Possible future enhancement NPD about jda HOT 7 OPEN

Thanks a lot!

from jda.

@inspirit
Hi, I'm studying face alignment now. I have replaced conventional pixel difference by normalized pixel difference. However, it seems that NPD made no contribution to accuracy. Is there any part I need to modify to match the new feature? Thanks!

from jda.

this is not that easy as just using NPD instead of pixel diffs. you should take into account that response values will be different so you should select threshold more careful. please refer to the paper i mentioned above

from jda.

Actually, I have selected threshold from -0.25 to 0.25 when using NPD, while -64 to 64 when using PD. But I made no improvement. Anyway, thanks!

from jda.

in my implementation i'm generating random splits with 2 thresholds one negative and one positive.
threshold values are in range -127 to 127. so for each split i try 10 random thresholds * 30 random pixel positions, and then select the one that reduce the error better than others.
the split check against 2 thresholds is:

const float npd = response(feature_responses);
// thresh1 is negative, thresh2 is positive in [0,127) range
return npd < thresh1 || npd > thresh2;

from jda.

Thank you very much, I have read the paper. Originally I thought that double thresholds were only used in duadratic tree. Now I figure it out. Thanks!

from jda.

I have some questions too..:
const float npd = response(feature_responses);
Do you used response (f1-f2)/ (f1+f2) or more complicated function?

// thresh1 is negative, thresh2 is positive in [0,127) range
return npd < thresh1 || npd > thresh2;
What scale is used for npd weighing?

from jda.