1. This paper proposes a novel attention mechanism for mobile networks called "coordinate attention" which embeds positional information into channel attention.
2. The coordinate attention factorizes channel attention into two 1D feature encoding processes that aggregate features along the two spatial directions, respectively.
3. Extensive experiments demonstrate that the coordinate attention is beneficial to ImageNet classification and down-stream tasks such as object detection and semantic segmentation.
The article is generally reliable and trustworthy, as it provides evidence for its claims in the form of extensive experiments demonstrating the effectiveness of its proposed coordinate attention mechanism for mobile networks. The article also provides a link to code which can be used to verify its claims further. However, there are some potential biases in the article, such as not presenting both sides equally (e.g., only focusing on the benefits of coordinate attention without exploring any potential drawbacks). Additionally, there may be some missing points of consideration or evidence for certain claims made in the article (e.g., not providing enough detail about how exactly coordinate attention works). Finally, there may be some promotional content in the article (e.g., emphasizing how effective coordinate attention is without providing enough detail about other existing methods).