6 DOF pose optimization

Description 

• Abstract

Part 1 Keypoint Localization:

The network takes as input an RGB image, and outputs a set of heatmaps with the intensity at each location of the heatmap indicating the prediction uncertainty of the respective keypoint. The network consists of two hourglass components, where each component can be further subdivided into two main processing stages. In the first step, we will train a heatmap-based neural network that given an image of an object, it estimates the location of the keypoints in the image.

Part 2 Pose Optimization:

Propose to fit a deformable shape model to the 2D detections while considering the uncertainty in keypoint predictions. However, in this part we need to generate another projected keypoints, since the plot scatter function multiple K there. For the second step, we use the coordinates of detected keypoints to estimate the 6DoF pose of the object.

​

​