DynAlign: Unsupervised Dynamic Taxonomy Alignment for Cross-Domain Segmentation
Authors: HAN SUN, Rui Gong, Ismail Nejjar, Olga Fink
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on the street scene semantic segmentation benchmarks GTA Mapillary Vistas and GTA IDD validate the effectiveness of our approach, achieving a significant improvement over existing methods. |
| Researcher Affiliation | Collaboration | Han Sun1 Rui Gong2 Ismail Nejjar1 Olga Fink1 1EPFL 2Amazon EMAIL EMAIL |
| Pseudocode | No | The paper describes the methodology in natural language and illustrates the framework with figures, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is publically available at https://github.com/hansunhayden/Dyn Align. |
| Open Datasets | Yes | Experiments on the street scene semantic segmentation benchmarks GTA Mapillary Vistas and GTA IDD validate the effectiveness of our approach... We evaluate our method using the synthetic dataset GTA as the source domain and two real-world datasets, Mapillary Vistas and India Driving Dataset (IDD), as target domains. GTA (Richter et al., 2016) is a synthetic dataset... Mapillary Vistas (Neuhold et al., 2017) contains 25k high-resolution images... IDD dataset (Varma et al., 2019), captured from Indian urban driving scenes... |
| Dataset Splits | Yes | For the unsupervised taxonomy-adaptive DA task, we use the GTA training set as the labeled source domain and the training sets of Mapillary Vistas and IDD as the unlabeled target domains. No target domain annotations are used during training, ensuring a fully unsupervised domain adaptation setup. ... Performance is reported on the validation set of each target dataset. ... The labeled GTA training set and unlabeled Mapillary Vistas/IDD training set are used to train the UDA Framework. |
| Hardware Specification | Yes | All experiments are conducted on NVIDIA A100-SXM4-80GB GPU. |
| Software Dependencies | No | The paper mentions several models and frameworks (e.g., GPT-4, SAM, CLIP, HRDA, Mask2Former) but does not provide specific version numbers for underlying software dependencies like programming languages (e.g., Python), libraries (e.g., PyTorch), or CUDA. |
| Experiment Setup | Yes | In our experiments, we develop an UDA model inspired by the architecture proposed by Hoyer et al. (2022b). ... we follow the default training parameters in Hoyer et al. (2022b). ... We set the confidence threshold to 0.5 by default for reassigning the class label. ... A self-training strategy with a teacher-student model is used, where the teacher generates pseudo-labels for the target domain, which are then weighted based on confidence estimates to account for uncertainty. These weighted pseudo-labels are used to further refine the model s performance on the target domain. The teacher model is updated using an exponential moving average (EMA) of the student model s weights, ensuring stable pseudo-labels over time. |