These are chat archives for menpo/menpo

Jun 2014
Nicolas DUBIEN
Jun 03 2014 17:34

Unexpected behaviour in DalalTriggsHOGdescriptor from repository menpo/menpo??

Line 294: block(double *) is set to 0.
Line 413: the loop can change the value of block[i][j][k]

We can imagine that for a given value of (x, y), block[i][j][k] is modified at line 426
But that for another value (after) (x', y'), block[i][j][k] is not modified (no else statement)
That means that the value of block[i][j][k] will not be the one corresponding to this (x', y') (and not 0.) but will still be taken into account for the remaining part of the process.

Is it really the expected behaviour ?

James Booth
Jun 03 2014 20:40
  1. what’s wrong with nosetests ./menpo/image/test/ You could configure your IDE to run these tests, or write a trivial Makefile. There is no more automated way than the above command though.
2 - Let’s discuss the meeting via email (will respond in a mo)
James Booth
Jun 03 2014 20:46
3 - really keen to meet up and discuss the progress you have already made. It would be good if you could brief us in that meeting about how you are looking to integrate CUDA with Python - this is a key design decision so something I want to get right. Once we are all happy I’d like to see a PR adding CUDA support to Menpo (or perhaps onto a special cuda branch for the time being) so we can move forward with small PR’s adding new CUDA functionality.
4 - Questions about the HOG implementation are best handled by @nontas who wrote the code - I’ll ask him to get back to you. Unfortunately he has injured his hand recently so might be out of action for a day or two!
Nicolas DUBIEN
Jun 03 2014 22:39
Thanks for your answers.
  1. done, I finally find how to use nosetests to carry out tests on image/features
  2. for the meeting I'm available whenever you want this week and next week. The sooner the better as it will help me to know what to do, how to test precisely..
  3. for the moment I created a file called to replace HOG.cpp. It contains a basic implementation of HOG algorithms in CUDA. I still need to check that the output remains the same as the one you get with your version but it compiles and passes tests.
  4. for questions on HOG I will refer to @nontas . Hope that your hand is going better
See you soon ^^