What's the difference between various image resizing algorithms in the module of PIL.Image?

As described in Image.resize, there are multiple mothods to resize an image. For eexample, PIL.Image.NEAREST, PIL.Image.BILINEAR, PIL.Image.BICUBIC et al. But in util.py for semantic segmentation, Image.ANTIALIAS is used when target image size is smaller than source image, and Image.BICUBIC is used when target image is larger, even Image.LINEAR is also used. What's the difference among them?