Point Cloud Saliency Maps Based on Non-Contribution Factors.

The special properties of the point cloud structure lead to difficulties in interpreting the features learned from Deep Neural Networks. In this paper, a method for obtaining point cloud saliency maps in point cloud target recognition models is proposed. Several free factors are randomly released in the target space and fed into the model. Then the pooled features output from the backbone of the model are made to deviate as much as possible from the pooled features output from the target point cloud during recognition process. The iterative factors contribute ”zero” to the prediction of the model, so that moving any point in the target to these factors has exactly the same effect on the recognition result as dropping the point. Since the movement of the point is differentiated, saliency maps can be obtained based on the gradient information. We generate saliency maps on PointNet, and the saliency maps are evaluated to be of good generality.