Ziqi Zhao

Hong Kong University of Science and Technology (HKUST)

First class honor

Quantized Neural Networks for 6D Pose Estimation

Master's Thesis supervised by Dr. Mathieu Salzmann in Computer Vision Lab. Score: 5.75/6.

• Applied network quantization to 6D pose estimation models to improve efficiency.
• Analyzed the sensitivity of different parts of the network against quantization.
• Proposed a multi-stage quantization algorithm to improve 1%-5% accuracy for different models.

Pushing the Limits of Optical Character Recognition on Complex Multilingual Documents

Semester Project supervised by Prof. Frédéric Kaplan in Digital Humanities Lab. Score: 5.75/6.

• Finetuned Google Tesseract's OCR model for ancient Greek texts to improve overall recognition performance.
• Decreased the character error rate (CER) of Greek texts from 4.55% to 1.97% and that of the whole commentary from 8.07% to 3.28%.
• Developed an image pre-processing pipeline to further decrease the CER by 1%.

Network Pruning in Adversarial Training

Semester Project supervised by Prof. Sabine Süsstrunk in Image and Visual Representation Lab. Score: 5.75/6.

• Proposed an algorithm to find robust subnetworks from randomly-initialized binary networks based on Strong Lottery Ticket Hypothesis.
• Designed an adaptive pruning strategy to balance subnetwork search space and expression power.
• The binary subnetworks showed competitive performance among state-of-the-art full-precision ones and saved more than 30% FLOPs.
• The work was extended and published on Advances in Neural Information Processing Systems (NeurIPS) 2022. See the work

Floor Identification using Crowdsourced RF Signals

Research Project supervised by Prof. Gary Shueng Han CHAN.

• Proposed a novel algorithm to efficiently and accurately identify a user's current floor using received Wi-Fi signals and limited labels.
• Modeled Wi-Fi signals as a bipartite graph and applied network embedding to solve the missing value problem.
• Achieved nearly 100% of identification accuracy and still have over 90% of accuracy when 70% of Wi-Fi access points are missing.
• The work was published on the 42nd IEEE International Conference on Distributed Computing Systems (ICDCS) 2022. See the work

Wi-Fi Geo-Fencing

Research Project supervised by Prof. Gary Shueng Han CHAN.

• Designed a novel and robust algorithm for identifying whether people go out of a predefined geo-fence.
• Helped collect the data from various settings.
• Part of the project work was published on IEEE Internet of Things Magazine Special Issue on Smart IoT Solutions for Combating the COVID-19 Pandemic. See the work

Indoor Crowdsourced Wi-Fi Fingerprinting with Network Embedding

Bachelor's Thesis supervised by Prof. Gary Shueng Han CHAN.

• Designed a novel algorithm to build up an indoor Wi-Fi fingerprinting system with high deployment efficiency.
• Conducted extensive experiments to evaluate the performance of the system.

Personalized Running Speed Estimation using Foot Inertial Sensor

Semester Project supervised by Prof. Kamiar Aminian in Laboratory of Movement Analysis and Measurement. Score: 6/6

• Implemented a running speed personalization model with machine learning.
• Analyzed the major features that can contribute most to the model and implement several models for comparison.