Sayeem26s/chart-element-detector

Chart Element Detector

A chart element detection model based on CACHED (Context-Aware Chart Element Detection).

Model Details

• Architecture: Cascade R-CNN + Swin Transformer + FPN
• Task: Chart element detection and localization
• Classes: 18 chart element classes
• Dataset: PMC Chart Dataset
• COCO AP: 0.729

Classes

x_tick_label, y_tick_label, x_tick, y_tick, x_axis_title, y_axis_title, chart_title, legend_marker, legend_label, legend_title, value_label, mark_label, tick_grouping, plot_area, x_axis_area, y_axis_area, legend_area, others

Output Format

{
  "chart": [
    {
      "x1": 10.0,
      "y1": 20.0,
      "x2": 100.0,
      "y2": 200.0,
      "score": 0.95,
      "class": "chart_title"
    }
  ]
}

Requirements

torch==1.13.1
mmdet==2.28.2
mmcv-full==1.7.0

Citation

@inproceedings{yan2023cached,
  title={CACHED: Context-Aware Chart Element Detection},
  author={Yan, Pengyu and Ahmed, Saleem and Doermann, David},
  booktitle={ICDAR},
  year={2023}
}