ICPR 2020 Competition on HArvesting Raw Tables (ICPR-2020-CHART-UB_PMC) (ICPR2020-CHART-Info)
This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License.
Locating and interpreting the axes of the chart is critical to transforming data point coordinates from units of pixels to the semantic units. Competing systems are expected to output the location and value of each tick mark on both the X-axis and Y-axis. Tick locations are represented as points and must be associated with the corresponding value (a string). Note that some sets of ticks are ordered or unordered discrete sets with textual non-numeric labels.
For this competition, X-axis will always refer to the axis that represents the independent variable shown, rather than the axis that is visually horizontal. For example, vertical bar and vertical box plots have an X-axis that is vertical. Similarly, the Y-axis is not always the axis that is vertical.
We use a modified F-measure to score each axis and then take the average F-measure over all axes. Each detected tick is scored for correctness, receiving a score between 0 and 1. Precision is then computed as the sum of the scores divided by the number of predictions. Recall is computed as the sum of the scores divided by the number of ground truth ticks.
A detected tick receives a score of 1 if the predicted point is close to the corresponding GT tick point, where correspondance between predictioned and GT ticks is based on the text BB and transcription. The threshold for close (scoring 1) and the threshold for far (scoring 0) is based on the distance between tick marks in the chart image. Predictions that are between the close and far thresholds are penalized linearly with distance.
No comments on this dataset yet.