1. FCN 预训练模型入门

这是一个在 PASCAL VOC 数据集上使用 GluonCV FCN 模型的快速演示。如果您尚未安装 MXNet 和 GluonCV，请按照安装指南进行安装。

import mxnet as mx
from mxnet import image
from mxnet.gluon.data.vision import transforms
import gluoncv
# using cpu
ctx = mx.cpu(0)

准备图像

下载示例图像

url = 'https://raw.githubusercontent.com/dmlc/web-data/master/gluoncv/segmentation/voc_examples/1.jpg'
filename = 'example.jpg'
gluoncv.utils.download(url, filename)

输出

Downloading example.jpg from https://raw.githubusercontent.com/dmlc/web-data/master/gluoncv/segmentation/voc_examples/1.jpg...

  0%|          | 0/24 [00:00<?, ?KB/s]
100%|##########| 24/24 [00:00<00:00, 11662.99KB/s]

加载图像

img = image.imread(filename)

from matplotlib import pyplot as plt
plt.imshow(img.asnumpy())
plt.show()

使用数据集均值归一化图像

from gluoncv.data.transforms.presets.segmentation import test_transform
img = test_transform(img, ctx)

加载预训练模型并进行预测

获取预训练模型

model = gluoncv.model_zoo.get_model('fcn_resnet101_voc', pretrained=True)

输出

Downloading /root/.mxnet/models/resnet101_v1s-bd93a83c.zip from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/resnet101_v1s-bd93a83c.zip...

  0%|          | 0/100120 [00:00<?, ?KB/s]
  0%|          | 101/100120 [00:00<02:01, 822.78KB/s]
  1%|          | 508/100120 [00:00<00:43, 2313.59KB/s]
  2%|2         | 2174/100120 [00:00<00:13, 7494.88KB/s]
  7%|7         | 7485/100120 [00:00<00:03, 23397.08KB/s]
 14%|#3        | 13569/100120 [00:00<00:02, 35844.96KB/s]
 22%|##1       | 21878/100120 [00:00<00:01, 49694.96KB/s]
 31%|###       | 30614/100120 [00:00<00:01, 61341.20KB/s]
 38%|###8      | 38117/100120 [00:00<00:00, 65534.44KB/s]
 45%|####5     | 45383/100120 [00:00<00:00, 67702.96KB/s]
 53%|#####2    | 52645/100120 [00:01<00:00, 69192.76KB/s]
 60%|######    | 60464/100120 [00:01<00:00, 71242.72KB/s]
 69%|######8   | 68796/100120 [00:01<00:00, 74709.63KB/s]
 77%|#######6  | 76924/100120 [00:01<00:00, 76586.12KB/s]
 85%|########4 | 84930/100120 [00:01<00:00, 77623.11KB/s]
 93%|#########2| 93082/100120 [00:01<00:00, 78787.56KB/s]
100121KB [00:01, 60169.24KB/s]
Downloading /root/.mxnet/models/fcn_resnet101_voc-12c2b9b3.zip from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/fcn_resnet101_voc-12c2b9b3.zip...

  0%|          | 0/197332 [00:00<?, ?KB/s]
  0%|          | 713/197332 [00:00<00:32, 6019.02KB/s]
  2%|1         | 3620/197332 [00:00<00:11, 17077.96KB/s]
  6%|5         | 11124/197332 [00:00<00:04, 41267.10KB/s]
 10%|9         | 19268/197332 [00:00<00:03, 56266.43KB/s]
 14%|#3        | 27197/197332 [00:00<00:02, 64272.42KB/s]
 18%|#8        | 35948/197332 [00:00<00:02, 71345.17KB/s]
 23%|##2       | 44545/197332 [00:00<00:02, 76018.37KB/s]
 27%|##6       | 52534/197332 [00:00<00:01, 77228.92KB/s]
 31%|###1      | 61368/197332 [00:00<00:01, 80665.47KB/s]
 35%|###5      | 69482/197332 [00:01<00:02, 51085.40KB/s]
 39%|###8      | 76908/197332 [00:01<00:02, 56108.35KB/s]
 42%|####2     | 83673/197332 [00:01<00:02, 54654.53KB/s]
 47%|####6     | 92345/197332 [00:01<00:01, 62310.44KB/s]
 51%|#####     | 100241/197332 [00:01<00:01, 64432.00KB/s]
 55%|#####5    | 108838/197332 [00:01<00:01, 70037.42KB/s]
 59%|#####9    | 117065/197332 [00:01<00:01, 73351.23KB/s]
 64%|######3   | 125602/197332 [00:01<00:00, 76705.05KB/s]
 68%|######7   | 133555/197332 [00:02<00:00, 75112.36KB/s]
 72%|#######1  | 142021/197332 [00:02<00:00, 77812.37KB/s]
 76%|#######5  | 149960/197332 [00:02<00:00, 65174.54KB/s]
 80%|########  | 158552/197332 [00:02<00:00, 70465.03KB/s]
 85%|########4 | 167545/197332 [00:02<00:00, 75663.15KB/s]
 89%|########9 | 175755/197332 [00:02<00:00, 77439.02KB/s]
 93%|#########3| 184167/197332 [00:02<00:00, 79328.51KB/s]
 98%|#########7| 192819/197332 [00:02<00:00, 80870.78KB/s]
197333KB [00:02, 67221.14KB/s]

使用单尺度进行预测

output = model.predict(img)
predict = mx.nd.squeeze(mx.nd.argmax(output, 1)).asnumpy()

添加颜色调色板用于可视化

from gluoncv.utils.viz import get_color_pallete
import matplotlib.image as mpimg
mask = get_color_pallete(predict, 'pascal_voc')
mask.save('output.png')

显示预测掩码

mmask = mpimg.imread('output.png')
plt.imshow(mmask)
plt.show()

更多示例