Skip to end of metadata
Go to start of metadata

2012 02 14

Hasan:

  • continued deformancy eval on type 4
  • code is more or less finished
  • need to merge it with older code
  • run it with nightly
  • using small data set of full book to egnerate results
  • 60 pages
  • will send some images

Zeshan:

  • feature based dewapring
  • based on text lines
  • will be integrated into final product using SIFT
  • optimizing SIFT approach for documents

Martin:

  • continued on 3d dewarping and improving it.
  • results look okay for dewarping
  • texture mapping is difficult but progressing

Jon:

  • continuing work on Exporter UI implementation and help with implementing export UI
  • will start working on a decapod genpdf spec to help plan work better.

Justin:

  • implementing the export UI
  • found bug in Uploader
  • having problem with Mercurial, not allowing to add files

2012 02 07

martin:

  • 2d dewarping algorithm implemented
  • algorithm can be used as a post processing function for 3d
  • distance to camera is neglected currently
  • with 3d dewarping, the page curl eliminated (see images)
  • working on a 3d approach hopefully for next week
  • the issue with 3d approach is texture mapping images to model
  • have a new approach but requires calibration using 6 points
  • these 6 points will give distance information

http://iupr-serv5.cs.uni-kl.de/scr1.png http://iupr-serv5.cs.uni-kl.de/scr2.png

Zeshan

Hasan:

  • genpdf - working on performance evaluation for type 4
  • have some results. Code has yet to be integrated with genpdf.
  • working on genpdf to work on larger books 60+ pages to see how constructed fonts look

Jon:

  • exporter redesign
  • DPI testing this week
  • email list about Ocropus upgrade

Justin:

  • server code modifications complete with unit tests
  • waiting for code review
  • looking into cherrypy installer options

Colin:

  • will look into the server code.

2012 01 31

martin:

  • background and foreground extraction
  • working on 3d approach, not successful so far
  • determine content by looking at image data and not 3d approach.

zeshan

  • experimenting on dewarping
  • dewarping is working, but not working exactly.

hasan:

  • working performance measure for type 4
  • running into issues with large books

justin

  • code review of server-side code
  • jon to test it out
  • server side validation for image types
  • updating install scripts for cherrypy no package for 10.4, using PIP

jonathan

  • continuing work on exporter UI design
  • will email list about upgrade to Ocropus.

2012 01 24

Present: Martin, Zeshan, Justin, Jon, Hasan

Dewarping:

  • working on sift features and outliers
  • working well and should work on a wide range of books
  • Stereo features part is pretty much done.
  • Dewarping still needs to be done for stereo
  • Working on segmenting the book from the background into automated procedure.
  • Dewarping working with perfectly appplicable surfaces
  • not producing good output.
  • working on procedures to work on near perfect surfaces
  • perfectly applicable surfaces - page that easily be transformed back into a plane
  • in reality we won't have perfectly applicable surfaces.

Genpdf:

  • resolution option: -r option to control image resolution before embedding
  • available in genpdf
  • token based pdf performance index now available.
  • creating performance measure for type 4 PDF
  • DPI only.

Decapod UI:

  • server's Cherrypy upgrade complete and is essentially working to what it was in 05a
  • writing unit tests
  • continuing Decapod 0.5 design work.

2012 01 17

Present: Zeshan, Justin, Jon, Hasan

Dewarping

  • Stereo reconstruction has been tested on a larger dataset. It had some problems in surface fitting. The model have been generalized to be able to cope up with variety of documents.
  • For Structured Light, the work is in progress for the automatic extraction of region of interest and separation of pages.
  • For Dewarping, conformal mapping has been fully implemented and tested. It works with perfectly applicable surfaces* (not always possible to reconstruct perfectly applicable). At the moment Implementing another dewarping method for nearly applicable surfaces. It will help in determining the final model which should be incorporated in Decapod.
  • Upgrade Numpy 0.1.1 -> 0.9

Genpdf

  • completed objective performance evaluation img2pdfper.py measurement for type 3 or 4
  • creates a pdf file that also lists their performance index
  • larger number more variance between the original character and the generated token.
  • Generates an index number for entire image.

Server

  • changes for RESTful architecture
  • implementing new scheme for dynamic URLs
  • code cleanup

UI

  • Designs progressing on new Export UI.

Q:
Is it possible to create a test to detect handwriting? (This way Decapod can quickly decide whether or not Type 2 or higher will succeed or not).

A:
If the objective is to create segmentable hand written text, it's possible but requires work. You would create a model based on specific hand writing but it will only work for one particular hand writing style.

If the objective is to detect which type of PDF to generate, then there have been discussions on such a tool, but not in the near future.

Genpdf:

  • downscaling and upscaling PDF output to be investigated
  • colour output option to be investigated

2012 01 10

Present: Martin, Zeshan, Justin, Jon.

Dewarping

  • Experiments giving good results
  • Will put all software pieces into a software module
  • Some parts are still manual process, but can be automated.
  • tests run using data from structured light and stereo.

Decapod Webapp

  • going over notes from recent Nigeria trip and figure out what new features will be added to the roadmap
  • architecture: implementing restful architecture requires changing CherryPy. Upgrading causes the server to stop working. Working through that.

New Ubuntu

  • need to decide if we will move to Ubuntu 12.04 LTS

New Ocropus installer

  • a new installer is available. Supposed to be better.
  • likely will break genpdf. Will wait until Hasan is back before we decide to move to that version.

Depenedencies

  • need a list of dependencies.

2011 12

No weekly meetings due to holidays.

2011 11 08

Present: martin, hasan, zeshan, jon, justin.

genpdf (hasan, jon)

3D models (martin and zeshan)

  • approximating real book surface based on stereo matching
  • smoothing out book shape and removing outliers
  • hope to get dewarping results soon
  • working on conformal mapping (See: http://en.wikipedia.org/wiki/Conformal_map) which will be part of the dewarping process

Test server (martin, hasan)

  • working continuing on test server

Decapod 0.5 (jon and justin)

  • re-prioritized features for the release to accomodate work needed for server architecture
  • Decapod 0.5a will be a simple application that imports images and outputs PDFs.
  • Decapod 0.5 will add page management.

Decapod Application Server (justin)

  • working on server architecture

Decapod Application Client (jon)

  • Working on UI implementation of Decapod 0.5a release

2011 10 25

Present: Martin, Hasan, Jon, Justin

genpdf (Hasan)

  • moved from autotrace to potrace
  • potrace produces much nicer results
  • experimenting with scaling images up then doing the trace. - get better results
  • straight lines end up looking better when scaled up then traced.
  • [see this PDF for the results|^gendpdf 2011-1024-28 (Decapod).pdf].

stereo (Martin)

  • working on stereo
  • no new results because of a report that was due last week
  • working on handling non-text regions next(i.e. images)

testing server (Martin)

  • still working debugging the testing machine.
  • removed most of the test images, but some images take more than 24 hours to generate a PDF.

structured light (Zeshan)

  • current RGB pattern can be occluded and in this case the algorithm can produce bad results
  • implementing more robust pattern detection that doesn't have this problem.
  • looking at a new dewarp algorithm as well.

Decapod Webapp

  • Import page is basically up and running, it will upload files to the server and store them on the file system named with UUIDs
  • Beginning to start the Image Management page, Currently working on rendering and reordering the thumbnails.
  • Need to think a bit about the model (the data to store the info about the pages of a book) and something to mediate the communication between the server and client on the client side
  • Ongoing UI styling and design

Cutting versions of Ocropus and genpdf

  • will need to cut tags of ocropus and genpdf soon
  • Jon will test genpdf a bit more and then determine a deadline to cut genpdf.
  • Martin to talk to Thomas about taggin ocropus versions because there has been a lot of refactoring in the past 3 months.

Administration

  • Propose to switch meeting dates to Tuesday same time

2011 10 18

Present: Martin, Zeshan, Hasan, Jon, Justin, Jess

stereo based reconstruction (Martin)

  • work is ongoing in creating good 3D models using stereo.

structured (Zeshan)

  • perfect 3d
  • working on improving orientation at capture time so it can capture at different angles
  • can share repo so we can test it out.

nightly test server (Martin)

  • memory consumption of tests crashes server
  • will stabilize it next week
  • in meantime scaled back the number of evaluations.

PDF generation (Hasan)

  • corrected errors mislaignment and type 3 and 4.
  • uploaded to bitbucket
  • improving quality of font reconstruction
  • considering to switch from autotrace to a different library

Decapod Server (Justin)

  • working on uploading images to the server. Progress is good.

Decapod UI (Jon and Justin)

  • working on design changes for the main Image Management UI
  • created a new Import UI
  • going to wire up the Import UI to the server so we can have uploading working.

Decapod Extension (Jess)

  • Proposing a no-cost extension to Nov 2012
  • financials have all been sorted out.

Notes from conversations between 2011 09 20 - 2011 10 14

  • Testing latest versions of Hasan's genpdf code.
    • identified some issues which has been addressed by Hasan's latest commit. Will need to be tested to verify.
    • Further testing required for Type 1 and Type 2 PDF to ensure reliability for 0.5 deliverable.
  • Migration of git has been scrubbed. We will continue to use Mercurial and bitbucket as a method of collaboration.

2011 09 19

Present: Martin, Zeshan, Hasan, Jonathan, Justin

Genpdf:

  • Current version in repository is not working.
  • Hasan has been working on fixes and will email Jon a copy.

Ubuntu version:

  • Currently 11.04 32bit is the platform to use.
  • However, the version may change depending on the version of scipy required for 3D modelling and dewarping (Martin and Zeshan to inform on this).

3D model:

  • Stereo progressing, but results are not yet as good as structured light.

2011 08 15

Present: Martin, Zeshan, Hasan, Jonathan

Decapod test suite:

3D models:

  • dense disparity was generated using the Fuji W3 in an uncalibrated setup, structured light experiments are yet to come.
  • results are online on Google+
  • now to get a mesh and start dewarp.

Genpdf:

  • Hasan testing genpdf. Found only working with Type 1 PDFs, and not type 2.
  • Jon suggested to try decapod-genpdfEXP.py inside the decapod-genpdf directory and use ocropus 0.4.4
  • Hasan is making contact w Michael Cutter to help understand work done

Dewarping

  • Zeshan has dense maps and working on dewarping
  • hope to have something by end of week

Administrative

  • Tom back in october 13/14/15
  • majority of IDRC team away due to Masters program launch

2011 08 08

Present: Martin, Zeshan, Hasan, Justin, Jonathan

3D progress:

  • Performing experimentation on pixel depth calculation
  • we have reasonable sparse 3d reconstructions using stereo images from the W3 (structured light is currently not yet in a usable state) and currently work on producing a dense mesh out of the 3d points for texture mapping & dewarping
  • dense reconstruction is currently not yet working good enough for using it in the pipeline, but we are in parallel trying to get some reasonable results there
  • Shan switched to work on dewarping of stereo images to get something deliverable for the demo (we'll investigate structured light after we have something demonstrable there)
  • Aim to have something deliverable for Decapod extension meeting (sometime in September 2011?)

Gen PDF testing

  • progressing on testing gen pdf using clean computer generated documents
  • results aren't particularly good for Type 2 or Type 3 PDF
  • To do Martin to find out from Ocropus devs what test samples would yield the best results
  • To do Jonathan to post findings to list and find out if we need to upgrade to Ocropus tip to get better results

Roadmap discussion

  • Jonathan went over the Milestones, in particular how it relates to deliverables for Nigeria.

Introducing Hasan

  • new to the iupr team
  • will be working on font grouping and pdf generation
  • Hasan has been added to the Decapod mailing list

2011 07 11

Present: Martin, Zeshan, Jess, Thomas
Decapod Fluid Toronto team is participating in the a11y hackathon

Solutions out there that smush books:

  • Roebook
  • Ion
    like book liberator, but still uses mechanical pressure on book to flatten rather than stereo capture
  • book liberator

Martin:
received the stereo camera – the W3?; doing calibration; new laptop set it up
will have stereo by September.

Zeshan:
waiting for projector to come and taking images from web and building software can acquire own images. process those images – structured light using existing algorithms

Thomas:
doing work on OCRopus
get done before summer ends – eliminate C++ library since it's not needed and replace with 1 or 2 smaller python libraries. image processing code that python already covers or is easier manipulated in python.
C++ command line programs will disappear
thinks everything uses Python CLI programs anyway

3D stuff should have new functionality by Sept. and old part should still be working…

flat originals should be able to capture just fine with Stereo stuff.
if layout is other than single or double column. One strategy with other layouts could be to give the user a tool where they can label the layout.
The solution can find text lines, but can't find reading order. If the user has labeled the reading order that would help.

For Decapod we're focusing on just generating PDFs.
always will get a PDF that looks ok – we promised to generate a nice PDF
OCR and layout won't work well – didn't promise great OCR because we can't do that.
even w/o layout analysis – it's still searchable.
layout is needed for reflow – reformatting for mobile e.g.

2011 07 04

Present: Martin, Zeshan, James, Jonathan

  • JH testing Decapod to determine its current capabilities.
  • Book Liberator no longer an active project.
  • Book Saver product similar to Book Liberator, may be useful for left-right capture.
    • JH to send email to determine specifications on product.
  • Fuji W3 camera received:
    • possible to create a preset calibration profile that should more or less work for all W3 cameras.
  • Logitech C900 HD webcams purchased to use with structured light (1080P, 10MP) - used for capturing book details.
  • Projector has not arrived.
  • Awaiting laptops on July 17th to begin testing the W3 and webcam hardware.
  • Questions to be answered:
    • Should we update to latest gphoto (currently on gphoto 2.4.9)?
    • Is Ocropus 0.4.4 the version we should still be using?

2011 06 27

  • Focus for development for next year by TMB's team will be structured light capture.
  • See Decapod User Workflow with Structured Light for overview how this process may look like.
  • Aim to have an initial python script that "does something" in 3 months.
  • Will be using a Fuji W3 3D camera to see if it's viable for Decapod's current Stereo capture method.
  • Stereo capture is still a viable mode of capture, but the complexities make structured light more attractive.
  • Possible capture modes for future consideration:
    • 90-degree book. 2 cameras: 1 page captured per camera. 3rd low resolution camera (i.e. webcam) to calculate depth.
    • Multiple camera capture. An array of cameras each capturing a different section of the book surface. Reassembled by software.
    • Light field capture – very promising and effective, but cost prohibitive.
  • No labels