Adversarial Style Augmentation for Domain Generalized Urban-Scene Segmentation
Authors: Zhun Zhong, Yuyang Zhao, Gim Hee Lee, Nicu Sebe
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on two synthetic-to-real semantic segmentation benchmarks demonstrate that Adv Style can significantly improve the model performance on unseen real domains and show that we can achieve the state of the art. |
| Researcher Affiliation | Academia | Zhun Zhong1 Yuyang Zhao2 Gim Hee Lee2 Nicu Sebe1 1 Department of Information Engineering and Computer Science, University of Trento 2 Department of Computer Science, National University of Singapore |
| Pseudocode | Yes | The detailed training procedure and Pytorch-like pseudo-code can be found in the Appendix. |
| Open Source Code | Yes | Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] See the Appendix. |
| Open Datasets | Yes | For the synthetic-to-real domain generalization (DG), we use one of the synthetic datasets (GTAV [28] or SYNTHIA [29]) as the source domain and evaluate the model performance on three real-world datasets (City Scapes [8], BDD-100K [39], and Mapillary [23]). |
| Dataset Splits | Yes | Following [7], the model obtained by the last training iteration is used to evaluate the m Io U performance on the three real-world validation sets. |
| Hardware Specification | No | The checklist indicates: 'Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [No]'. |
| Software Dependencies | No | The paper mentions 'Pytorch-like pseudo-code' but does not specify exact software versions or dependencies like Python, PyTorch, or CUDA versions. |
| Experiment Setup | Yes | We adopt SGD optimizer with an initial learning rate 0.01, momentum 0.9 and weight decay 5 10 4 to optimize the model. The polynomial decay [19] with the power of 0.9 is used as the learning rate scheduler. The learning rate of Adv Style γ is set to 3. All models are trained for 40K iterations with a batch size of 16. Four widely used data augmentation techniques are used during training, including color jittering, Gaussian blur, random cropping and random flipping. The input image is randomly cropped to 768 768 for training. |