We provide the following datasets used to train DeepLab with weak bounding box annotations. The datasets can be downloaded from this link. See below for details and also please refer to our paper, Weakly- and Semi-Supervised Learning of a Deep Convolutional Network for Semantic Image Segmentation.

  • We generate bounding box segmentations from the pixel-level annotations by fitting tight bounding boxes.
  • densecrf has been employed to perform foreground/background segmentation for each bounding box annotation. We here provide the segmentation annotations after applying densecrf.

Note that please transform the annotations by removing the colormap (see the provided matlab script matlab/my_script/SavePngAsRawPng.m) so that OpenCV can read the annotations correctly.