Image segmentation dataset generated from dash cam videos, consisting of 700 images (and their associated annotations. The dataset comprises 32 labels.
Original dataset: images (557MB), annotations (16MB), label colors
Conversion from original data into dataset storing the label index in the blue channel of the RGB PNG image (1-32):
Download the images and annotations archives and extract them, resulting in the 701_StillsRaw_full and LabeledApproved_full directories.
Create bluechannel directory at the same level as extracted archives.
Run this Python script (requirements.txt) in the top-level directory to create images/annotations in the bluechannel directory: ./venv/bin/python bluechannel.py
NB: This conversion script skips image/annotation pair Seq05VD_f02610 as it contains colors other than the defined labels, which reduces the number of images from the original 701 to 700.
Citation
Brostow, Gabriel J., Jamie Shotton, Julien Fauqueur, and Roberto Cipolla. "Segmentation and recognition using structure from motion point clouds." In European conference on computer vision, pp. 44-57. Springer, Berlin, Heidelberg, 2008.
Gabriel J. Brostow, Julien Fauqueur, Roberto Cipolla, Semantic object classes in video: A high-definition ground truth database, Pattern Recognition Letters, Volume 30, Issue 2, 2009, Pages 88-97, ISSN 0167-8655, https://doi.org/10.1016/j.patrec.2008.04.005.
License
Unclear.