Research map / Continual learning / Domain and prompt-based
DPCore: Dynamic Prompt Coreset 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 DPCore, a continual test-time adaptation method that learns visual prompts per domain group and manages them in a dynamic prompt coreset, targeting a new Continual Dynamic Change setting where domains recur with varying frequency and duration.
Problem
Continual test-time adaptation degrades when domains recur irregularly with varying durations (the CDC setting), causing convergence issues with brief exposures, catastrophic forgetting, and negative transfer across dissimilar domains.
Approach
Learn visual prompts from scratch by aligning batch feature statistics with source statistics, store (prompt, statistics) pairs in a coreset, and at each batch either refine a distance-weighted composition of existing prompts (similar domain) or add a new prompt (dissimilar domain), so only a small set of prompts is trained and kept.
Key results
- On ImageNet-to-ImageNet-C in the CSC setting, DPCore reaches 39.9% mean error (+15.9% over source), versus 43.4% for ViDA.
- In the CDC setting on ImageNet-C, DPCore's error rises only from 39.9% to 42.1%, while ViDA rises from 43.4% to 52.1%.
- DPCore uses 0.08M trainable parameters versus 93.70M for ViDA and runs at 1.8x Tent's time versus 2.8x for ViDA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DPCore instead of the source model | ViT-Base · ImageNet-to-ImageNet-C CSC | error rate: 55.8 -> 39.9 | Helped |
| Use DPCore instead of ViDA | ViT-Base · ImageNet-to-ImageNet-C CDC | error rate: 52.1 -> 42.1 | Helped |
| Use DPCore instead of ViDA | Segformer-B5 · Cityscapes-to-ACDC CSC | mIoU: 61.9 -> 62.5 | Helped |
| Use DPCore instead of ViDA | Segformer-B5 · Cityscapes-to-ACDC CDC | mIoU: 59.2 -> 61.0 | Helped |
| Use DPCore instead of ViDA | CIFAR10 model · CIFAR10-to-CIFAR10C CSC | error rate: 20.7 -> 15.4 | Helped |
| Use DPCore instead of ViDA | CIFAR10 model · CIFAR10-to-CIFAR10C CDC | error rate: 23.6 -> 17.1 | Helped |
| Add Prompt Coreset to VPA and Dynamic Update (ablation) | DPCore · ImageNet-to-ImageNet-C CSC | error rate: 50.8 -> 39.9 | Helped |
| Add Dynamic Update to VPA and Prompt Coreset (ablation) | DPCore · ImageNet-to-ImageNet-C CSC | error rate: 48.3 -> 39.9 | Helped |
| Add Visual Prompt Adaptation instead of using NormLayer parameters (ablation) | DPCore with Prompt Coreset and Dynamic Update · ImageNet-to-ImageNet-C CSC | error rate: 45.1 -> 39.9 | Helped |
| Replace flexible coreset management with a fixed K=15 coreset that discards the oldest prompt (ablation) | DPCore · ImageNet-to-ImageNet-C CSC, 10 domain orders | error rate: 40.2 -> 42.3 | Hurt |
| Replace flexible coreset management with a fixed K=15 coreset that merges the most similar prompts (ablation) | DPCore · ImageNet-to-ImageNet-C CSC, 10 domain orders | error rate: 40.2 -> 41.1 | Hurt |
| Transfer a Gaussian Noise-learned prompt to other corruption domains | ImageNet-to-ImageNet-C · Shot Noise | error rate: 51.8 -> 40.4 | Helped |
Surprising findings
- A single prompt learned on Gaussian Noise transfers well to Shot and Impulse Noise but increases error on Contrast (91.4% to 95.7%), showing negative transfer.
- Tent performs worse than the source model in the CDC setting on ImageNet-C (57.1% vs 55.8%).
Limitations
- Relies on source statistics computed from 300 source examples, so source data access is still required, though a STL10 fallback is evaluated.
- Coreset size grows with unseen domains and the theoretical analysis uses simplified well-separated cluster assumptions.
Experimental setup
Datasets: ImageNet-to-ImageNet-C CIFAR10-to-CIFAR10C CIFAR100-to-CIFAR100C Cityscapes-to-ACDC STL10
Benchmarks: ImageNet-to-ImageNet-C CIFAR10-to-CIFAR10C CIFAR100-to-CIFAR100C Cityscapes-to-ACDC
Metrics: classification error rate mIoU trainable parameters relative computation time
Baselines: Tent SAR CoTTA VDP DePT RoTTA EcoTTA ViDA BDG C-MAE RDumb
Builds on: Tent CoTTA ViDA Visual Prompt Tuning Online K-Means
Closest papers by meaning
- BECoTTA: Input-dependent Online Blending of Experts for Continual Test-time Adaptation (ICML 2024 · similarity 0.90)
- ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptation (ICLR 2024 · similarity 0.90)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.90)
- A Versatile Framework for Continual Test-Time Domain Adaptation: Balancing Discriminability and Generalizability (CVPR 2024 · similarity 0.90)
- Class-aware Domain Knowledge Fusion and Fission for Continual Test-Time Adaptation (NeurIPS 2025 · similarity 0.89)
- PALM: Pushing Adaptive Learning Rate Mechanisms for Continual Test-Time Adaptation (AAAI 2025 · 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.