tag:blogger.com,1999:blog-15418143.post3775278291392261459..comments2024-03-09T05:42:18.102-05:00Comments on Tombone's Computer Vision Blog: learning to "borrow" examples for object detection. Lim et al, NIPS 2011Tomasz Malisiewiczhttp://www.blogger.com/profile/17507234774392358321noreply@blogger.comBlogger6125tag:blogger.com,1999:blog-15418143.post-23048300887531116122012-09-20T14:08:05.025-05:002012-09-20T14:08:05.025-05:00dear sir,
iam working on my senior project
my pro...dear sir,<br />iam working on my senior project<br />my project is about car detection using matlab ( video processing)<br /> iam not sure if you can help me, but i face some problems with the code, iam asking for your help pleas<br />best regards <br />laianhttps://www.blogger.com/profile/06235242249039856400noreply@blogger.comtag:blogger.com,1999:blog-15418143.post-17196537229960173852012-02-27T12:14:23.948-05:002012-02-27T12:14:23.948-05:00Thank you for your reply. I am blocked by GFW, so ...Thank you for your reply. I am blocked by GFW, so I can't read your blog often. What you say makes sense. I think I should make some experiments on this topic. Thanks again.loveispnoreply@blogger.comtag:blogger.com,1999:blog-15418143.post-31343414185525477222011-12-29T16:42:43.208-05:002011-12-29T16:42:43.208-05:00like your blog and all the papers you reviewed. Go...like your blog and all the papers you reviewed. Good job!Zachttps://www.blogger.com/profile/00146723204669087695noreply@blogger.comtag:blogger.com,1999:blog-15418143.post-41077922487754850782011-12-20T13:02:17.257-05:002011-12-20T13:02:17.257-05:00@ loveisp
One will never be able to collect enoug...@ loveisp<br /><br />One will never be able to collect enough images to cover all of the world's visual concepts. In practice, dealing with thousands of images for negative data is sufficient. Remember that in the detection problem, the goal is to localize objects within images? What this means is that if I give you a single image and tell you that it doesn't contain any cats, any subwindow in that image can be treated as a non-cat. Most of those subwindows will be non-object patches. This means that a single negative image gives rises to ~20,000 negative data points.<br /><br />Generative methods have a longer history than discriminative methods, and there are many popular generative methods around. However, it is still a matter of research to tell which method (or combination of) will prevail.Tomasz Malisiewiczhttps://www.blogger.com/profile/17507234774392358321noreply@blogger.comtag:blogger.com,1999:blog-15418143.post-72846076171734109592011-12-20T06:44:14.955-05:002011-12-20T06:44:14.955-05:00Hi Tomasz, I want to ask u a basic problem.
If I ...Hi Tomasz, I want to ask u a basic problem.<br /><br />If I want to train a cat detector, how can I get non-cat images to cover so many different non-cat labels? It is a confusing problem for me.<br /><br />For example, dog is a non-cat label, car also a non-cat label, television, cup, table, cellphone... all also non-cat labels. How do I collect all images from all these labels? I think they are infinite. But if I only get some non-cat images, say 1000 images, to train the detector, how do the detector measure the difference between cat and the label not covered by the 1000 images?<br /><br />There is another related problem. Why do we not use the method only model cat to detect cat, like anomaly detection? Is it maybe the difference between discriminative model and generative model?loveisphttps://www.blogger.com/profile/14715016474601662093noreply@blogger.comtag:blogger.com,1999:blog-15418143.post-21572494415938285012011-12-20T06:40:31.517-05:002011-12-20T06:40:31.517-05:00This comment has been removed by the author.loveisphttps://www.blogger.com/profile/14715016474601662093noreply@blogger.com