Deep Learning for Computer Vision

The scale-space concept

The scale-space is the concept of using images that are of various sizes. An image is reduced to smaller size, hence bigger objects can be detected with the same-sized window. An image can be resized to some sizes with decreasing sizes. The resizing of images by removing alternative pixels or interpolation may leave some artefacts. Hence the image is smoothened and resized iteratively. The images that are obtained by smoothening and resizing are scale space.

The window is slide on every single scale for the localization of objects. Running multiple scales is equivalent to running the image with a bigger window. The computational complexity of running on multiple scales is high. Localization can be sped up by moving ...