Relabeled EPFL RGB-D Pedestrian Dataset

The EPFL RGB-D Pedestrian dataset consists of over 5000 RGB + Depth images acquired from an RGB camera and Kinect V2 sensor setup. However, the annotations for this dataset do not include persons that are highly occluded. This is why we decided to relabel the entire dataset manually, providing bounding boxes for every person in the image - independent of their occlusion level - and instead adding an occlusion percentage value. This relabeled dataset was used to assess the added value of fusing RGB + depth in a single-pass detection network.

Publication year:2018

Accessibility:open

Publisher:EAVISE

License:MIT

Format:YAML

Keywords: RGB-D, dataset, object detection, pedestrian