Concept | Object detection#

Dataiku provides object detection modeling, a computer vision technology that can recognize, label, and draw bounding boxes around objects within images.

Just as with image classification models, object detection models are pre-trained on millions of images, and then can be fine-tuned on a particular set of images to focus on specific objects.

By drawing bounding boxes and classifying objects, the model can tell you:

  • Whether an object is present in an image.

  • If so, where it is located within an image.

  • Which class it belongs to.

Examples of image classification (labeling images as either lion or tiger) and object detection (identifying a tiger in an image).

Object detection has many potential applications and is widely used in satellite imagery, autonomous driving, agriculture, and manufacturing, among many other industries.

In the example below, a model was trained to find and label different types of computer microcontrollers, such as Arduinos and Raspberry Pis.

Note

You can build your own object detection model using the microcontrollers images in our tutorial, and you can view the final project in the Dataiku Gallery.

An example of object detection available in Dataiku tutorials and the gallery.

R-CNN models#

As covered, Concept | Pre-trained image classification models, Dataiku uses CNNs, or convolutional neural networks, to classify images.

Similarly, Dataiku also uses a CNN approach to object detection. However, this CNN is region-based, or an R-CNN, meaning it looks at a finite set of regions for each image, which greatly cuts down on computational complexity.

The provided model is a Faster R-CNN, an iteration of the R-CNN that is built to perform much more quickly.

Conceptual image of a region-based CNN.