DaDA: Distortion-aware Domain Adaptation for Unsupervised Semantic Segmentation

Authors: Sujin Jang, Joohan Na, Dokwan Oh

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results highlight the effectiveness of our approach over state-of-the-art methods under unknown relative distortion across domains. We present extensive experimental results to validate our distortion-aware domain adaptation (Da DA) framework for semantic segmentation in the presence of both visual and geometric domain shifts.
Researcher Affiliation Industry Sujin Jang Samsung Advanced Institute of Technology s.steve.jang@samsung.com Joohan Na Samsung Advanced Institute of Technology joohan.na@samsung.com Dokwan Oh Samsung Advanced Institute of Technology dokwan.oh@samsung.com
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No Datasets and more information are available at https://sait-fdd.github.io/. This statement does not explicitly mention that the source code for the methodology is provided at the URL.
Open Datasets Yes The Cityscapes dataset contains..., The GTAV dataset contains..., The Woodscape dataset consists of..., all followed by citations. Woodscape [38], Cityscapes [8], and GTAV [29] are commonly used public datasets.
Dataset Splits Yes We use front and rear camera scenes containing 4, 023 images in our experiments. The images are randomly split into a training set with 3, 023 images and a validation set with 1, 000 images. We randomly pulled 974 of validation images and remaining 2, 923 images are used for the training.
Hardware Specification Yes All of our codes are written in Py Torch and trained on a single NVidia RTX A6000 GPU with 48 GB memory.
Software Dependencies No All of our codes are written in Py Torch. While PyTorch is mentioned, a specific version number is not provided, which is required for reproducibility.
Experiment Setup Yes We trained all networks with the Adam [18] solver with a batch size of 4. The learning rate is 0.2 10 5 for M and DM and 0.1 10 6 for G and DG. We set the weight factors of losses in Eq.(6) as: β1 = 100.0, β2 = 10.0, β3 = 10.0 for Cityscapes Woodscape (or FDD); and β1 = 100.0, β2 = 1.0, β3 = 100.0 for GTAV Woodscape (or FDD).