Progress Log.txt

Apr 22:
Begin pseudocode, look at getting Baxter Dataset

Apr 23:
Completed high level pseudocode.
Completed basic data storage spec for cost volume. 
Pseudocode continues. 
Spent time looking at example rendering code, postponing usage until later.
(Maintenance) Install PTAM, Newest OpenCV, ROS Hydro on new workstation.

Apr 24:
    Level 2 pseudocode (nearly runnable, full idealized interfaces) begins
    Discovered two major issues: 
    1. PTAM is not compatibly licenced with OpenCV
    2. DTAM wants threads to run properly, but OpenCV does not allow 
        internal threading. 
    Spent most of day trying to figure out how to resolve these issues.

Apr 25:
    Devised new design for program to stay within guidelines
    DTAM will consist of a Mapper and Tracker. These will run as seprate threads
    from the rest of the program. The threads are thin wrappers around 
    the Map and Track classes, which are the internal machinery of DTAM, but without
    any internal driving threads. These classes are the ones that will be in OpenCV
    and it will be the user's responsibility to ensure that they are powered by 
    seperate threads, or by a single thread so fast it avoids dropping frames.

Week 2
Apr 28:
    Nearly completed level 2 pseudocode on map side. Just need the projection step and a way 
    to automatically detect new keyframes. So overall we're ~ 1 day behind schedule,
    which is okay because of flexibility built in, but still a bit unnerving.

Progress Report 
End of Week Two (Pseudocode Phase)

    Schedule was to have pseudocode for both tracking and mappig done by end of two weeks. 
    Initially had a setback of a day on mapping, but tracking was completed a day 
    faster than expected.
    
    Overall, pseudocode was completed, except that reprojection was not explicitly written. On the 
    other hand, a working Matlab version of the core tracker algorithm was completed, and works on 
    synthetic data. I have discovered that the ESM tracker appears to be most effective when the 
    amount of motion is already nearly known for synthetic images. For whatever reason, convergence
    appears to be much better on real images though.
    
    Next three weeks are writing up actual code.
    


DTAM (Dense Tracking and Mapping) GSoC Progress Report, Week 3:

    Since this is my first blog posted progress report, I'll do a quick recap:
    -Weeks 1 and 2 were writing a pseudocode skeleton of the program, both tracking and mapping.
    -Pseudocode for tracking turned out to be runnable Matlab, so I have a working implementation of the 
    tracker core, though of course it still needs to be converted to OpenCV.
    -Overall schedule was maintained for weeks 1-2.
    
    Week 3:
    -This week was the first "real" coding week, and the first of three assigned to writing the base CPU 
     implementation of the mapper. The plan was to write reprojection and optimiztion in week 3, 
     infrastructure and API code in week 4, and compile-test-debug week 5.
    -Overall, two days were lost due to unexpected "writer's block" on the reprojection algorithm. 
     This is just barely within the allocated "flex time" in the schedule.
        Issues stemmed from the initializtion procedure for the cost volume, where it was unclear which
        depth steps to choose because it is impossible to determine true scale with a monocular system. 
        As a stop-gap, it was decided to use 64 inverse depth steps starting from half the distance of the
        nearest estimated point in the initializaiton, and a base scale of the largest motion observed in 
        first 10 s of the input or 1/10th the distance to the nearest point, whichever is smaller.
        There remains an issue of how to deal with cells and rows in the cost volume that have never been
        observed. For now, they are simply left out of the optimization.
    -The plan for week 4 is to catch up to schedule. By the end of week 4 code should be in a 
     ready-to-compile state.
    -Doc. VR pointed out a pair of upcoming works at ICRA, REMODE and SVO, which provide capabilities 
     addressing the same needs as DTAM and will also be open source. We are looking into a 
     collaboration, and a way to differentiate our product from theirs.