tag:blogger.com,1999:blog-6141980.post3467056896827061686..comments2024-03-20T12:28:35.004-05:00Comments on Nuit Blanche: How does the Rice one pixel camera work ?Igorhttp://www.blogger.com/profile/17474880327699002140noreply@blogger.comBlogger7125tag:blogger.com,1999:blog-6141980.post-33609019914507004522013-06-06T06:46:15.020-05:002013-06-06T06:46:15.020-05:00Thanks for the very clear explanation.Thanks for the very clear explanation.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-6141980.post-80693477723312734112011-04-10T17:37:12.478-05:002011-04-10T17:37:12.478-05:00I know this is a bit old, but I wanted to point ou...I know this is a bit old, but I wanted to point out that a larger single pixel with back thinning, and an electron multiplying register could do well to improve single exposures, thereby reducing the acquisition time. - I have some descriptions of EM cameras at www.austinblanco.com/blogUnknownhttps://www.blogger.com/profile/05427828887504222396noreply@blogger.comtag:blogger.com,1999:blog-6141980.post-43935003454338953462009-12-22T13:39:27.454-06:002009-12-22T13:39:27.454-06:00Thanks for the answer. I have one more question.
...Thanks for the answer. I have one more question. <br /><br />As I understand, the random mask represents a projection matrix 'U' consisting of 0's and 1's, and image data 'x' (from spatial domain) is projected onto these bases. <br /><br />Do we need that the data in original space should be sparse (i.e. 'x' should be sparse in the current representation)? or it can be sparse in any basis?<br /><br />Another question is - suppose 'x' is not sparse in current representation. Do we need to design the random projection matrices so that they bring out the sparsity in data? or it can be any random matrix of 0's and 1's - which are ofcourse incoherent with cosine bases?<br /><br />thanks a lot,<br />abhishek.abhisheknoreply@blogger.comtag:blogger.com,1999:blog-6141980.post-22587275366025759032009-12-20T04:41:41.146-06:002009-12-20T04:41:41.146-06:00Abhishek,
Yes, as far as I can tell you do need m...Abhishek,<br /><br />Yes, as far as I can tell you do need m*t exposure time.<br /><br />Cheers,<br /><br /><br />Igor.Igorhttps://www.blogger.com/profile/17474880327699002140noreply@blogger.comtag:blogger.com,1999:blog-6141980.post-23776515898800028862009-12-19T23:52:02.530-06:002009-12-19T23:52:02.530-06:00Do we need a longer exposure time for these single...Do we need a longer exposure time for these single pixel cameras? If 't' is the exposure time needed for traditional camera, then for single pixel camera do we need 'm*t' exposure time where 'm' is the number of distinct random masks?abhishekhttp://www.cs.utah.edu/~abhiknoreply@blogger.comtag:blogger.com,1999:blog-6141980.post-3060980220037361562007-07-20T10:26:00.000-05:002007-07-20T10:26:00.000-05:00Aleks,I guess your question is:- The number of mea...Aleks,<BR/><BR/>I guess your question is:<BR/>- The number of measurements is lower<BR/>- but you need to also transmit all the "masks" that produced the random measurement.<BR/>-> the total amount of information is much higher, because you have to send both the "masks" and the measurements.<BR/><BR/>Random needs to be taken in the following understanding: <BR/>- at the lab, you use your favorite random number generator and produce these masks ahead of time.<BR/>- you install those "masks" as tables/memory in the sensor<BR/>- when the sensor is in the field, it recalls these masks from its memory. <BR/>- when transmitting the compressed measurement somewhere, you can add a bit or two to let the receiver know which of the random mask was used.<BR/><BR/>So in effect, it is random, but it is not generated on the fly.<BR/><BR/>For the random intelligently, some people are doing it and the bayesian compressive sensing would fall in that category. However, in these types of system, doing any computation uses a lot of power and so the thinking is that if you want to have those for years, you do not want them to draw on the power side of things.<BR/><BR/>And then there is the kicker as mentioned by Richard Baraniuk and his team. Since you know these random measurements are pretty much universal, the information you are retrieving from them may not be the same as the ones you are retrieving from them ten years down the road where we will have much better reconstruction techniques. If you put intelligence in it today, you are most likely decreasing your ability to get anything interesting out these data in the future.<BR/><BR/>Igor.Igorhttps://www.blogger.com/profile/17474880327699002140noreply@blogger.comtag:blogger.com,1999:blog-6141980.post-55855667184937746172007-07-20T09:47:00.000-05:002007-07-20T09:47:00.000-05:00How would one be able to use this concept for comp...How would one be able to use this concept for compression: you would have to communicate the locations of those pixels too, which would take a lot of bits?<BR/><BR/>Second, why sample randomly, why not sample intelligently?Aleks Jakulinhttps://www.blogger.com/profile/14056353067485759304noreply@blogger.com