Video Instance Segmentation (CVPRW)
Conducted at CyberCore when I were a technical project manager
- Time: Apr 2021 – Jun 2021
- Role: Leader of a team with 4 members.
- Description: We participated the challenge as a company team. The problem is simultaneously detecting, segmenting, and tracking visual objects. Our solution is unifying the three tasks into a single model.
- Result: 1st rank in the challenge [Paper link] [Video link].
Transformer-based Visual Perception
Conducted at CyberCore when I were a technical project manager
- Time: Sep 2020 – Mar 2021
- Role: Leader of a team with 3 members.
- Description: Research on (Multi-head attention) Transformer-based methods for Object Detection and Multiple Object Tracking problems, benchmarked on Autonomous Driving datasets.
- Result: Delivered to customer.
Zalo AI challenge 2020
Conducted at CyberCore when I were a technical project manager
- Time: Nov 2020 – Dec 2020
- Role: Leader of a team with 4 members.
- Description: We participated the challenge as a company team. The problem is detecting traffic signs on the road, which is systematically a visual module in autonomous-driving vehicles.
- Result: Rank 3rd in the Traffic Sign Detection track.
Line detection and segmentation
Conducted at CyberCore when I were a technical project manager
- Time: Jun 2020 – Nov 2020
- Role: Leader of a team with 3 members.
- Description: Cooperate with Toda Construction to develop a portable device for verifying steel structure in construction. We design a fast and robust framework for processing Full-HD images on CPU with 2-10 FPS.
- Result: Delivered to customer.
3D Dangerous Object Detection using Milliwave Radar
Conducted at CyberCore when I were a technical project manager
- Time: Jun 2020 – now
- Role: Leader of a team with 6 members.
- Description: Cooperate with Taiyo Yuden to develop a security product for early alert at airports. It uses a network of various milliwave radars to detect dangerous objects (e.g., knife, gun) inside clothes.
- Result: The project is postponed due to the impact of COVID-19.
Waymo Open Dataset Challenge 2020
https://github.com/thuyngch/DSTNet
Conducted at CyberCore when I were a machine-learning engineer
- Time: Apr 2020 – May 2020
- Role: Leader of a team with 3 members.
- Description: We participated the challenge as a company team. The problem is detecting and tracking objects (e.g., car, pedestrian, and cyclist) on the road, which is systematically a visual module in autonomous-driving vehicles.
- Result: Rank 5th in 2D-Tracking and Rank 10th in 2D-Detection.
Out-of-distribution Object Detection
Conducted at CyberCore when I were a machine-learning engineer
- Time: Jan 2020 – Mar 2020
- Role: Major contributor of a team of 4 members, in which, I were responsible for reading, implementing, and improving SOTA papers.
- Description: We have researched techniques to make a network being able to detect unknown objects (not seen in training set and may be harmful if being detected wrongly). This project is a contract between CyberCore and Toyota Research Institute Advanced Development for building a module in autonomous cars.
- Result: The project was passed the PoC phase.
Object-Detection Network Compression
Conducted at Cyber Core when I were a machine-learning engineer
- Time: Jun 2019 – Dec 2019
- Role: Member of a team of 6 members, in which, I were responsible for reading and implementing SOTA papers.
- Description: We had researched techniques to compress a powerful network to a tiny one with 8% of the original GFLOPs without sacrificing the original accuracy. This project is a contract between CyberCore and Toyota Research Institute Advanced Development for building a module in autonomous cars.
- Result: The project was finalized and delivered to Toyota.
Vehicle Management System
Conducted personally
- Time: Aug 2019 – Dec 2019
- Role: Leader of a team with 4 members. I designed system architecture and was particularly responsible for AI core.
- Description: We had built a system to manage in/out information (vehicle image, licence plate, in-time, out-time, etc.) of vehicles in the Unilever factory.
- Result: The project was finalized and delivered to Unilever Vietnam. The system has been deployed live at Unilever factory in Cu Chi district, HoChiMinh city.
Image Forgery Classification and Segmentation: A Unified Deep-Learning Approach
Bachelor thesis
- Time: Jan 2019 – Jun 2019
- Role: Leader of a team with two members.
- Description: We proposed a unified deep-learning network which can perform classification and segmentation simultaneously. Besides, we also derived a loss function for overcoming data imbalance.
- Result: In comparison, there are two tasks. The proposed method surpasses recent methods (up to 2018) for 3/5 public datasets on task1, and 5/5 public datasets on task2.
Pornographic Activity Detection
Conducted at Zalo AILab when I were a data-mining collaborator
- Time: Mar 2018 – May 2019
- Role: Main contributor (e.g., collecting data, organizing data, training model, deploying model, evaluating model).
- Description: I had built up a pipeline, including image-classification network, text-detection network, OCR network, and text-classification model, to detect users who were carrying out pornographic activities on Zalo.
- Result: The pipeline has been deployed live in Zalo.
Human Segmentation
Conducted at Zalo AILab when I were a data-mining collaborator
- Time: Dec 2018 – Apr 2019
- Role: Share an equal contribution with a colleague.
- Description: We had explored and developed a realtime mobile app which can segment human body and replace background in livestream videos.
- Result: A base model is available (25 FPS on single core, 91% mIoU on our dataset).
Face Attendance Checking System
https://github.com/thuyngch/Face-Attendance-System
- Time: Sep 2018 – Nov 2018
- Role: Leader of a team with 6 members.
- Description: My team had designed an Attendance Checking application using face to distinguish individuals.
- Result: The algorithm can be run realtime on popular laptops (in CPU mode). It is also accurate at 96.5%.
Things Classifier
Conducted at Zalo AILab when I were a data-mining collaborator
- Time: Jul 2018 – Nov 2018
- Role: Major contributor (e.g., collecting data, organizing data, training model, evaluating model).
- Description: I had built up a module that can automatically classify images uploaded by users. It is part of a system to understand users’ interests.
- Result: The module has been deployed live in Zalo.
Image Forgery Detection using Deep Learning
https://github.com/thuyngch/Image-Forgery-using-Deep-Learning
- Time: Jun 2018 – Nov 2018
- Role: Leader of a team with 3 members.
- Description: This was a research contract with HCMUT. My team had researched on Deep Learning techniques applied for Image Forgery Detection problem.
- Result: Our model can detect forged images with high accuracy of 95.15%.
Event Information Extraction from Flyers
https://github.com/thuyngch/Event-info-extraction-from-flyers
- Time: Sep 2017 – Oct 2017
- Role: Algorithm development.
- Description: This was a project with 2 classmates in the course ”Digital Image Processing”. Our work was reading a paper from the Standford university and then realizing it into code.
- Result: A Matlab GUI extracting important information of events as text (e.g. time and location) from poster or flyer images.
Iris Recognition
https://github.com/thuyngch/Iris-Recognition
- Time: Jun 2017 – Dec 2017
- Role: Leader of a team with 6 members.
- Description: My team had conducted a research on Iris Recognition to recognize identities of people by using their iris images.
- Result: Two versions of Iris Recognition application in both Python and Matlab. An MIT-license repository is published in GitHub and received many interests from the community.