GaDocNet: Rethinking the Anchoring Scheme and Loss Function in Vietnamese Document Images

In recent years, page object detection has received much attention from document image understanding. However, its application has many limitations in Vietnamese document images. In this paper, we address the page object detection problem in the Vietnamese document image. Specially, we experiment with four state-of-the-art object detection methods: Dynamic Faster R-CNN, Guided Anchoring Faster R-CNN, PointRend, and CascadeTabNet on the Vietnamese image document dataset named UIT-DODV. UIT-DODV dataset is the first Vietnamese document image dataset with four objects: Table, Figure, Caption, and Formula. In addition, we further evaluate the bounding box regression loss functions of the IoU family. Then we propose the EIoU loss function for efficiently page object detection in Vietnamese document images. Based on the preliminary experimental results, we present GaDocNet along with the EIoU loss function. The proposal achieves 76.1%, which is 1.6% higher than the baseline on the UIT-DODV dataset. Moreover, we evaluate with Deformable DETR, PAA, Reppoints, Foveabox, FSAF, and ATSS on UIT-DODV. The empirical evaluation points out the advantages of our approach, which is the foundation for further works.