BECoTTA: Input-dependent Online Blending of Experts for Continual Test-time Adaptation
Authors: Daeun Lee, Jaehong Yoon, Sung Ju Hwang
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate that our method outperforms multiple CTTA scenarios, including disjoint and gradual domain shits, while only requiring 98% fewer trainable parameters. We also provide analyses of our method, including the construction of experts, the effect of domain-adaptive experts, and visualizations. Project Page: https://becottactta.github.io/. We compare our proposed method with strong baselines, including SAR (Niu et al., 2023), De PT (Gao et al., 2022), VDP (Gan et al., 2023), and Eco TTA (Song et al., 2023), on multiple CTTA scenarios and our suggested CGS benchmark. Our BECo TTA achieves +2.1%p and +1.7%p Io U enhancement respectively on CDS-Hard and CDS-Easy scenarios... |
| Researcher Affiliation | Collaboration | Daeun Lee * 1 Jaehong Yoon * 2 Sung Ju Hwang 3 4 ... 1Statistics, Korea University 2Computer Science, UNC-Chapel Hill 3Korea Advanced Institute of Science and Technology 4Deep Auto. Correspondence to: Daeun Lee <goodgpt@korea.ac.kr>, Jaehong Yoon <jhyoon@cs.unc.edu>, Sung Ju Hwang <sjhwang82@kaist.ac.kr>. |
| Pseudocode | Yes | To clarify our whole CTTA process, we provide the whole pipeline of BECo TTA at Algo. 1. According to the different initialization steps (Line 1 9), BECo TTA can be initialized in various ways including SDA or not. Algorithm 1 Continual Test-time Adaptation Pipeline Input: Source domain Xs, a sequence of target domains Xt = {X1 t , X2 t , . . .}, source model f, trainable parts of Mo DE W d g , W d noise, W down, W up, number of experts N, number of domain routers D. |
| Open Source Code | Yes | Project Page: https://becottactta.github.io/. For future reproducibility, we will publicly share the file list of our scenario. |
| Open Datasets | Yes | To reflect various domain shifts, we adopt balanced weather shifts (CDS-Easy) and imbalanced weather & area shifts (CDS-Hard) scenarios. For the CDS-Easy, we utilize the Cityscapes-ACDC setting used in previous work (Wang et al., 2022a): Cityscapes (Cordts et al., 2016) is used as the source domain, and ACDC (Sakaridis et al., 2021) as the target domain, consisting of four different weather types (fog, night, rain, snow). For the CDS-Hard, we propose a new imbalanced scenario considering both weather and geographical domain shifts. We also add clear and overcast weather from BDD-100k (Yu et al., 2020) to the existing target domain to mimic the real-world variety. |
| Dataset Splits | No | The paper mentions warming up the architecture for 10 epochs and using a 'training set' for target domains, but it does not specify a separate validation split or how hyperparameters were tuned using a validation set for reproducibility. |
| Hardware Specification | No | The paper does not specify the exact hardware (e.g., GPU model, CPU model, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'mmsegmentation framework' and refers to 'Table 7: Baseline method hyperparameters' and 'Table 8: Our method hyperparameters' in the appendix, but the provided text does not list specific version numbers for any software components. |
| Experiment Setup | Yes | We provide the implementation details utilized in our experiments in Tabs. 7 and 8. Table 7: Baseline method hyperparameters. Eco TTA (Song et al., 2023) K=4, λ=0.5, H0=0.4 Ours λd=0.1, λm=0.0005, κ=0.4. Table 8: Our method hyperparameters. Warm-up TTA Dataset SDA Target domains Optimizer Adam W Adam Optimizer momentum (β1, β2) = (0.9, 0.999) Epoch 10 Online Batch size 1 Learning rate 0.00006 0.00006/100 Label accessibility Yes No |