Different confidence results between pjreddie/darknet and alexeyab/darknet and wrapper

[ghost]

Hello,

I trained yoloV3 with alexeyab/darknet with a custom dataset in order to get a person detector on thermal images.

Database : FLIR thermal images avaiable at https://www.flir.com/oem/adas/adas-dataset-form

Change made in yolov3.cfg :

number_filters -> 18
width,height -> 608,608

I test on 3 specific images :

FalsePositive1.jpg : no object FalsePositive2.jpg : no object TruePositive1.jpg : 2 objects

I test theses images with alexeyab/darknet

FalsePositive1.jpg where alexeyab/darknet detects a false-positif with a confidence of 27% Falsepositive2.jpg where alexeyab/darknet detects a false-positif with a confidence of 56% TruePositive1.jpg where alexeyab/darknet detects 2 true positif with 100% confidence

I test theses images with pjreddie/darknet

FalsePositive1.jpg where pjreddie/darknet detects nothing, no false-positif Falsepositive2.jpg where pjreddie/darknet detects nothing, no false-positif TruePositive1.jpg where pjreddie/darknet detects 2 true positif with 100% confidence

Besides I test theses images with alexeyab/darknet and C++ wrapper yolo_v2_class.cpp and i get :

FalsePositive1.jpg where C++ wrapper detects a false-positif with a confidence of 58% Falsepositive2.jpg where C++ wrapper detects a false-positif with a confidence of 57% TruePositive1.jpg where C++ wrapper detects 2 true positif with 100% confidence

Of course, theses 3 methods detect the same number of objects at the same place (true positive and false positive). But it's seems confidence is very different for false positive. It's very strange.

I uploaded .cfg .data .names .weights and the 3 test images at : https://drive.google.com/open?id=1DyNmM0z8fCmT4TSUWRYgjgNNpWAYy9Ou

Thanks a lot

[ghost]

Somebody has an idea or could test my weights ?

Thanks

[Fetulhak]

@ghost did you get any solution back then?