and lidar corruptions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BFTT3D instead of SHOT | PointNet · ModelNet-40C | classification error (%): 34.25 -> 33.78 | Helped |
| use BFTT3D instead of SHOT | DGCNN · ModelNet-40C | classification error (%): 30.17 -> 29.33 | Helped |
| use BFTT3D instead of SHOT | Curvenet · ModelNet-40C | classification error (%): 19.79 -> 19.23 | Helped |
| use BFTT3D instead of LAME | PointNet · ScanObjectNN-C | classification error (%): 55.67 -> 54.46 | Helped |
| use BFTT3D instead of BN | DGCNN · ScanObjectNN-C | classification error (%): 45.11 -> 43.91 | Helped |
| use BFTT3D instead of TENT | Curvenet · ScanObjectNN-C | classification error (%): 65.77 -> 57.80 | Helped |
| use 100% rather than 25% of source features as prototype memory (ablation) | BFTT3D · ScanObjectNN-C | classification error (%): 43.91 -> 43.75 | Helped |
| use 75% rather than 25% of source features as prototype memory (ablation) | BFTT3D · ScanObjectNN-C | classification error (%): 43.91 -> 43.80 | Helped |
| use 50% rather than 25% of source features as prototype memory (ablation) | BFTT3D · ScanObjectNN-C | classification error (%): 43.91 -> 43.91 | No clear effect |
| add TCA subspace learning instead of no subspace learning (ablation) | BFTT3D · ScanObjectNN-C | classification error (%): 56.84 -> 54.46 | Helped |
| use JDA instead of TCA for subspace learning (ablation) | BFTT3D · ScanObjectNN-C | classification error (%): 54.46 -> 54.87 | Hurt |
| use CORAL instead of TCA for subspace learning (ablation) | BFTT3D · ScanObjectNN-C | classification error (%): 54.46 -> 56.80 | Hurt |
Surprising findings
- The LAME baseline degrades strongly on several corruptions (e.g., 83.16% mean error on ModelNet-40C with PointNet vs 35.88% source-only).
- On ScanObjectNN-C with DGCNN, SHOT is worse than source-only (57.53% vs 46.09%), while BFTT3D reaches 43.91%.
Limitations
- Relies on a non-parametric source-prototype memory, which requires access to source features before test time.
- Gains on some corruptions are small or negative (e.g., background and rot on some backbones compared with TENT/SHOT).
- Evaluated only on point cloud classification with synthetic corruptions.
Experimental setup
Datasets: ModelNet-40 ScanObjectNN
Benchmarks: ModelNet-40C ScanObjectNN-C
Metrics: mean classification error (%)
Baselines: TENT BN adaptation SHOT LAME DSS MATE
Builds on: Point-NN TCA TENT SHOT
Closest papers by meaning
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.89)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.88)
- Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization (AAAI 2024 · similarity 0.88)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.87)
- Test-time Correlation Alignment (ICML 2025 · similarity 0.87)
- Test-Time Model Adaptation with Only Forward Passes (ICML 2024 · similarity 0.87)
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.