Closed ascbot closed 5 years ago
Original Redmine Comment Author Name: Moses Milazzo (Moses Milazzo) Original Date: 2011-06-29T22:31:31Z
Jigsaw has the ability to weight the measures according to a number of accuracy measures. The registration algorithms should inform jigsaw of this accuracy in some way. Included in this could be the knowledge of the differences in viewing geometry, including lighting and viewing conditions.
Desire: Ability for a human to teach the algorithm to modify its parameters to achieve "best" results.
Original Redmine Comment Author Name: Moses Milazzo (Moses Milazzo) Original Date: 2011-07-12T21:26:13Z
ISIS2 can register Titan images, but ISIS3 isn't working well (too many false positives).
Original Redmine Comment Author Name: Stuart Sides (@scsides) Original Date: 2012-03-01T18:02:59Z
Updated Type Major to Recommendation. Priority remained at High
Original Redmine Comment Author Name: Travis Addair (Travis Addair) Original Date: 2012-04-09T17:06:44Z
I have been moved off this issue for lack of time remaining to work on it.
Original Redmine Comment Author Name: Stuart Sides (@scsides) Original Date: 2012-04-11T16:56:50Z
Moved to acknowledged due to departing developer
Original Redmine Comment Author Name: Jeff Anderson (Jeff Anderson) Original Date: 2012-06-20T00:04:32Z
Sounds right to me.
Orrin -----Jeff Anderson janderson@usgs.gov wrote: -----
To: Raad A Saleh <rsaleh@usgs.gov>
From: Jeff Anderson <janderson@usgs.gov>
Date: 05/22/2012 12:23PM
cc: Orrin H Thomas <othomas@usgs.gov>
Subject: Re: info
Hi Raad,
Orrin did most of the work on this so I will summarize what I know
1) The original claim by the users was the sub-pixel computation was not
accurate enough.
2) Orrin added a new sub-pixel algorithm using centroiding.
3) Orrin felt the original sub-pixel computation was pretty good (within
1/2 pixel)
4) His new algorithm may have improved the subpixel results by 10%
5) After the sub-pixel analysis, the real problem appears to be was the
whole pixel accuracy was not close enough for sub-pixel to succeed in
many cases
6) Orrin look at outlier rejection using RANSAC but it was too slow and
was not generalized for line scan cameras
7) Then Orrin worked on detecting and eliminating outliers in jigsaw
(maximum likelihood estimation) which proved fruitful
8) Travis also developed outlier rejection by comparing at forward and
reverse coregistration. Measure 1 held fixed and Measure 2 registered
then Measure 2 held fixed and Measure 1 registered. If the two
registrations deviated by a user tolerance the measures were rejected in
"pointreg"
9) The new rejection techniques allow the user to more easily eliminate
bad points
I'm cc'ing to Orrin so he can give his concurrence on this summary.
Thanks, Jeff
Jeffery Anderson
Supervisory Computer Scientist
Astrogeology Science Center
2255 N Gemini Drive
Flagstaff, Arizona
928-556-7167
On 5/21/12 1:19 PM, Raad A Saleh wrote:
>
>
> Hello Jeff,
>
> I need your kind help, if you can. This is a paragraph I am writing:
>
> ISIS has embedded routines and stand-alone functions for point matching,
> such as pointreg and coreg. These routines and functions are fundamental
> to production of high precision products, as we will discuss in more
> details at the Perceived Impact section. Our interest in improving the
> performance of ISIS matching capabilities was partly stimulated by the
> recent discovery of specific problems that prevented achieving subpixel
> match accuracy. Some of these specific problems have been addressed by
> recent modifications.
>
> I wonder if you can you provide me with the following:
> A bried description of these problem; and
> What are these recent modifications, and how have they (or how would they)
> address the specific problems.
>
> Thanks in advance,
>
> Ra'ad
Original Redmine Comment Author Name: Jeff Anderson (Jeff Anderson) Original Date: 2012-06-20T00:11:09Z
Finished effort hours set to 1. These tasks were set as other mantis tickets for Orrin and Travis.
Author Name: Moses Milazzo (Moses Milazzo)
Original Assignee: Jeff Anderson
1) Determine differences between ISIS2 and ISIS3 subpixel registration (coreg, subpreg)
Steps to reproduce:
Compare the results of ISIS2 CALCREG/REGCHIP to the ISIS3 results. Compare both ISIS2 and ISIS3 with human-picked subpixel registration. Team to do these tests needs to be identified.
Kris has implemented the Gruen algorithm in ISIS3--needs testing.
Based on testing, we will revisit what the fixes need to be.
Will probably want to add brute force subpixel registration algorithm to ISIS3.
In the future we could look at other algorithms and certainly should involve Raad Saleh in this discussion.