Simple scanning-window detectors process image windows independently one by one. I am aware of two general approaches going beyond the independent processing.
One is by sharing computation of features: Feature-Centric Evaluation (Schneiderman 2004, Dalal 2005), feature-targeted preprocessing (Dollar 2009), detection of local features or parts (Felzenszwalb 2010, Leibe 2007, Vedaldi 2009, Lampert 2010), and, in a sense, convolutional neural networks.
The second approach is by exploiting smoothness of detector responses (Vedaldi 2009, Lampert 2010, Chum 2007).
Are there any others? Could you point me to good articles? I would appreciate even good articles which fit the two groups I already mentioned.
Hi,
Well you question is rather strange as by definition a scanning window detector is a scanning window detector! So i think what you mean is "Is there any alternative to scanning window detector for object detection"
And the answer would be yes, there are lots.
here's one which is most like scanning window, but guides the scanning process with context.
Searching for objects driven by context:
http://groups.inf.ed.ac.uk/calvin/Publications/alexe12nips.pdf
Here's a nice article that you should find very usdeful in answering your question, it covers many techniques for object localisation and classification:
Object class detection: A survey
http://dl.acm.org/citation.cfm?id=2522978
Paul, thanks for both of the references. The NIPS paper is definitely relevant and the survey will be useful as well.
I admit that "scanning window" might be misleading, but my vocabulary and taxonomy of detectors are letting me down. I would say that I am looking for detection methods which are "close" to "scanning window" detectors.
Michal,
Yes the vocabulary surrounding these techniques is very confusing and is not in anyway consistent as of yet, even two people sitting next each other in the same lab may have different vocabulary on the matter!
If your are looking for things like scanning window, then perhaps you are looking for ways to process the response maps of a scanning window, in a more pleasant way than bog standard greedy local non-max suppression?
If so then you are probably heading towards so called "structured output prediction" with a scanning window response maps "stacked" as the features.
in which case this 2009 Marr prize winner! will be of great interest to you:
Discriminative models for multi-class object layout
http://www.ics.uci.edu/~fowlkes/papers/drf-iccv09.pdf
Cheers
- Badges
- Science topic