51 ms·
This algorithm is relatively old, but there is an interesting question about it: given how well it works (eg on landscape images), one wonders if and how it cou
by ur-whale 7y ago
This algorithm is relatively old, but there is an interesting question about it: given how well it works (eg on landscape images), one wonders if and how it could be recast in a deep learning / differentiable programming setting.
The problem is, the dynamic programming and/or median extraction techniques that the algorithm uses (IIRC) are not very differentiable.
- jhurliman 7y agoGiven an energy map of an image and a target number of pixels to remove, the algorithm is deterministic. The step to apply ML would be the energy map generation. Saliency map generation with CNNs is an active area of research and it could replace a slew of more specific algorithms you would want to apply (face detection, foreground/background segmentation, edge detection, etc).