Keypoint Detection

This repository contains examples and best practice guidelines for building keypoint detection systems. It also shows how to use a pre-trained model for human pose estimation.

Keypoints are defined as points-of-interests on objects. For example, one might be interested in finding the position of the lid on a bottle. Another example is to find body joints (hands, shoulders, etc.) for human pose estimation.

We use an extension of Mask R-CNN which simultaneously detects objects and their keypoints. The underlying technology is very similar to our approach for object detection, ie. based on Torchvision’s Mask R-CNN. The example notebook for keypoint localization is therefore in the detection folder.

Detecting the top and bottom on our fridge objects	Detecting various keypoints on milk bottles	Human pose estimation using the provided pre-trained model

Frequently asked questions

See the FAQ.md in the object detection folder.

Notebooks

Notebook name	Description
03_keypoint_rcnn.ipynb	Notebook which shows how to (i) run a pre-trained model for human pose estimation; and (ii) train a custom keypoint detection model.

Contribution guidelines

See the contribution guidelines in the root folder.