---------------------------------------------------------------- Block-Based Ground Truth Dataset for Scene Text Detection scene-osakafu-u-GT4 (Rev.20090417) ---------------------------------------------------------------- 1. About This Dataset This dataset contains 4-class Ground Truth data for the natural scene images with text provided at: http://www.cs.osakafu-u.ac.jp/document/ This dataset is intended to be used for evaluations of block-based text detection algorithms. Please refer to the following paper for details. [HG2008IJDAR] Hideaki Goto, "Redefining the DCT-based feature for scene text detection -- Analysis and comparison of spatial frequency-based features," IJDAR, Vol.11, No.1, pp.1-8 (2008). The original package of this dataset can be found on our website: http://www.imglab.org/db/ ---------------------------------------------------------------- Copyright (C) 2005-2009 Hideaki Goto All Rights Reserved. You may use, copy, modify, merge, and distribute this dataset without restriction and free of charge, subject to the following conditions. * The above copyright notice and this permission notice shall be included in all copies or substantial portions of the dataset. * Modification(s) made to the dataset and the reason(s) of the modification(s) must be clearly explained in a document, and the document shall be included in all copies or substantial portions of the dataset. THE DATASET IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND. THE AUTHOR(S) OR COPYRIGHT HOLDER(S) WILL NOT BE RESPONSIBLE FOR ANY DAMAGE CAUSED BY THIS DATASET. ---------------------------------------------------------------- 2. Description The dataset is based on the Ground Truth data used in our research group (http://www.sc.isc.tohoku.ac.jp/). The original files of scene images can be found at: http://www.cs.osakafu-u.ac.jp/document/ (C) 2005 Koichi Kise The basic block size is 16x16 (pixels). Each block was manually classified into one of the following four classes. value class 0 text block 127 intermediate block 200 large text block 255 non-text block The criteria used are as follows. a) If a block does not contain any character stroke at all, the block is classified as "non-text block." b) If a character is smaller than 8x8 pixels, we cannot expect typical OCRs to recognize such a small character accurately. Therefore, if a block contains such a character only, the block is classified as "intermediate blocks." c) If a character is larger than double of the block size, its strokes often do not fit in a block. Therefore, if a block contains such a large character only, the block is classified as "large text blocks." d) If a block contains only a small portion of a stroke, the block is classified as "intermediate block." e) The other blocks are classified into "text blocks." 3. Changes Rev.20090417 Document update. Rev.20060316 004.pgm has been fixed because it had some non-255 white blocks. Rev.20051025 First release. 4. Credits This Ground Truth dataset was created by the following people. Tohoku University, Japan Seiji Saito Tomoyuki Saoi Hiroki Shiratori Hideaki Goto Contact: Assoc. Prof. Hideaki Goto Cyberscience Center, Tohoku University, Sendai 980-8578, JAPAN E-mail: hgot_(at)_isc.tohoku.ac.jp (remove underscores) WWW : http://www.sc.isc.tohoku.ac.jp/~hgot/