This method involves randomly selecting small cubes within an image or 3D volume (if working with voxel data) and setting their values to zero. In the context described, these cubes are referred to as ‘zero-value cubes’. This strategy aims to obscure parts of the data in a random manner. However, as noted in your text, the random mask may lead to poor performance because it only contains local, incoherent parts of the foreground, lacking a comprehensive representation of the image or volume.
a zero-centered mask likely refers to a masking strategy where the zero-value region (or obscured area) is centrally located within the data. This approach might be designed to ensure that the central part of the data is occluded, forcing the model to learn from the peripheral regions. This can encourage the model to integrate more context or background information in its learning process.
In this strategy, the shape of the zero-value region is βH × W × L, with β set to 8/27. This approach seems to create a larger, contiguous obscured area, as opposed to the dispersed small cubes in the random mask. The contact mask might provide better integrity of foreground information than the random mask, as it doesn’t fragment the view as much. However, it’s noted to perform worse than the zero-centered mask because it allows less interaction between the foreground and the background. 接触掩膜定义了复制部分的形状和位置。
These masking strategies are likely used in the context of semi-supervised learning or unsupervised learning, where part of the input data is deliberately obscured or masked to train models to better understand or reconstruct the underlying information. The effectiveness of these strategies can depend on factors like the nature of the data and the specific goals of the model training.