PingAn-VCGroup's Solution for ICDAR 2021 Competition on Scientific Literature Parsing Task B: Table Recognition to HTML

The ICDAR 2021 competition on scientific literature parsing task B is to reconstruct the table image into an HTML code. In this competition, PubTabNet dataset (v2.0.0) [3] is provided as the official evaluation data, and Tree-Edit-Distance-based similarity (TEDS) metric is used for evaluation. The PubTabNet data set consists of 500,777 training samples, 9,115 validation samples, 9,138 samples for the development stage, and 9,064 samples for the final evaluation stage. For the training and validation data, the ground truth HTML codes and the position of non-empty table cells are provided to the participants. Participants of this competition need to develop a model that can convert images of tabular data into the corresponding HTML code, which should correctly represent the structure of the table and the content of each cell. The labels of samples for the development and the final evaluation stages are preserved by the organizers. We divide this task into four sub-tasks: table structure recognition, text line detection, text line recognition, and box assignment. And several tricks are tried to improve the model. The details of each sub-task will be discussed in the following section.