Author Image

Hi, I am Zekun

Zekun Li

Ph.D. Student at University of Minnesota, Twin Cities

I am a computer science Ph.D. student at the University of Minnesota (UMN), working with Professor Yao-Yi Chiang. My research interest is on the geospatial data analysis with computer vision and natural language processing techniques. I have worked on text detection of historical map labels, connecting separated text labels, linking recognized place names to existing knowledge bases (entity linking) and label type inference (entity typing). Previously at the University of Southern California (USC), I also worked on object detection projects in the Information Sciences Institute (ISI) and Spatial Sciences Institute (SSI).

pytorch
keras
docker
ubuntu
centos
raspberrypi

Experiences

1
Text Detection and Recognition for Historical Maps
Spatial Sciences Institute, USC

Dec 2019 - Aug 2020, Los Angeles

Responsibilities:
  • Built a deep neural network for detecting text of various font size, style and orientation angles on historical map patches. The network is able to handle text regions of arbitrary shapes
  • Designed the network to highlight probable text regions and then predict accurate bounding boxes given both map features and text probability distributions in a coarse-to-fine manner

Generating Historical Maps from online Maps
Spatial Sciences Institute, USC

Aug 2018 - Dec 2019, Los Angeles

Responsibilities:
  • Synthesized historical maps from Open Street Map tiles with conditional generative adversarial networks
  • The network generated background and foreground separately using different targets to solve the content mismatch problem in online maps and historical maps
  • Used the synthesized historical maps as the base-map and automatically place text labels on them to provide a large amount of training data for text detection networks
2

3
Synthetic Face Generation for Facial Landmark Detection
Amazon

May 2020 - Aug 2020, Seattle (Remote)

Responsibilities:
  • Built a pipeline to generate synthetic face images with landmark annotations using 3D modeling application Makehuman and rendering application Blender
  • Rendered the images from 3D models with various poses, camera setting, lighting conditions and backgrounds
  • Verified that the 2D landmark detection task and the 3D mesh prediction task can both benefit from the large amount of generated synthetic images

Automated Visual Data Extraction from Chart Images
Microsoft Research Asia

May 2019 - Aug 2019, Beijing, China

Responsibilities:
  • Built a pipeline to automatically infer numerical values for each chart given the column chart images
  • Applied trident-net to extract the chart object heights. Designed a ruler encoding module to interpret the y-axis information to convert the objects from pixel-space to ruler space to generate reading
  • The ruler encoding module focuses on the minimum and maximum values of the ruler to decide the numerical range that the charts represent
4

Publications

SpaBERT Pretrained Language Models on Geographic Data for Geo-Entity Representation

The paper proposes a novel spatial language model called SpaBERT that captures the spatial context of named geographic entities (geo-entities) in geospatial data. The model is based on the hypothesis that the characteristics of a geo-entity can be inferred by its surrounding entities, similar to word meanings in linguistic context.

ACE: Anchor-free Corner Evolution for Real-time Arbitrarily-oriented Object Detection

The paper proposes a novel model for detecting arbitrarily-oriented objects, such as texts/hands or objects in aerial images. The model evolves the axis-aligned bounding box to an oriented quadrilateral box using contour information.

ChartOCR: Data Extraction from Charts Images via a Deep Hybrid Framework.

The paper proposes an unified method, called ChartOCR, to extract data from various types of chart images, including bar charts, line charts, and pie charts. We combine deep learning and rule-based methods to achieve generalization ability and obtain accurate and semantic-rich intermediate results.

An Automatic Approach for Generating Rich, Linked Geo-Metadata from Historical Map Images

We present an end-to-end approach that automatically processes historical map images to extract their text content and generate a set of metadata linked to large geographic databases. The approach combines OCR with geocoding to accurately identify location phrases and assign geospatial coordinates.

Synthetic Map Generation to Provide Unlimited Training Data for Historical Map Text Detection

Many text detection algorithms have been proposed to locate text regions in map images automatically, but most of the algorithms are trained on out-of-domain datasets. This paper introduces a method to automatically generate an unlimited amount of annotated historical map images for training text detection models.

Combining Remote-Sensing-Derived Data and Historical Maps for Long-Term Back-Casting of Urban Extents.

The paper proposes a framework that uses remote sensing data (the Global Human Settlement Layer, GHSL) and georeferenced historical maps to generate historical urban extents for the early 20th century.

Weighted Feature Pooling Network in Template-Based Recognition

The paper proposes a template-based learning approach for computer vision tasks, where multiple instances of a concept are available. The method dynamically predicts weights that consider noise and redundancy to aggregate image-level features into a single template-level representation.

Selected Projects

mapKurator system
mapKurator system
Team Lead March 2020 - Present

A deep learning tool to process scanned historical maps. Performs text detection & recoginition, postOCR correction and entity linking.

SpaBERT
SpaBERT
First Author Jan 2021 - Nov 2022

SpaBERT extends BERT to capture linearized spatial context, while incorporating a spatial coordinate embedding mechanism to preserve spatial relations of entities in the 2-dimensional space.

Synthetic Map Generation
Synthetic Map Generation
First Author Jun 2019 - Aug 2020

Utilized cycle-GAN to convert open street map (OSM) images into historical map style.

Accomplishments

Ordnance Survey Award

Mishmash: A Mix of Old and New won the Ordnance Survey Award 2022. I adopted cycleGAN to generate synthetic historical maps from Open Street Map (OSM) vector data.

AI for Critical Mineral Assessment Competition
DARPA December 2022

The Map Feature Extraction Challenge required us to identify and label map features–lines, polygons, and points– that appear in the legend of historical maps. Our team isi-umn won the first place.

Recent Posts