Skip to content

Latest commit

 

History

History
32 lines (28 loc) · 2.51 KB

CHANGELOG.md

File metadata and controls

32 lines (28 loc) · 2.51 KB
Raw

13/09/17

  • Added emergency mode (setting threshold lower when score < 0, so that tracking snaps back on after a full occlusion)
  • Added some more adaptability to the program
  • Added more ideas in comments ...
  • Implemented first (untested) darknet usage through shell
  • Forgot to say that I added a gitignore for the video file

14/09/17

  • Gitignore for the video file is now working, original video: https://www.youtube.com/watch?v=PNCJQkvALVc (for the full occlusion test, I added a black bar in the center to fake the occlusion)
  • Added tracker.py as class, that runs one MDNet for tracking and gets fed with a frame and values :)
  • Added variable length of result & result_bb in Tracker.py so that we're able to do a frame by frame analyzation
  • Refactored the Tracker.py class with the code from run_tracker.py so that self.* contains class specific variables like "model"
  • fixed bug where pytorch would try to use gpu even though opts['use_gpu'] equaled False
  • Realized that my Laptop is way too slow for this kind of work
  • Split model code into two functions and got it working so that frame by frame analyzation is able to work (seems slower though, have to double check)
  • Figured that I need asynchronous callbacks for the Tracker.py class (namely: startTracking() and updateFrame()) so that the code doesn't get blocked by multiple trackers at once, also because then I can have a callback that feeds the next frame into each Tracker because they will definitely not run equally fast

15/09/17

  • Figured that indentation is important in python (srsly, spent 20 minutes on figuring out why some functions didn't work)
  • Async sucks (as it already did in JS, but at least I knew how to properly do it there)

18/09/17

  • Async doesn't suck as much, thought the bug was async-related, was actually python3 related. So ... Python sucks (i guess?) -> fixed it though, iteritems()->items()
  • Also want to note that OSX sucks. Trying this on a Windows computer now, if that doesn't work I'll install Ubuntu ...

25/09/17

  • got a pc I can code on now
  • fixed a bug where numpy.int64 was given but float was needed
  • fixed some small bugs where class-specific variables weren't referenced properly
  • ran first frame-by-frame analyzation, worked out fine, but I broke the output somehow -> doesn't matter though
  • did matter, because result didn't contain the result ... let's just use result_bb :)
  • tried running every 3rd frame, everything above fails after some time -> added a TODO to check whether the network has to get to know the object more at first