Geometry-Aware Network for Domain Adaptive Semantic Segmentation

Authors: Yinghong Liao, Wending Zhou, Xu Yan, Zhen Li, Yizhou Yu, Shuguang Cui

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Qualitative and quantitative results demonstrate that our model outperforms state-of-the-arts on GTA5 Cityscapes and SYNTHIA Cityscapes.
Researcher Affiliation Academia 1 The Future Network of Intelligence Institute, The Chinese University of Hong Kong (Shenzhen) 2 School of Science and Engineering, The Chinese University of Hong Kong (Shenzhen) 3 Department of Computer Science, The University of Hong Kong {yinghongliao@link., lizhen@}cuhk.edu.cn
Pseudocode No The paper includes diagrams illustrating the network architecture and training schemes (Figures 1, 2, 3, 4), but it does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access to its source code, such as a repository link or an explicit statement about code release in supplementary materials.
Open Datasets Yes Datasets. In our experiments, we adopt the GTA5 and SYNTHIA datasets as the source synthetic datasets, respectively, and exploit the Cityscapes dataset as the target real scene dataset. The GTA5 dataset (Richter et al. 2016) has 24,966 synthetic photo-realistic scenes with a resolution of 1914 × 1052. SYNTHIA (Ros et al. 2016) is a large-scale synthetic urban image dataset, containing 9,400 synthetic urban images with the resolution of 1280 × 760 in model training. The Cityscapes dataset (Cordts et al. 2016) is a realistic dataset of the 2048 × 1024 street scenes from 50 cities, where 2,975 images for training and 500 ones for validation.
Dataset Splits Yes The Cityscapes dataset (Cordts et al. 2016) is a realistic dataset of the 2048 × 1024 street scenes from 50 cities, where 2,975 images for training and 500 ones for validation.
Hardware Specification Yes We conduct the experiments on 4 Tesla V100 GPUs.
Software Dependencies No The paper states, 'Our model is implemented in PyTorch' and mentions using 'Deep Labv2' and 'Seg Former' as architectures, but it does not specify version numbers for PyTorch or any other software libraries or dependencies.
Experiment Setup Yes The Cityscapes dataset (Cordts et al. 2016) is a realistic dataset of the 2048 × 1024 street scenes from 50 cities, where 2,975 images for training and 500 ones for validation. [...] We utilize the Deep Labv2 (Chen et al. 2018) and Seg Former (Xie et al. 2021) as the segmentation architectures. [...] We train the whole stage using the following total loss with the weight µ: Ldepth-aware = Ls seg + Ls depth + µLadv. (6) [...] where λ is the loss weight. Ls SILog is the Scale-Invariant Logarithmic loss for depth estimation (Yuan et al. 2022): [...] where ψ is the scale constant set as 10 and γ denotes the variance minimizing factor set as 0.85. [...] Thus, the total loss of the geometry-aware adaptation stage is: Lgeometry-aware = Ls seg + Lt seg + αLs SGC + βLt SGC, (10) where α and β control the weights of the SGC loss on each domain.