Objects on images indexing

There is an enormous amount of visual information in the world. The most actual task is to find exactly what you need. Instead of Google image search, there are a lot of specific tasks with a limited number of object types for image tagging and indexing.

Let’s look at the computer vision technologies usage example for detecting wedding accessories and classifying them, providing information on different accessories characteristics provided in the table below.

Technology

The first step of this project is data labeling. The customer has provided a dataset of wedding photos without any labeling. And we used Amazon MTurk service to provide such.

Amazon pipeline consists оf :

1. HTML template, which workers used for drawing bounding boxes and selecting the label.
2. Python script, which checks the similarity of labels, approves or rejects the assignment.
3. Python script, which generates proper labeling for training the neural net.

The second step was to train a neural network for category detection. We tested several neural net architectures, such as MobileNet, NasNet, ResNet, and YOLO, to find the best accuracy/speed ratio. Finally, the YOLOv2 neural net was trained on the labeled dataset.

The third step was to train an ensemble of models to make a classification of detected categories by tags. We used PyTorch pre-trained vgg16 models to train it on a labeled dataset.