Research map / Continual learning / Domain and prompt-based
BECoTTA: Input-dependent Online Blending of Experts for Continual 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 BECoTTA, a continual test-time adaptation method that places input-routed low-rank Mixture-of-Domain experts (MoDE) in each ViT block and adds a domain-expert synergy loss, and also introduces a gradual-shift CTTA benchmark (CGS).
Problem
Continual test-time adaptation (CTTA) for semantic segmentation and classification must balance forgetting and adaptation while remaining computationally and memory efficient, and existing benchmarks mostly assume disjoint domain shifts rather than gradual ones.
Approach
Mixture-of-Domain low-rank experts with multiple domain routers (one per pseudo-domain from a domain discriminator) so only a few domain-relevant experts are updated per input, plus a mutual-information-style domain-expert synergy loss; a new CGS benchmark with gradually shifting weather distributions.
Key results
- On CDS-Hard, BECoTTA (L) reaches 51.3 avg mIoU over 10 rounds vs. 50.2 for CoTTA (using about 17x more parameters), and BECoTTA+ (L) reaches 54.2 with SDA.
- On CDS-Easy, BECoTTA+ (L) with SDA reaches 63.0 mean mIoU vs. 58.5 for CoTTA and 55.8 for EcoTTA-style baselines (round-averaged table).
- On CGS, BECoTTA+ (L) with SDA reaches 60.23 mean mIoU vs. 57.04 for EcoTTA, and parameter count stays around 3M.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use BECoTTA (L) instead of CoTTA | CTTA · CDS-Hard, round 10 | mIoU: +2.1%p | Helped |
| Use BECoTTA+ (L) instead of EcoTTA | CTTA · CDS-Hard, w/ SDA | mIoU: +16.8%p | Helped |
| Use BECoTTA (S) instead of EcoTTA | CTTA · CDS-Easy, w/o SDA | mIoU: +5%p | Helped |
| Use BECoTTA (L) instead of EcoTTA | CTTA · CGS | mIoU: +5.5%p | Helped |
| Use BECoTTA instead of EcoTTA | CIFAR100-C classification · CIFAR100-C, severity 5 | average error rate: -2.47%p | Helped |
| Add source domain augmentation (SDA) (ablation) | BECoTTA (S) · CDS-Easy | mIoU: 58.6 -> 60.2 | Helped |
| Add source domain augmentation (SDA) (ablation) | BECoTTA (M) · CGS | mIoU: 55.60 -> 59.13 | Helped |
| Add source domain augmentation (SDA) (ablation) | BECoTTA (L) · CGS | mIoU: 57.13 -> 60.23 | Helped |
| Use SDA with source and Bright domains rather than no SDA (ablation) | BECoTTA+ · CDS-Hard, A-Night | IoU: +8.5%p | Helped |
| Use DD alone rather than source-only (ablation) | BECoTTA · Component ablation | IoU: +0.04%p | Helped |
| Add MoDE (ablation) | BECoTTA · Component ablation | IoU: +5.36%p | Helped |
| Add domain-expert synergy loss (ablation) | BECoTTA · Component ablation | IoU: +6.12%p | Helped |
Surprising findings
- Domain-augmented initialization with domains that do not match the target domains still improves adaptation, suggesting routers are differentiated regardless of SDA-target relevance.
- TENT and SAR degrade substantially over rounds on CDS-Hard (TENT drops from 47.9 to 34.6 mIoU), showing severe forgetting of fully domain-agnostic entropy minimization.
Limitations
- Relies on a pseudo-domain discriminator trained on pre-defined SDA domains; routing quality depends on the discriminator.
- Requires a warm-up/initialization phase on source data before deployment, with source-free assumption only after initialization.
Experimental setup
Datasets: Cityscapes ACDC BDD-100k CIFAR10-C CIFAR100-C GTAV Synthia Mapillary
Benchmarks: CDS-Easy (Cityscapes-to-ACDC) CDS-Hard Continual Gradual Shifts (CGS) CIFAR10-to-CIFAR10C CIFAR100-to-CIFAR100C zero-shot DG (BDD100k, Mapillary, GTAV, Synthia)
Metrics: mIoU classification error rate parameter count
Baselines: TENT CoTTA SAR EcoTTA DePT VDP TTN BN Stats Adapt TTT++ EATA NOTE SWR
Builds on: EcoTTA CoTTA TENT sparse mixture-of-experts with top-k routing SAR
Closest papers by meaning
- DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptation (ICML 2025 · similarity 0.90)
- Class-aware Domain Knowledge Fusion and Fission for Continual Test-Time Adaptation (NeurIPS 2025 · similarity 0.89)
- A Versatile Framework for Continual Test-Time Domain Adaptation: Balancing Discriminability and Generalizability (CVPR 2024 · similarity 0.88)
- ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptation (ICLR 2024 · similarity 0.88)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.87)
- Active Test-Time Adaptation: Theoretical Analyses and An Algorithm (ICLR 2024 · similarity 0.86)
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.