To detect the same picture, why is AB-version much better than PJ-version?

[pascal1129]

I'm interseted in human detection. I detected the same picture both with AB-version and PJ-version, i found AB-version is much better, i,m quite confused, what are differences between the two versions?

Here is my test picture:

I uesd origin pj-version:
git clone https://github.com/AlexeyAB/darknet.git
wget https://pjreddie.com/media/files/yolov3.weights
I set Makefile: GPU=1, CUDNN=1, OPENCV=1, OPENMP=1, DEBUG=0

The serult is notgood:

Then, i run ab-version:
git clone https://github.com/pjreddie/darknet
wget https://pjreddie.com/media/files/yolov3.weights
I set Makefile: GPU=1, CUDNN=1, CUDNN_HALF=0, OPENCV=1, AVX=1, OPENMP=1, LIBSO=0

The result is perfect to me:

So why is so differnent, i,m quite confused.

[fabioac]

pascal1129,
whch GPU are you using? when you use CUDNN_HALF cause you to improve detection speed 3x on Titan V / Tesla V100 / DGX-2 and later. See in instruction on how to compile on linux.

[pascal1129]

@fabioac
GTX1080ti * 2, Ubuntu 16.04

[fabioac]

set CUDNN_HALF=0 and run again to try

[pascal1129]

@fabioac In fact, that's what i set, you can see it again.

[AlexeyAB]

@pascal1129
May be due to these differences: #232 (comment)
In my resize-approach objects sizes are larger than in the original resize-approach(letterbox).