Thanks for the explanation. This kind of background is essential to
understand the methods and thus use them right.
Just to fully clarify, for a given keypoint in image1, is the ratio of
"the distance to the closest to the distance to the second closest"
computed
to all keypoints of image 2 or just to those within a radius equal to
precision only? If the ratio is computed to all keypoints of image 2
we could often have
points with similar radiometry, thus would get a high ratio and
erroneously discard a correct matching. While the probability of
having a non-homologous point with similar radiometry within the
radius of precision would be low and thus the erroneous rejection
would be rare.
Also, why is "the signature is a 128 dimension vector" ? Are you using
a window instead of the actual pixel of the keypoint?
Agus
On Mon, Aug 31, 2015 at 3:12 PM, Julien Michel <[email protected]
<mailto:[email protected]>> wrote:
Le 31/08/2015 14:29, Agustin Lobo a écrit :
Actually, I do not quite understand:
Is mfilter stating whether homologous points are filtered
according to the value
of "precision"?
yes
What do you mean by " padded according to the user defined
*precision*.?
I mean that in geobins mode, we consider square tiles in reference
image and predict the location of the same tile on the secondary
image. We take a slightly larger region on the secondary image, to
account for sensor model precision (given by parameter
"precision"). The idea is the following : the naive matching cost
is O(n1xn2), where n1 and n2 are the number of keypoints found in
each image. Therefore if you try to match large images, it will
take forever ... But if you divide your reference image in bins,
and predict roughly where the bin is in the second image, you only
match a few point at a time, and therefore gain speed (and
robustness, since it will constrain the range within which false
matches occur).
No padding is done in full mode (obviously), which is the one you
use.
What is threshold? Is it a radiometric distance or an spatial
distance?
Threshold is used during matching. Each keypoint has a signature,
and the matching algorithm works as follows :
- Euclidean distance is computed between the signatures of each
keypoint of the reference image and each keypoint of the secondary
image (hence the O(n1xn2) cost).
- A naive matching approach would consider that the correct match
has the minimum distance
- However, this does not allow to discard matches (each point will
have a match). But we know that some points should not have
matches (because they were not detected in both images). We
therefore need a way to discard them.
- We could filter them by considering that the distance should be
minimum AND bellow a given threshold. However, the signature is a
128 dimension vector, and setting such a threshold is not easy,
- A much more robust way of finding matches is by thresholding the
ratio from the distance to the closest to the distance to the
second closest. In fact, the more this ratio is close to 1 the
more the match is ambiguous (because there exist a second point in
about the same distance). The thresold parameter is exactly this :
the thresold above with the match will be discarded by the
matching algortihm.
Let me explain. 0.6 is a fair default value and used in many
papers / implementaiton. By setting threshold to 10, you discard
the filter on distance ratio (remember the max is 1). Therefore,
you have one match for each of the detected point in image 1.
You therefore end up with a lot of false matches, and this is
where mfilter comes : since you are using this option, all
matches that are more that "precision" away of the the homologous
point predicted by sensor modelling get discarded (therefore, you
discard 11469 points).
To cut it short : with threshold = 0.6, you have a few homologous
points, but none of them get discarded by the posteriori
filtering, probably because they are good.
With threshold = 10, you have a lot of homologous points, but most
of them get discareded by the a posteriori filter.
I have tested and by increasing threshold I get more homologous
points but also more of them get discarded (all other parameters
equal):
$ otbcli_HomologousPointsExtraction -in1 IMG_0062.tif -band1 2
-in2 IMG_0062.tif -band2 1 -mode full -precision 50 -mfilter 1
-algorithm surf -out IMG_0062homSURF2.txt -threshold 0.6
2015 Aug 31 13:58:34 : Application.logger (INFO) Elevation
management: setting default height above ellipsoid to 0 meters
2015 Aug 31 13:58:34 : Application.logger (INFO) Doing update
2015 Aug 31 13:58:40 : Application.logger (INFO) Found 15650
surf points in image 1.
2015 Aug 31 13:58:47 : Application.logger (INFO) Found 15722
surf points in image 2.
2015 Aug 31 13:59:14 : Application.logger (INFO) Found 740
homologous points.
2015 Aug 31 13:59:14 : Application.logger (INFO) 0 points discarded
$ otbcli_HomologousPointsExtraction -in1 IMG_0062.tif -band1 2
-in2 IMG_0062.tif -band2 1 -mode full -precision 50 -mfilter 1
-algorithm surf -out IMG_0062homSURF2.txt -threshold 10
2015 Aug 31 13:59:26 : Application.logger (INFO) Elevation
management: setting default height above ellipsoid to 0 meters
2015 Aug 31 13:59:26 : Application.logger (INFO) Doing update
2015 Aug 31 13:59:32 : Application.logger (INFO) Found 15650
surf points in image 1.
2015 Aug 31 13:59:38 : Application.logger (INFO) Found 15722
surf points in image 2.
2015 Aug 31 14:00:07 : Application.logger (INFO) Found 15650
homologous points.
2015 Aug 31 14:00:08 : Application.logger (INFO) 11469 points
discarded
Thanks!
Agus
On Wed, Jun 10, 2015 at 2:52 PM, Julien Michel
<[email protected] <mailto:[email protected]>> wrote:
Hi Agus,
The readthedoc documentation export is experimental (Rashad
did this a while ago), and surely has some unsolved issues
for now. If you refer to the official documentation here :
https://www.orfeo-toolbox.org/CookBook/CookBooksu103.html#x136-6690005.6.7
You will see that the table is right. For instance :
precision
Float
Estimated precision of the colocalisation function (in pixels).
mfilter
Boolean
Filter points according to geographical or sensor based
colocalisation
Regarding the precision parameter documentation, it is used
at several steps of the algorithm, as described in the
Detailed Description of the application :
"In this mode, the corresponding spatial bin in the second
image is estimated using geographical transform or sensor
modelling, and is padded according to the user defined
*precision*."
"Last, in both modes the application can filter matches whose
colocalisation in first image exceed this *precision*."
But we can update the documentation of the parameter if you
think this is not clear enough.
Regards,
Julien
Le 10/06/2015 10:00, Agustin Lobo a écrit :
Doc on Homologous Point Extraction
http://otbcb.readthedocs.org/en/latest/Applications/app_HomologousPointsExtraction.html
is incomplete.
For example, the only information provided on argument mfilter
is that it is a Boolean. No definition, no default value, no example.
Instead, the command line help page is more informative:
"-mfilter <boolean> Filter points according to
geographical or sensor based colocalisation (optional, off by
default)"
although it is not clear if the user has to write "on" or "1" to turn
it on. And it is not clear either what kind of filter (mean filter?)
and which radius
Regarding "precision", there is also no definition in the table (the
information on
Parameter Type is just repeated) but it is later defined as in the
command line help page:
"Estimated precision of the colocalisation function (in pixels).
Estimated precision of the colocalisation function in pixels."
I do not say this definition is wrong, but it is no appropriate. The
definition must
state the role of the argument in the process. In this case, I
understand that the
algorithm searches, for a given point image in1 band1, an homologous
point on
image in2 band2 within a radius equal to the value of precision in
pixels.
Correct?
Also, what happens if the user sets precision to i.e. 10 and mfilter is
off?
Thanks
Agus
--
Julien MICHEL
CNES - DCT/SI/AP - BPI 1219
18, avenue Edouard Belin
31401 Toulouse Cedex 09 - France
Tel: +33 561 282 894 - Fax: +33 561 283 109
--
--
Check the OTB FAQ at
http://www.orfeo-toolbox.org/FAQ.html
You received this message because you are subscribed to the
Google
Groups "otb-users" group.
To post to this group, send email to
[email protected] <mailto:[email protected]>
To unsubscribe from this group, send email to
[email protected]
<mailto:otb-users%[email protected]>
For more options, visit this group at
http://groups.google.com/group/otb-users?hl=en
---
You received this message because you are subscribed to the
Google Groups "otb-users" group.
To unsubscribe from this group and stop receiving emails from
it, send an email to [email protected]
<mailto:[email protected]>.
For more options, visit https://groups.google.com/d/optout.
--
Julien MICHEL
CNES - DCT/SI/AP - BPI 1219
18, avenue Edouard Belin
31401 Toulouse Cedex 09 - France
Tel: +33 561 282 894 - Fax: +33 561 283 109
--
--
Check the OTB FAQ at
http://www.orfeo-toolbox.org/FAQ.html
You received this message because you are subscribed to the Google
Groups "otb-users" group.
To post to this group, send email to [email protected]
<mailto:[email protected]>
To unsubscribe from this group, send email to
[email protected]
<mailto:otb-users%[email protected]>
For more options, visit this group at
http://groups.google.com/group/otb-users?hl=en
---
You received this message because you are subscribed to the Google
Groups "otb-users" group.
To unsubscribe from this group and stop receiving emails from it,
send an email to [email protected]
<mailto:[email protected]>.
For more options, visit https://groups.google.com/d/optout.
--
--
Check the OTB FAQ at
http://www.orfeo-toolbox.org/FAQ.html
You received this message because you are subscribed to the Google
Groups "otb-users" group.
To post to this group, send email to [email protected]
To unsubscribe from this group, send email to
[email protected]
For more options, visit this group at
http://groups.google.com/group/otb-users?hl=en
---
You received this message because you are subscribed to the Google
Groups "otb-users" group.
To unsubscribe from this group and stop receiving emails from it, send
an email to [email protected]
<mailto:[email protected]>.
For more options, visit https://groups.google.com/d/optout.
