tag:blogger.com,1999:blog-15418143.post8144480098588221164..comments2024-03-09T05:42:18.102-05:00Comments on Tombone's Computer Vision Blog: One Part Basis to Rule them All: Steerable Part ModelsTomasz Malisiewiczhttp://www.blogger.com/profile/17507234774392358321noreply@blogger.comBlogger1125tag:blogger.com,1999:blog-15418143.post-88445077566560595382012-04-24T03:55:32.680-05:002012-04-24T03:55:32.680-05:00I see strong connections with this general line of...I see strong connections with this general line of work (taking HOG cells and learning dictionaries) and deep learning. A single HOG cell is exactly equivalent to a one-layer conv net. Specifically, with simple cells in layer (S1) with 9 linear (rectified) filters, each detecting an edge of different orientation, and a complex layer cell (C1) that does average pooling across the simple cell responses. A deep network would go on to learn the second simple cells layer (S2) on top of these C1 cells by learning a dictionary again in a slightly extended receptive field. Here, that is equivalent to learning a dictionary for a grid of HOG Cells. In other words, there is the same theme of building alternating layers of AND operations and OR operations -- selective operations and invariant ones -- simple cells or complex cells, etc. It's funny how these two communities have their own dictionaries, but there is often a 1-1 translation and they are computing basically equivalent operations on images.<br /><br />In this particular paper though, it seems that the dictionary is not obtained by simply running a version of Vector Quantization on the hog cell grids, so I'm not quite sure about the mapping there. Will have to read up a bit more on this.<br /><br />Though, btw, in deep learning people sometimes tend to prefer using MAX operation as opposed to AVG for complex cells, as they sometimes find it works better. I wonder -- what kind of results do you get if, in the HOG code, you don't do a histogram of edges, but instead set each bin to be max of edge response in the particular direction?Andrej Karpathynoreply@blogger.com