Segmentation and Tracking of Out of Distribution Objects

Описание: This video visualizes our OOD segmentation method as well as the tracking of the resulting segments. Top left, the original images from the Street Obstacle Sequences dataset (first two sequences) and from the CARLA WildLife Sequences dataset (https://zenodo.org/communities/buw-oo...) are shown with overlays of the ground truth semantic segmentation (here for classes road and OOD object) if available. The top right panel depicts the entropy heatmap obtained by the softmax output of semantic segmentation network. The corresponding semantic segmentation prediction of class road is given bottom left. The segmentation is obtained from a Nvidia DeepLabV3+ network (see http://github.com/sthalles/deeplab_v3). In addition, the predicted OOD objects placed on the road are obtained by entropy thresholding. In the bottom right panel, a visualization of our segment tracking approach is given applied to the resulting OOD objects (segments with low probability of being a true OOD object are removed). A given color represents a segment's ID over the course of image frames.

For further reading, please refer to https://arxiv.org/abs/2210.02074.