Fig 1.
For different distributions of scenes, the proposed GDAAR adaptively employs different weights on the feature map.
For instance, the feature map of input with complex scene distribution (a) may need more attention to obtain low-level details, and simple scene distribution (b) may same attention to get different-level features. Red circle and blue circle in diagrams denote the higher weight and the lower weight respectively, and black quadrangles denote the attention regulator of the corresponding feature node in feature maps.
Fig 2.
The proposed global domain adaptation attention with data-dependent regulator framework for scene segmentation.
To ensure stability in the system, it is common practice to fix certain components in the architecture, such as the beginning STEM (preprocessing) and the final upsample block. Top: The global domain adaptation attention space with feature maps of L layers. The dotted lines among red circles or blue circles mean same feature transformations from N feature nodes. Red circles and blue circles denote feature transformations, and yellow quadrangles denote the regulator of the corresponding feature node on feature maps. Bottom: Given the input from the former layer, we first generate global domain adaptation attention through obtaining global scope relations. To further improve the performance of our approach, we design a data-dependent regulator that adjusts the attention weight on the feature map during inference. By incorporating this additional mechanism, our model is able to better adapt to diverse input domains and enhance segmentation accuracy for challenging images.
Table 1.
Effect on global domain adaptation attention.
Fig 3.
Visualization results of global spatial relation on Cityscapes validation set.
Fig 4.
Visualization results of global channel relation on Cityscapes validation set.
Table 2.
Performance on both STEM block and regulator.
Fig 5.
Visualization results of our GDAAR method on Cityscapes validation set.
Table 3.
Segmentation comparisons on cityscapes dataset.
Table 4.
Segmentation comparisons on PASCAL context.
Table 5.
Segmentation comparisons on COCO stuff dataset.