Fei Wu

• Since 09/2023:
  Postdoctoral researcher at the Pattern Recognition Lab, Department of Computer Science 5, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany.
• 09/2017 – 07/2023:
  Doctor of Philosophy, School of Electrical, Electronics and Communication Engineering, University of Chinese Academy of Sciences, Beijing, China.
• 09/2013 – 07/2017:
  Bachelor of Engineering, School of Computer Science and Engineering, Central South University, Changsha, China.

2023

• Research on handwriting analysis, object tracking and segmentation based on machine learning

  (Own Funds)

  Term: November 16, 2023 - December 1, 2024
  Abstract

  As an important issue in the first step of digitizing scanned documents, this project will focus on line segmentation and text recognition. Line segmentation can be regarded as instance segmentation or polygon detection. In this work, we will first assess the performance of our recently proposed model: AMD-HookNet and HookFormer. After making a comparison with the current state-of-the-art line segmentation methods, deeper research based on these two baseline models is required. Both architecture improvements and novel global-local interaction strategies will be investigated. Furthermore, the text recognition technique will be developed as a unified end-to-end segmentation-free approach for addressing the unnecessary two-phrase recognition problem.

  This project will investigate state-of-the-art algorithms for achieving accurate and stable object tracking and segmentation. Nowadays, Siamese trackers dominate the tracking field. The balanced fast inference speed and relatively high performance have caught the researchers’ attention. However, Siamese trackers mostly rely on large dataset offline training to learn the general representative capability for an arbitrary given target object. This ignores the target context relationship from adjacent frames. In addition, both CNNs and ViTs are used as feature extractors while the combination of local fine-grained and global coarse representation is still unexplored. We will implement CNNs- and ViTs-based improvements on a baseline tracker (TransT or a pure ViT) and then evaluate them on several well-known public datasets to validate their effectiveness.

  →More information

2024

• Competition Winner:
  The International Conference on Document Analysis and Recognition (ICDAR) Competition on Few-Shot and Many-Shot Layout Segmentation of Ancient Manuscripts (SAM)