YoloV3 detection results are worse compared with pjreddie master

[PIXEVIA]

Detection results are worse compared with pjreddie master with all tested images using YoloV3 configuration and YoloV3 weights

test image: http://www.pixevia.com/media/img/items.jpeg

Main (pjreddie) branch results:

AlexeyAB branch results:

width=416
height=416

current darknet branch, command:
~/workspace/darknet3/darknet detect yolov3.cfg yolov3.weights ~/items.jpeg -thresh 0.05
bottle: 27%
book: 20%
book: 9%
book: 8%
chair: 99%
bottle: 99%
bottle: 66%
person: 100%

GPU=1
CUDNN=1
OPENCV=0
OPENMP=0
DEBUG=0

current AlexeyAB branch, command:
~/workspace/darkneta/darknet detect yolov3.cfg yolov3.weights ~/items.jpeg -thresh 0.05
bench: 12% (left_x: 57 top_y: 248 width: 708 height: 596)
chair: 95% (left_x: 71 top_y: 276 width: 535 height: 557)
person: 100% (left_x: 118 top_y: 58 width: 611 height: 747)
person: 10% (left_x: 344 top_y: 185 width: 343 height: 591)
bottle: 96% (left_x: 628 top_y: 96 width: 70 height: 182)
bottle: 87% (left_x: 703 top_y: 187 width: 45 height: 110)
bottle: 7% (left_x: 742 top_y: 190 width: 30 height: 102)
bottle: 54% (left_x: 754 top_y: 191 width: 45 height: 104)
bottle: 22% (left_x: 780 top_y: 199 width: 33 height: 101)

GPU=1
CUDNN=1
CUDNN_HALF=0
OPENCV=0
AVX=0
OPENMP=0
LIBSO=0

Cuda compilation tools, release 9.0, V9.0.176
/usr/local/cuda-9.0/targets/x86_64-linux/lib/libcudnn.so.7

[AlexeyAB]

The difference in resize approach: #232 (comment)
In some cases better the first approach, in other better the second.

In this repository https://github.com/AlexeyAB/darknet do simple change:

• comment this line:

  darknet/src/detector.c

  Line 1124 in 3e856ec

  image sized = resize_image(im, net.w, net.h);

• un-comment this line:

  darknet/src/detector.c

  Line 1125 in 3e856ec

  //image sized = letterbox_image(im, net.w, net.h); letterbox = 1;

darknet.exe detector test data/coco.data yolov3.cfg yolov3.weights items.jpeg -ext_output -thresh 0.05

chair: 99%      (left_x:   74   top_y:  314   width:  520   height:  534)
person: 100%    (left_x:  131   top_y:   39   width:  579   height:  773)
bottle: 99%     (left_x:  633   top_y:   84   width:   65   height:  205)
bottle: 59%     (left_x:  701   top_y:  184   width:   47   height:  111)
bottle: 26%     (left_x:  753   top_y:  181   width:   51   height:  116)
book: 20%
book: 7%        (left_x:  778   top_y:  187   width:   34   height:  111)
book: 6%        (left_x:  811   top_y:  187   width:   61   height:  113)

[PIXEVIA]

Thank you! It works perfectly now:)
With v2 we got very similar results, but with v3 and v3-spp the results with lower thresholds were worse. Most likely a neural network should be trained with the same resize function too, to achieve better performance.

[AlexeyAB]

@PIXEVIA Yes. Because for my datasets the resize() is better than letterbox(), then Training and Detection using resize() give me the best result.