Competition on HArvesting Raw Tables (CHART) 2019 - PubMedCentral (ICDAR-CHART-2019-PMC)

Research Tasks

Text Detection and Recognition

2019-06-18 (v. 1)

Contact author

Kenny Davila

University at Buffalo


This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License.


This sub-task concentrates on detecting and recognizing the text within the chart image. Competing systems are expected to produce tight bounding boxes and transcriptions for each text block. Examples of individual text blocks individual titles, tick labels, legend labels. Text blocks may be a single line, multiple lines (due to text wrapping), and may be horizontal, vertical, or rotated.


A predicted bounding box matches a GT bounding box if their Intersection Over Union (IOU) is at least 0.5, and tighter IOU criteria will be used to resolve ties when multiple predictions can match a single GT bounding box. There are two evalaution metrics for detection and recognition respectively. For detection, we sum the per-block IOU and divide by max(#predicted, #GT) for each image. For recognition, we average normalized Character Error Rate (CER) for each text block in an image. By normalized CER, we mean that the number of character edits to transform a predicted word to GT word is divided by the length of the GT block. False positive and false negative text block detections will be assigned a normalized CER of 1 and an IOU of 0. We will use the same procedure as the ICDAR Robust Reading Competitions to handle split/merged boxes.


No comments on this dataset yet.

Add your comment

In order to comment on a dataset you need to be logged on
Register Now!


In order to rate this dataset you need to be logged on
Register Now!