Figure 2: GOD framework. The model takes an image as input and generates the feature map.RPN proposes the region based on the feature map. Detection network uses the proposed regions and feature map to detect various graphical objects

Table I: Performances of different backbone models on ICDAR-POD 2017 with IoU threshold 0.6. Deeper pre-trained model obtains higher detection accuracy. AP : average precision. Bold value indicates the best result.

Table II: Greater IoU threshold reduces the performance of the GOD (Faster R-CNN ) model on ICDAR-POD 2017 due to reduction in number of true positives.

Table III: Comparison with state-of-the-arts based on mAP and Ave F1 with IoU = 0.8 and 0.6, respectively on ICDAR-POD 2017 data set. Our GOD (Mask R-CNN ) is better than state-of-the-arts while IoU = 0.8 and IoU = 0.6. Bold value indicates the best result.

Table IV: Comparison with state-of-the-arts based on recall, precision and F1 with IoU = 0.5 on ICDAR-2013 data set. Bold value indicates the best result.

Table V: Comparison with state-of-the-arts on UNLV data set based on recall, precision and F1 with IoU = 0.5. Bold value indicates the best result.

Table VI: Performance of the trained GOD model on GO-IIIT-5K with IoU = 0.5. GOD: trained on public benchmark ICDAR-2013, ICDAR-POD 2017, UNLV and Marmot data sets and tested on test images of GO-IIIT-5 K data set. GOD†: trained on the public benchmark ICDAR -2013, ICDAR-POD 2017, UNLV and Marmot data sets, then fine-tuned on training images and tested on test images of GO-IIIT-5K data set. ‘*’ indicates that data set is randomly divided into training and test sets. ‘**’ indicates that data set is divided into training and test sets based on company. Bold value indicates the best result.

Figure 3: (a) Results of graphical objects: table, figure and equation localization using the GOD (Mask R-CNN ) on ICDAR-POD 2017 data set. Blue, Green and Red colors represent the predicted bounding boxes of table, figure and equation, respectively. (b) Results of table localization using the GOD (Mask R-CNN ) in the document images of ICDAR-2013 data set. Blue color represents the predicted bounding box of the table.

Figure 4: (a) Results of the table localization using GOD (Mask R-CNN ) in the document images of UNLV data set. (b) Results of table localization using the GOD (Mask R-CNN ) in the document images of GO-IIIT-5K data set. Blue color represents the predicted bounding box of the table.

Figure 5: Examples where the GOD fails to accurately localize the graphical objects. (a) figure is detected as table. (b) table is partially detected (here, object of interest is table). (c) some equations are not detected. (d) figure is detected as both figure and table. (e) single figure is detected as multiple figures. (f) paragraph is detected as table. Blue, Green and Red colors represent the predicted bounding boxes of table, figure and equation, respectively.