02. Predict with pre-trained Faster RCNN models

This article shows how to play with pre-trained Faster RCNN model.

First let’s import some necessary libraries:

from matplotlib import pyplot as plt
import gluoncv
from gluoncv import model_zoo, data, utils

Load a pretrained model

Let’s get an Faster RCNN model trained on Pascal VOC dataset with ResNet-50 backbone. By specifying pretrained=True, it will automatically download the model from the model zoo if necessary. For more pretrained models, please refer to Model Zoo.

The returned model is a HybridBlock gluoncv.model_zoo.FasterRCNN with a default context of cpu(0).

net = model_zoo.get_model('faster_rcnn_resnet50_v1b_voc', pretrained=True)

Out:

Downloading /root/.mxnet/models/faster_rcnn_resnet50_v1b_voc-447328d8.zip from https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/models/faster_rcnn_resnet50_v1b_voc-447328d8.zip...

  0%|          | 0/121887 [00:00<?, ?KB/s]
  0%|          | 93/121887 [00:00<03:03, 663.30KB/s]
  0%|          | 508/121887 [00:00<01:00, 2005.37KB/s]
  2%|1         | 2188/121887 [00:00<00:18, 6525.08KB/s]
  6%|5         | 7014/121887 [00:00<00:05, 19720.29KB/s]
 11%|#         | 13360/121887 [00:00<00:03, 32056.23KB/s]
 17%|#7        | 21247/121887 [00:00<00:02, 46074.86KB/s]
 23%|##3       | 28540/121887 [00:00<00:01, 52338.12KB/s]
 30%|##9       | 36313/121887 [00:00<00:01, 59769.65KB/s]
 36%|###5      | 43696/121887 [00:01<00:01, 61654.87KB/s]
 42%|####2     | 51593/121887 [00:01<00:01, 66655.90KB/s]
 48%|####8     | 58912/121887 [00:01<00:00, 66305.85KB/s]
 55%|#####4    | 66693/121887 [00:01<00:00, 69609.64KB/s]
 61%|######    | 73888/121887 [00:01<00:00, 67838.95KB/s]
 67%|######7   | 81756/121887 [00:01<00:00, 70943.11KB/s]
 73%|#######2  | 88977/121887 [00:01<00:00, 68911.94KB/s]
 79%|#######9  | 96874/121887 [00:01<00:00, 71787.73KB/s]
 85%|########5 | 104106/121887 [00:01<00:00, 68564.59KB/s]
 92%|#########1| 112007/121887 [00:02<00:00, 71520.70KB/s]
 98%|#########7| 119219/121887 [00:02<00:00, 69455.14KB/s]
121888KB [00:02, 56629.65KB/s]

Pre-process an image

Next we download an image, and pre-process with preset data transforms. The default behavior is to resize the short edge of the image to 600px. But you can feed an arbitrarily sized image.

You can provide a list of image file names, such as [im_fname1, im_fname2, ...] to gluoncv.data.transforms.presets.rcnn.load_test() if you want to load multiple image together.

This function returns two results. The first is a NDArray with shape (batch_size, RGB_channels, height, width). It can be fed into the model directly. The second one contains the images in numpy format to easy to be plotted. Since we only loaded a single image, the first dimension of x is 1.

Please beware that orig_img is resized to short edge 600px.

im_fname = utils.download('https://github.com/dmlc/web-data/blob/master/' +
                          'gluoncv/detection/biking.jpg?raw=true',
                          path='biking.jpg')
x, orig_img = data.transforms.presets.rcnn.load_test(im_fname)

Out:

Downloading biking.jpg from https://github.com/dmlc/web-data/blob/master/gluoncv/detection/biking.jpg?raw=true...

  0%|          | 0/244 [00:00<?, ?KB/s]
100%|##########| 244/244 [00:00<00:00, 24081.94KB/s]

Inference and display

The Faster RCNN model returns predicted class IDs, confidence scores, bounding boxes coordinates. Their shape are (batch_size, num_bboxes, 1), (batch_size, num_bboxes, 1) and (batch_size, num_bboxes, 4), respectively.

We can use gluoncv.utils.viz.plot_bbox() to visualize the results. We slice the results for the first image and feed them into plot_bbox:

box_ids, scores, bboxes = net(x)
ax = utils.viz.plot_bbox(orig_img, bboxes[0], scores[0], box_ids[0], class_names=net.classes)

plt.show()