search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.65k stars 7.96k forks source link

Grayscale label problem in darknet test prediction #6879

Open merterhk opened 3 years ago

merterhk commented 3 years ago

  1. screenshot - bad (yolo v4 custom, channel=1) resim

  2. screenshot - good (my another network result with same image - yolo v3 tiny 3l, channel=3) resim

D:\set-2>darknet.exe detector test data/urine.data.txt yolov4_custom_1024.cfg yolov4_custom_1024_best_89.weights -dont_show set-2-test\10357500106_190831_02_28_53_-_01.jpg   & rename predictions.jpg 10357500106_190831_02_28_53_-_01-sonuc.jpg
 CUDA-version: 10020 (10020), cuDNN: 7.6.4, GPU count: 1
 OpenCV version: 4.3.0
 compute_capability = 750, cudnn_half = 0
net.optimized_memory = 0
mini_batch = 1, batch = 1, time_steps = 1, train = 0
   layer   filters  size/strd(dil)      input                output
   0 conv     32       3 x 3/ 1   1024 x 768 x   1 -> 1024 x 768 x  32 0.453 BF
   1 conv     64       3 x 3/ 2   1024 x 768 x  32 ->  512 x 384 x  64 7.248 BF
   2 conv     64       1 x 1/ 1    512 x 384 x  64 ->  512 x 384 x  64 1.611 BF
   3 route  1                                      ->  512 x 384 x  64
   4 conv     64       1 x 1/ 1    512 x 384 x  64 ->  512 x 384 x  64 1.611 BF
   5 conv     32       1 x 1/ 1    512 x 384 x  64 ->  512 x 384 x  32 0.805 BF
   6 conv     64       3 x 3/ 1    512 x 384 x  32 ->  512 x 384 x  64 7.248 BF
   7 Shortcut Layer: 4,  wt = 0, wn = 0, outputs: 512 x 384 x  64 0.013 BF
   8 conv     64       1 x 1/ 1    512 x 384 x  64 ->  512 x 384 x  64 1.611 BF
   9 route  8 2                                    ->  512 x 384 x 128
  10 conv     64       1 x 1/ 1    512 x 384 x 128 ->  512 x 384 x  64 3.221 BF
  11 conv    128       3 x 3/ 2    512 x 384 x  64 ->  256 x 192 x 128 7.248 BF
  12 conv     64       1 x 1/ 1    256 x 192 x 128 ->  256 x 192 x  64 0.805 BF
  13 route  11                                     ->  256 x 192 x 128
  14 conv     64       1 x 1/ 1    256 x 192 x 128 ->  256 x 192 x  64 0.805 BF
  15 conv     64       1 x 1/ 1    256 x 192 x  64 ->  256 x 192 x  64 0.403 BF
  16 conv     64       3 x 3/ 1    256 x 192 x  64 ->  256 x 192 x  64 3.624 BF
  17 Shortcut Layer: 14,  wt = 0, wn = 0, outputs: 256 x 192 x  64 0.003 BF
  18 conv     64       1 x 1/ 1    256 x 192 x  64 ->  256 x 192 x  64 0.403 BF
  19 conv     64       3 x 3/ 1    256 x 192 x  64 ->  256 x 192 x  64 3.624 BF
  20 Shortcut Layer: 17,  wt = 0, wn = 0, outputs: 256 x 192 x  64 0.003 BF
  21 conv     64       1 x 1/ 1    256 x 192 x  64 ->  256 x 192 x  64 0.403 BF
  22 route  21 12                                  ->  256 x 192 x 128
  23 conv    128       1 x 1/ 1    256 x 192 x 128 ->  256 x 192 x 128 1.611 BF
  24 conv    256       3 x 3/ 2    256 x 192 x 128 ->  128 x  96 x 256 7.248 BF
  25 conv    128       1 x 1/ 1    128 x  96 x 256 ->  128 x  96 x 128 0.805 BF
  26 route  24                                     ->  128 x  96 x 256
  27 conv    128       1 x 1/ 1    128 x  96 x 256 ->  128 x  96 x 128 0.805 BF
  28 conv    128       1 x 1/ 1    128 x  96 x 128 ->  128 x  96 x 128 0.403 BF
  29 conv    128       3 x 3/ 1    128 x  96 x 128 ->  128 x  96 x 128 3.624 BF
  30 Shortcut Layer: 27,  wt = 0, wn = 0, outputs: 128 x  96 x 128 0.002 BF
  31 conv    128       1 x 1/ 1    128 x  96 x 128 ->  128 x  96 x 128 0.403 BF
  32 conv    128       3 x 3/ 1    128 x  96 x 128 ->  128 x  96 x 128 3.624 BF
  33 Shortcut Layer: 30,  wt = 0, wn = 0, outputs: 128 x  96 x 128 0.002 BF
  34 conv    128       1 x 1/ 1    128 x  96 x 128 ->  128 x  96 x 128 0.403 BF
  35 conv    128       3 x 3/ 1    128 x  96 x 128 ->  128 x  96 x 128 3.624 BF
  36 Shortcut Layer: 33,  wt = 0, wn = 0, outputs: 128 x  96 x 128 0.002 BF
  37 conv    128       1 x 1/ 1    128 x  96 x 128 ->  128 x  96 x 128 0.403 BF
  38 conv    128       3 x 3/ 1    128 x  96 x 128 ->  128 x  96 x 128 3.624 BF
  39 Shortcut Layer: 36,  wt = 0, wn = 0, outputs: 128 x  96 x 128 0.002 BF
  40 conv    128       1 x 1/ 1    128 x  96 x 128 ->  128 x  96 x 128 0.403 BF
  41 conv    128       3 x 3/ 1    128 x  96 x 128 ->  128 x  96 x 128 3.624 BF
  42 Shortcut Layer: 39,  wt = 0, wn = 0, outputs: 128 x  96 x 128 0.002 BF
  43 conv    128       1 x 1/ 1    128 x  96 x 128 ->  128 x  96 x 128 0.403 BF
  44 conv    128       3 x 3/ 1    128 x  96 x 128 ->  128 x  96 x 128 3.624 BF
  45 Shortcut Layer: 42,  wt = 0, wn = 0, outputs: 128 x  96 x 128 0.002 BF
  46 conv    128       1 x 1/ 1    128 x  96 x 128 ->  128 x  96 x 128 0.403 BF
  47 conv    128       3 x 3/ 1    128 x  96 x 128 ->  128 x  96 x 128 3.624 BF
  48 Shortcut Layer: 45,  wt = 0, wn = 0, outputs: 128 x  96 x 128 0.002 BF
  49 conv    128       1 x 1/ 1    128 x  96 x 128 ->  128 x  96 x 128 0.403 BF
  50 conv    128       3 x 3/ 1    128 x  96 x 128 ->  128 x  96 x 128 3.624 BF
  51 Shortcut Layer: 48,  wt = 0, wn = 0, outputs: 128 x  96 x 128 0.002 BF
  52 conv    128       1 x 1/ 1    128 x  96 x 128 ->  128 x  96 x 128 0.403 BF
  53 route  52 25                                  ->  128 x  96 x 256
  54 conv    256       1 x 1/ 1    128 x  96 x 256 ->  128 x  96 x 256 1.611 BF
  55 conv    512       3 x 3/ 2    128 x  96 x 256 ->   64 x  48 x 512 7.248 BF
  56 conv    256       1 x 1/ 1     64 x  48 x 512 ->   64 x  48 x 256 0.805 BF
  57 route  55                                     ->   64 x  48 x 512
  58 conv    256       1 x 1/ 1     64 x  48 x 512 ->   64 x  48 x 256 0.805 BF
  59 conv    256       1 x 1/ 1     64 x  48 x 256 ->   64 x  48 x 256 0.403 BF
  60 conv    256       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 256 3.624 BF
  61 Shortcut Layer: 58,  wt = 0, wn = 0, outputs:  64 x  48 x 256 0.001 BF
  62 conv    256       1 x 1/ 1     64 x  48 x 256 ->   64 x  48 x 256 0.403 BF
  63 conv    256       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 256 3.624 BF
  64 Shortcut Layer: 61,  wt = 0, wn = 0, outputs:  64 x  48 x 256 0.001 BF
  65 conv    256       1 x 1/ 1     64 x  48 x 256 ->   64 x  48 x 256 0.403 BF
  66 conv    256       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 256 3.624 BF
  67 Shortcut Layer: 64,  wt = 0, wn = 0, outputs:  64 x  48 x 256 0.001 BF
  68 conv    256       1 x 1/ 1     64 x  48 x 256 ->   64 x  48 x 256 0.403 BF
  69 conv    256       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 256 3.624 BF
  70 Shortcut Layer: 67,  wt = 0, wn = 0, outputs:  64 x  48 x 256 0.001 BF
  71 conv    256       1 x 1/ 1     64 x  48 x 256 ->   64 x  48 x 256 0.403 BF
  72 conv    256       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 256 3.624 BF
  73 Shortcut Layer: 70,  wt = 0, wn = 0, outputs:  64 x  48 x 256 0.001 BF
  74 conv    256       1 x 1/ 1     64 x  48 x 256 ->   64 x  48 x 256 0.403 BF
  75 conv    256       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 256 3.624 BF
  76 Shortcut Layer: 73,  wt = 0, wn = 0, outputs:  64 x  48 x 256 0.001 BF
  77 conv    256       1 x 1/ 1     64 x  48 x 256 ->   64 x  48 x 256 0.403 BF
  78 conv    256       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 256 3.624 BF
  79 Shortcut Layer: 76,  wt = 0, wn = 0, outputs:  64 x  48 x 256 0.001 BF
  80 conv    256       1 x 1/ 1     64 x  48 x 256 ->   64 x  48 x 256 0.403 BF
  81 conv    256       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 256 3.624 BF
  82 Shortcut Layer: 79,  wt = 0, wn = 0, outputs:  64 x  48 x 256 0.001 BF
  83 conv    256       1 x 1/ 1     64 x  48 x 256 ->   64 x  48 x 256 0.403 BF
  84 route  83 56                                  ->   64 x  48 x 512
  85 conv    512       1 x 1/ 1     64 x  48 x 512 ->   64 x  48 x 512 1.611 BF
  86 conv   1024       3 x 3/ 2     64 x  48 x 512 ->   32 x  24 x1024 7.248 BF
  87 conv    512       1 x 1/ 1     32 x  24 x1024 ->   32 x  24 x 512 0.805 BF
  88 route  86                                     ->   32 x  24 x1024
  89 conv    512       1 x 1/ 1     32 x  24 x1024 ->   32 x  24 x 512 0.805 BF
  90 conv    512       1 x 1/ 1     32 x  24 x 512 ->   32 x  24 x 512 0.403 BF
  91 conv    512       3 x 3/ 1     32 x  24 x 512 ->   32 x  24 x 512 3.624 BF
  92 Shortcut Layer: 89,  wt = 0, wn = 0, outputs:  32 x  24 x 512 0.000 BF
  93 conv    512       1 x 1/ 1     32 x  24 x 512 ->   32 x  24 x 512 0.403 BF
  94 conv    512       3 x 3/ 1     32 x  24 x 512 ->   32 x  24 x 512 3.624 BF
  95 Shortcut Layer: 92,  wt = 0, wn = 0, outputs:  32 x  24 x 512 0.000 BF
  96 conv    512       1 x 1/ 1     32 x  24 x 512 ->   32 x  24 x 512 0.403 BF
  97 conv    512       3 x 3/ 1     32 x  24 x 512 ->   32 x  24 x 512 3.624 BF
  98 Shortcut Layer: 95,  wt = 0, wn = 0, outputs:  32 x  24 x 512 0.000 BF
  99 conv    512       1 x 1/ 1     32 x  24 x 512 ->   32 x  24 x 512 0.403 BF
 100 conv    512       3 x 3/ 1     32 x  24 x 512 ->   32 x  24 x 512 3.624 BF
 101 Shortcut Layer: 98,  wt = 0, wn = 0, outputs:  32 x  24 x 512 0.000 BF
 102 conv    512       1 x 1/ 1     32 x  24 x 512 ->   32 x  24 x 512 0.403 BF
 103 route  102 87                                 ->   32 x  24 x1024
 104 conv   1024       1 x 1/ 1     32 x  24 x1024 ->   32 x  24 x1024 1.611 BF
 105 conv    512       1 x 1/ 1     32 x  24 x1024 ->   32 x  24 x 512 0.805 BF
 106 conv   1024       3 x 3/ 1     32 x  24 x 512 ->   32 x  24 x1024 7.248 BF
 107 conv    512       1 x 1/ 1     32 x  24 x1024 ->   32 x  24 x 512 0.805 BF
 108 max                5x 5/ 1     32 x  24 x 512 ->   32 x  24 x 512 0.010 BF
 109 route  107                                            ->   32 x  24 x 512
 110 max                9x 9/ 1     32 x  24 x 512 ->   32 x  24 x 512 0.032 BF
 111 route  107                                            ->   32 x  24 x 512
 112 max               13x13/ 1     32 x  24 x 512 ->   32 x  24 x 512 0.066 BF
 113 route  112 110 108 107                        ->   32 x  24 x2048
 114 conv    512       1 x 1/ 1     32 x  24 x2048 ->   32 x  24 x 512 1.611 BF
 115 conv   1024       3 x 3/ 1     32 x  24 x 512 ->   32 x  24 x1024 7.248 BF
 116 conv    512       1 x 1/ 1     32 x  24 x1024 ->   32 x  24 x 512 0.805 BF
 117 conv    256       1 x 1/ 1     32 x  24 x 512 ->   32 x  24 x 256 0.201 BF
 118 upsample                 2x    32 x  24 x 256 ->   64 x  48 x 256
 119 route  85                                     ->   64 x  48 x 512
 120 conv    256       1 x 1/ 1     64 x  48 x 512 ->   64 x  48 x 256 0.805 BF
 121 route  120 118                                ->   64 x  48 x 512
 122 conv    256       1 x 1/ 1     64 x  48 x 512 ->   64 x  48 x 256 0.805 BF
 123 conv    512       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 512 7.248 BF
 124 conv    256       1 x 1/ 1     64 x  48 x 512 ->   64 x  48 x 256 0.805 BF
 125 conv    512       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 512 7.248 BF
 126 conv    256       1 x 1/ 1     64 x  48 x 512 ->   64 x  48 x 256 0.805 BF
 127 conv    128       1 x 1/ 1     64 x  48 x 256 ->   64 x  48 x 128 0.201 BF
 128 upsample                 2x    64 x  48 x 128 ->  128 x  96 x 128
 129 route  54                                     ->  128 x  96 x 256
 130 conv    128       1 x 1/ 1    128 x  96 x 256 ->  128 x  96 x 128 0.805 BF
 131 route  130 128                                ->  128 x  96 x 256
 132 conv    128       1 x 1/ 1    128 x  96 x 256 ->  128 x  96 x 128 0.805 BF
 133 conv    256       3 x 3/ 1    128 x  96 x 128 ->  128 x  96 x 256 7.248 BF
 134 conv    128       1 x 1/ 1    128 x  96 x 256 ->  128 x  96 x 128 0.805 BF
 135 conv    256       3 x 3/ 1    128 x  96 x 128 ->  128 x  96 x 256 7.248 BF
 136 conv    128       1 x 1/ 1    128 x  96 x 256 ->  128 x  96 x 128 0.805 BF
 137 conv    256       3 x 3/ 1    128 x  96 x 128 ->  128 x  96 x 256 7.248 BF
 138 conv     33       1 x 1/ 1    128 x  96 x 256 ->  128 x  96 x  33 0.208 BF
 139 yolo
[yolo] params: iou loss: ciou (4), iou_norm: 0.07, cls_norm: 1.00, scale_x_y: 1.20
nms_kind: greedynms (1), beta = 0.600000
 140 route  136                                            ->  128 x  96 x 128
 141 conv    256       3 x 3/ 2    128 x  96 x 128 ->   64 x  48 x 256 1.812 BF
 142 route  141 126                                ->   64 x  48 x 512
 143 conv    256       1 x 1/ 1     64 x  48 x 512 ->   64 x  48 x 256 0.805 BF
 144 conv    512       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 512 7.248 BF
 145 conv    256       1 x 1/ 1     64 x  48 x 512 ->   64 x  48 x 256 0.805 BF
 146 conv    512       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 512 7.248 BF
 147 conv    256       1 x 1/ 1     64 x  48 x 512 ->   64 x  48 x 256 0.805 BF
 148 conv    512       3 x 3/ 1     64 x  48 x 256 ->   64 x  48 x 512 7.248 BF
 149 conv     33       1 x 1/ 1     64 x  48 x 512 ->   64 x  48 x  33 0.104 BF
 150 yolo
[yolo] params: iou loss: ciou (4), iou_norm: 0.07, cls_norm: 1.00, scale_x_y: 1.10
nms_kind: greedynms (1), beta = 0.600000
 151 route  147                                            ->   64 x  48 x 256
 152 conv    512       3 x 3/ 2     64 x  48 x 256 ->   32 x  24 x 512 1.812 BF
 153 route  152 116                                ->   32 x  24 x1024
 154 conv    512       1 x 1/ 1     32 x  24 x1024 ->   32 x  24 x 512 0.805 BF
 155 conv   1024       3 x 3/ 1     32 x  24 x 512 ->   32 x  24 x1024 7.248 BF
 156 conv    512       1 x 1/ 1     32 x  24 x1024 ->   32 x  24 x 512 0.805 BF
 157 conv   1024       3 x 3/ 1     32 x  24 x 512 ->   32 x  24 x1024 7.248 BF
 158 conv    512       1 x 1/ 1     32 x  24 x1024 ->   32 x  24 x 512 0.805 BF
 159 conv   1024       3 x 3/ 1     32 x  24 x 512 ->   32 x  24 x1024 7.248 BF
 160 conv     33       1 x 1/ 1     32 x  24 x1024 ->   32 x  24 x  33 0.052 BF
 161 yolo
[yolo] params: iou loss: ciou (4), iou_norm: 0.07, cls_norm: 1.00, scale_x_y: 1.05
nms_kind: greedynms (1), beta = 0.600000
Total BFLOPS 269.936
avg_outputs = 2228726
 Allocate additional workspace_size = 52.43 MB
Loading weights from yolov4_custom_1024_best_89.weights...
 seen 64, trained: 149 K-images (2 Kilo-batches_64)
Done! Loaded 162 layers from weights-file
set-2-test\10357500106_190831_02_28_53_-_01.jpg: Predicted in 152.255000 milli-seconds.
oksalat: 88%
epitel: 98%
oksalat: 96%

As seen on screenshots labels on predictions.jpg are grayscale and "oksalat" named label is full black. My other (channel=3) networks work well with same images and no problem with labels and colors.

I have 2 question: 1- Is grayscale label caused by 1 channel train? 2- Why my "oksalat" label seen in black? Maybe it has black background? How can i colorise my labels as others?

sunzhoujun commented 3 years ago

如果你的图片本来是灰度的，训练时channel=1就行，部署的时候读取图片照样以灰度图读取然后检测，检测完成之后如果直接将结果画在图上然后保存就会出现你第一张图片上的情况，但是如果检测完成以后将灰度图转换成为RGB图片保存，这样画出的框就是彩色的了。