Deliberated Domain Bridging for Domain Adaptive Semantic Segmentation

Authors: Lin Chen, Zhixiang Wei, Xin Jin, Huaian Chen, Miao Zheng, Kai Chen, Yi Jin

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on adaptive segmentation tasks with different settings demonstrate that our DDB significantly outperforms state-of-the-art methods.
Researcher Affiliation Academia Lin Chen1,2 Zhixiang Wei1 Xin Jin3 Huaian Chen1 Miao Zheng2 Kai Chen2 Yi Jin1 1 University of Science and Technology of China 2 Shanghai AI Laboratory 3 Eastern Institute for Advanced Study
Pseudocode Yes These two steps are conducted iteratively, and the ending of each round will serve as the beginning for the next round (see detailed algorithm in supplemental material).
Open Source Code Yes Code is available at https: //github.com/xiaoachen98/DDB.git.
Open Datasets Yes We use four publicly available semantic segmentation benchmarks for validation, including two synthetic scenes and two real-world scenes. In detail, GTA5 [48] is a synthetic dataset of 24,966 labeled images obtained from a video game. Synscapes [56] is also a synthetic dataset of 25,000 images created by photo-realistic rendering techniques, and its style is closer to real-world driving scenes than GTA5. Cityscapes [9] is a real-world urban dataset collected from European cities, with 2,975 images for training and 500 images for validation. Mapillary Vista [42] is a large-scale dataset collected by various imaging devices worldwide and includes 18,000 images for training and 2,000 images for validation.
Dataset Splits Yes Cityscapes [9] is a real-world urban dataset collected from European cities, with 2,975 images for training and 500 images for validation. Mapillary Vista [42] is a large-scale dataset collected by various imaging devices worldwide and includes 18,000 images for training and 2,000 images for validation.
Hardware Specification Yes We use the mmsegmentation [8] codebase and train models on RTX 3090Ti GPUs.
Software Dependencies No The paper mentions 'mmsegmentation', 'Deep Lab-v2', 'Res Net101', and 'Adam W' but does not specify their version numbers.
Experiment Setup Yes We set the learning rate as 6e-5 for the backbone and 6e-4 for the decoder head, use a weight decay of 0.01 and a linear learning rate warmup followed by 1.5k iterations linear decay. All experiments are trained on a batch of 512x512 random cropped images for 40k iterations. We set the batch size to 2 for analysis and experiments in Tab. 1 and Tab. 6, and set batch size to 4 for other results. Following [50], we use the same augmentation parameters and set τ = 0.968. For Cut Mix [62], the ratio of the selected region for cross-domain pasting is experimentally set to 0.3. For Class Mix [43], half of categories in the source domain are selected for cross-domain pasting.