Unsupervised Domain Adaptation for Semantic Segmentation by Content Transfer
Authors: Suhyeon Lee, Junhyuk Hyun, Hongje Seong, Euntai Kim8306-8315
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that the proposed method achieves the state-of-the-art performance in semantic segmentation on the major two UDA settings. We compare our method with existing state-of-the-art UDA methods Adapt Seg Net (Tsai et al. 2018), CLAN (Luo et al. 2019), Max Square (Chen, Xue, and Cai 2019), SSFDAN (Du et al. 2019), DISE (Chang et al. 2019), Adv Ent+Min Ent (Vu et al. 2019), APODA (Yang et al. 2020), Max Square+IW+Multi (Chen, Xue, and Cai 2019), Patch Alignment (Tsai et al. 2019), Weak Seg DA (Paul et al. 2020), and BDL (Li, Yuan, and Vasconcelos 2019). We report the results for GTA5 Cityscapes in Table 1 and the results for SYNTHIA Cityscapes in Table 2. |
| Researcher Affiliation | Academia | Suhyeon Lee, Junhyuk Hyun, Hongje Seong, Euntai Kim* School of Electrical and Electronic Engineering, Yonsei University, Seoul, Korea {hyeon93, jhhyun, hjseong, etkim}@yonsei.ac.kr |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statement regarding the release of open-source code for the described methodology, nor does it provide a link to a code repository. |
| Open Datasets | Yes | In this works, we consider two common synthetic-to-real domain adaptation problems; GTA5(Richter et al. 2016) Cityscapes(Cordts et al. 2016) and SYNTHIA(Ros et al. 2016) Cityscapes. |
| Dataset Splits | Yes | For both synthetic-to-real experiments, we evaluate the segmentation model in the validation set of the Cityscapes. The dataset consists of a training set with 2975 images, a validation set with 500 images, and a test set with 1525 images. |
| Hardware Specification | Yes | We implement with Py Torch (Paszke et al. 2017), and all experiments are conducted using a single Nvidia Titan RTX GPU and an Intel Core i7-9700K CPU. |
| Software Dependencies | No | The paper mentions 'Py Torch (Paszke et al. 2017)' but does not provide a specific version number for PyTorch or any other software dependencies required to replicate the experiments. |
| Experiment Setup | Yes | During the training, we resize the input image to a size of 1024 512 and then randomly crop to a 512 256 size patch. We train the model with a batch size 2 and use the weights provided in (Chang et al. 2019) as initial weights. The batch normalization layers are freeze. We use the SGD solver with an initial learning rate of 2.5 10 6 and a momentum of 0.9 for the segmentation model and the Adam solver with an initial learning rate of 1.0 10 5 and betas β1 = 0.5, β2 = 0.999 for the generators. We use the Adam solver with an initial learning rate of 1.0 10 6 to optimize discriminators for segmentation and image, and they have betas β1 = 0.9, β2 = 0.99 and β1 = 0.5, β2 = 0.999, respectively. All the learning rates decrease according to the polynomial policy. |