Peekaboo

There is apparently a "new" trend in computer vision that I seem to have missed. It's explicit feature maps . So what is this about? We all know and love the kernel trick: have some vector representation X of your data and some algorithm (like SVMs) that rely only on dot products in the input space. Replace the dot products by some positive definite kernel and by Mercer's Theorem there is some Hilbert space of functions Y and some mapping phi: X -> Y such that your kernelized algorithm works as if you mapped your input to phi. This extremely powerfull method is now used everywhere, most of all for SVMs. But it has several significant drawbacks in this case: Training using a non-linear kernel is a lot slower [O(n^3)] than training a linear SVM [O(n)]. Kernel SVMs take a lot of memory to store, since they are "non-parametric" in a sense. In addition to the Lagrange-Multipliers alpha, all the support vectors have to be stored. Recall is painfully...

I was just reading the paper in the title ( pdf ), which Alexei Efros talked about at CVML. It's from this years ICCV. The basic idea of the paper is to do object detection by training a linear SVM on hog features for each positive example. I thought the idea was pretty cool and wanted to write something about it... and then I saw there is already a post by the first author, Tomasz Malisiewicz, in his own blog . He also discusses some of his (matlab :( ) code for non-maximum suppression. So check out his blog for more details on this cool paper :)

Somehow I neglected writing about the CVML summer school (which is now what, 3 weeks ago?). I didn't really have the time to go through all of my notes yet. When I was at my notes on Alexei Efros' talk on Large Scale Recognition, I saw that he mentioned some joint work with Antonio Torralba on dataset biases in the computer vision community. The paper is called "Unbiased Look at Dataset Bias" ( pdf ) and appeared at CVPR this year. If you haven't heard of it: it is a MUST READ! Read it now! From the paper: "Disclaimer: No graduate students were harmed in the production of this paper. Authors are listed in order of increasing procrastination ability."

Last Thursday, Ivan Laptev talked about "Human Action Recognition" at the CVML. This is not really my area so while I really liked his very enthusiastic talk, I won't say much about it ;) What I liked about it most was how it was motivated. From a computer vision perspective, I felt human action recognition was somewhat peripheral to current research - surely with many interesting applications but not central to the field. However, Ivan Laptev had two arguments for human action recognition to be a center piece of visual understanding: Most of the data out there - in particular videos - show people: 35% of the pixels in TV and movies belong to people, 40% on youtube. Laptev concludes that video analysis is human action analysis. The semantics of objects can often be inferred from humans interacting with it. Instead of the classical "chair" example, Laptev showed a "luggage train": the thing in the airport you pick up your luggage from. Even if ...

This is my favourite part of every converence and workshop: quotes and fun facts :) There are some things researchers will never write in a paper but that they really like to tell you. Also many professors actually have a pretty good humor (or at least one that is as nerdy as mine). Please not that even though I put the quotes into quotation marks, they might not be completely accurate. Most lecturers can talk pretty fast and I am usually taking notes in the old-school paper way... Ponce: About learned dictionaries and filters: "Dictionary elements don't have semantic meaning. People like to look at them, I don't know why." About denoising using structured sparsity: "We don't know anything about image processing. The finish guys are way better. But the sparse model still works better." Lambert: About 1-vs-all training for multi class classification: "Everyone is using that. But no one knows why it works." Francis Bach (?): Abo...

The last talk of the CVML summer school (yes, I'm starting from the back ;) was by Martial Hebert about using scene geometry for recognition and segmentation. His work focuses mainly on pictures of man-made environment but is not exclusive to it. Heberts talk was two hours long and spanned many of his past and recent work which I can not all repeat here. I will focus on some main messages and things I got from his presentation. One of the first works Hebert talked about was classifying regions of an image into "surface orientation"  categories. These categories are roughly "ground plane", "sky", "vertical facing right", "vertical facing left", "vertical facing camera". I had seen several works in this direction but never found them to be very interesting. "Why this task?" is what I was always asking myself - I never quite understood the motivation. In his talk, Hebert made the motivation very clear: This i...

In one of the lectures at the CVML summer school, Josef Sivic proposed the IKEA problem: Finding all Ikea furniture in all of YouTube. Seems like a pretty cool task to me. And not completely unrealistic. It is a bit more challenging as it seems, though: There is the obvious large-scale aspect of trying to analize all YouTube videos - and with it come the restrictions such as the need for linear classifiers and efficient features. But another challenge are the objects themselves: Much of the Ikea furniture has no real texture - at least not on an interesting scale. A kitchen Table from ikea might have some small wood structure. But basically its a rectangle with 4 little straigt legs. From very varying viewpoints. This calls for a shape descriptor. The only one I am aware of that is currently used is Hog. But by nature hog is far from viewpoint invariant - it is not supposed to be. Fulkerson uses a mixture of Hogs for cars on streets - a task with arguably less variation in vi...

There were many posters on the CVML summer school and I won't talk about all of them. Actually 8 of them got prizes (in form of hand-signed CV and ML books). I knew some of the work form NIPS2010 but there were some things that were new to me: Alexander Vezhnevets presented work on Multi Image Model for Semantic Segmentation with Different Levels of Supervision. I don't know how I could miss that before. This is amazing work on weakly supervised semantic scene segmentation on MSRC. It makes use of CRFs, boosted texton forests and superpixels. The CRF does not only connect neighbouring superpixels but also superpixels in different images that look similar. Super pixel labels are treated as latent variables and only a very simple contraint between image label and superpixel label is enforced. Since I am looking at a very similar task at the moment, even though the other posters were very good, this one was definitely the best for me. Yang Hua presented work on Contextual...

Thanks to my institute and the B-IT , I can attend the CVML Summer school organised by ENS and INRIA in Pairs. It started on Monday and features many great speakers, for example Jitendra Malik, Cordelia Schmid, Andrew Zisserman, Jean Ponce and may others. I was pretty busy with the program until now but I hope I'll find some time to write about all the great lectures here. The lectures at the beginning of the week were about standard topics like Francis Bach's tutorial on SVMs and kernel methods and Cordelia Schmid's and Josef Sivic's introduction to interest points, features and visual words. During the week talks ranged from details about current state of the art in object recognition and large scale learning to bigger picture talks and directions for future research. Many of the professors put their slides online. I definitely suggest to have a look at those. Hopefully I have the time to write about all of the talks but I doubt it a little.