[pgpointcloud] RLE and SIGBITS heuristics
Rémi Cura
remi.cura at gmail.com
Fri Apr 17 05:43:12 PDT 2015
Hey,
For instance in every Lidar point cloud, the order is paramount, because
the sensing is structured (lines usually).
So ordering can give you neighbourhood relation for free.
Also, ordering may be important when you want to improve your point cloud
registration .
I don't mean point cloud global geolocalisation, but each point
localisation in the sensor reference system.
Typically points are sensed in the sensor reference system, then knowing
the sensor trajectory in WGS, you put the point in WGS (for instance).
Sadly sensor trakectory is always false to a point. Correcting it allow to
re-register the points and improve points precision.
Sorry my answer was a little cryptic.
(I talk about LOD in [this presentation](
https://github.com/Remi-C/Postgres_Day_2014_10_RemiC/raw/master/presentation/A%20PostgreSQL%20Server%20for%20Point%20Cloud%20Storage%20and%20Processing.pdf)
, p49 and 50)
For ordering points I build an octree, then for each level of the octree (0
to N), I take the point closest to the center of the cell.
This gives you at most 8^K points at the level K.
(here illustration in 2D on real data. Biggest point = most important point
= lower Level point. )
I don't build an octree in fact, because you can get what point goes in
which cell simply by looking at the binary coordinates of the points (after
offset and scale).
This is cool if the user read the whole level.
But you can do something better by allowing continuous-LOD.
In this scenario the user can read only a part of a level, and we still
want that he gets well spaced points.
For instance the level 2 would have 64 points, maybe the user read points 1
to 30.
If you order points within a level by morton curve, the user will only get
points in the upper part of the patch (for instance).
So the user would have no points in the lower part of the patch !
We want to do precisely the opposite, and that the user gets points well
spaced in the patch.
I used random because it was easy, but it is not deterministic.
The inverse morton code provides a cool order to dispatch points and kinda
solve this problem (as well as being easy & fast to write in low level
langages).
We could do more subtle I'm sure.
Cheers
2015-04-17 13:07 GMT+02:00 Peter van Oosterom <P.J.M.vanOosterom at tudelft.nl>
:
> Hi Rémi,
>
> Good to hear that other persons find our work of interest. Thanks!
>
> Fair remark about the fact that original order of points can be
> significant.
> (could you give one example where this is really clear?)
>
> I do not understand what you mean with 'for LOD I considered using an
> inverse morton curve'.
> Options for selecting points for higher LODs could be random (not so
> smart, but easy) or some spatial analysis (figure our relative relevance/
> importance of points, smarter, but non-trivial). So, in general two nearly
> equal (very close) points should not both go to higher LOD, so better look
> for the different point to push to higher level. You could do this with
> real distance in 3D space, but it may be easier with distance of 1D morton
> code. So, when points are processed (in original sequence), you compute
> their morton code, and when next point is significantly different wrt
> morton, it goes up (and otherwise stay at lowest level). Is this what you
> mean with 'inverse morton for LOD'?
>
> We really need LOD and smart data organization for efficient use, current
> test of 640 billion points we need it (and there is a 30 trillion data set
> announced for Netherlands, which we also would like to use efficiently in
> interactive manner).
>
> Kind regards, Peter.
>
>
> On 17-4-2015 12:35, Rémi Cura wrote:
>
> Hey Oscar (and Peter),
> great series of articles, the benchmark was really interesting !
>
> I personally am adamantly against reordering the points for storage
> purpose, because the order can have a strong meaning (physical or
> theoretical).
>
> For instance I re-ordered the points of my patch so that the points
> follows a LOD schema.
> So reading more and more points gives a better and better approximation
> of the patch.
> This gives free LOD facilities when all the patch have been re-ordered
> (free CPU wise, and no data duplication).
>
> Now I have to admit that efficient storage and LOD are in conflict,
> because storing is about putting the similar points together somehow, and
> LOD is about the opposite (kind of sequential vs random access).
> This has a strong impact, for instance reordering patches randomly will
> greatly increase patch storage size in pgpointcloud.
> It is funny because for LOD I considered using an inverse morton curve
> (interleave X Y Z , then read is backward ! ).
>
> I'm well aware that we could simply store the original order in an
> attribute, and so compression based on morton on XYZ might be transparent
> for the user.
>
> Anyway, just to say that pgpointcloud is ultimately not about storing
> points, but easing point cloud usage in my opinion.
>
> Cheers,
> Rémi-C
>
>
>
>
> 2015-04-17 11:02 GMT+02:00 Oscar Martinez Rubi <o.martinezrubi at tudelft.nl>
> :
>
>> Hi,
>>
>> About the XYZ binding for better compression. In our research in the NL
>> escience center and TU Delft we have been thinking (not testing yet though)
>> about one possible approach for this.
>>
>> It is based on using space filling curves. So, once you have the points
>> that go in a block you could compute the morton/hilbert code of the XYZ.
>> Since all the points are close together such codes will be extremely
>> similar, so one could store only the increments which could fit in many few
>> bits. We have not tested or compared this with any of the other
>> compressions but we just wanted to share it with you just in case you find
>> it useful!
>>
>> An additional improvement would be to sort the points within the blocks
>> according to the morton code. Then, when doing crop/filter operations in
>> the blocks one can use the morton codes for the queries similarly to what
>> we presented in our papers with the flat table (without blocks), I attach
>> one of them (see section 5.2). In a nutshell: You convert the query region
>> into a set of quadtree/octree nodes which can be also converted to morton
>> code ranges (thanks to relation between morton/hilbert curve and a
>> quadtree/octree). You scale down the ranges to increments (like you did
>> when storing the point of the block) and then you simply do range queries
>> in sorted data with a binary algorithm. In this way you avoid the
>> decompression of the morton code for most of the block. This filtering is
>> equivalent to a bbox filter so it still requires a point in polygon check
>> for some of the points.
>>
>> Kind Regards,
>>
>> Oscar.
>>
>>
>>
>> On 16-04-15 18:15, Rémi Cura wrote:
>>
>> epic fail ! I had avoided html just for you
>>
>> Dataset |subset size | compressing | decompressing |
>> |(Million pts)|(Million pts/s)|(Million pts/s)|
>> Lidar | 473.3 | 4,49 | 4,67 |
>> 21-atributes | 105.7 | 1,11 | 2,62 |
>> Stereo | 70 | 2,44 | 7,38 |
>>
>> Cheers
>>
>> 2015-04-16 17:42 GMT+02:00 Sandro Santilli <strk at keybit.net>:
>>
>>> On Thu, Apr 16, 2015 at 05:30:12PM +0200, Rémi Cura wrote:
>>> > OUps
>>> >
>>> > Dataset | subset size(Million pts) | compressing (Million
>>> pts/s) |
>>> > decompressing (Million pts/s)
>>> > Lidar | 473.3 | 4,49
>>> > | __4,67__
>>> > 21 attributes | 105.7 |
>>> > 1,11 | 2,62
>>> > Stereo | 70 | 2,44
>>> > | 7,38
>>>
>>> These tables aren't really readable here.
>>> Could you make sure to use a fixed-width font to write those tables
>>> and to keep lines within 70 columns at most ?
>>>
>>> --strk;
>>>
>>
>>
>>
>> _______________________________________________
>> pgpointcloud mailing listpgpointcloud at lists.osgeo.orghttp://lists.osgeo.org/cgi-bin/mailman/listinfo/pgpointcloud
>>
>>
>>
>
>
> --
> Peter van Oosterom P.J.M.vanOosterom at tudelft.nl
> Section GIS technology (room 00-west-520) Department OTB
> Faculty of Architecture and the Built Environment, TU Delft
> tel (+31) 15 2786950 Julianalaan 134, 2628 BL Delft, NL
> fax (+31) 15 2784422 P.O. Box 5043, 2600 GA Delft, NLhttp://geomatics.tudelft.nl MSc Geomaticshttp://www.msc-gima.nl MSc GIMA (Geo-Info Management&Appl)http://www.gdmc.nl
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/pgpointcloud/attachments/20150417/4b7fbf49/attachment.html>
More information about the pgpointcloud
mailing list