Thanh Do

Key Skills

Software

Anno-Mage

CVAT

Datature

Labelbox

LabelImg

Label Studio

Roboflow

SuperAnnotate

Top Subject Matter

No subject matter listed

Top Data Types

Image

Text

Video

Top Task Types

Bounding Box

Classification

Object Detection

Point Key Point

Polygon

Freelancer Overview

I have extensive experience in developing and optimizing AI models that rely on high-quality labeled data, particularly in computer vision and deep learning. Throughout my projects — such as YOLOv8-based systems for PPE detection, fire detection, and public behavior analysis — I have managed the full data pipeline: collecting, cleaning, labeling, and validating large-scale image and video datasets to ensure precise training outcomes. My hands-on understanding of feature extraction, data preprocessing, and annotation consistency has been crucial to achieving accurate, real-time model performance. In addition, I have worked closely with AI training data for diverse applications including object detection, pose estimation, and OCR. My technical expertise with PyTorch, ONNX, and DeepStream, combined with practical experience in model deployment and dataset optimization, allows me to bridge the gap between data preparation and production-ready AI systems. This end-to-end experience ensures that I can design labeling workflows that maximize both data quality and training efficiency.

Labeling Experience

Traffic violation monitoring

LabelimgImageBounding BoxObject Detection

The project focuses on traffic monitoring and red-light violation detection using computer vision. The main objective is to develop a high-quality dataset for vehicle and traffic signal recognition under various real-world conditions such as daytime, nighttime, and different weather scenarios. The dataset covers multiple categories of road users and emergency vehicles to ensure a comprehensive understanding of traffic environments. The labeling tasks include identifying and annotating all visible objects related to traffic flow and control systems within each frame. The dataset includes tens of thousands of labeled images collected from traffic cameras at intersections and roads with diverse angles and lighting conditions.

The project focuses on traffic monitoring and red-light violation detection using computer vision. The main objective is to develop a high-quality dataset for vehicle and traffic signal recognition under various real-world conditions such as daytime, nighttime, and different weather scenarios. The dataset covers multiple categories of road users and emergency vehicles to ensure a comprehensive understanding of traffic environments. The labeling tasks include identifying and annotating all visible objects related to traffic flow and control systems within each frame. The dataset includes tens of thousands of labeled images collected from traffic cameras at intersections and roads with diverse angles and lighting conditions.

2025 - 2025

Detecting violations of labor protection

LabelimgImageBounding BoxObject Detection

The PPE (Personal Protective Equipment) project focused on labeling safety compliance equipment in industrial environments. The dataset consisted of six distinct classes: helmet, vest, shoe, no_helmet, no_vest, and no_shoe. The main objective was to accurately identify whether workers were properly equipped with the required safety gear or not. The data labeling tasks involved annotating images using bounding boxes to classify each detected object into one of the six PPE categories. Special attention was given to precise object boundaries and correct differentiation between positive and negative cases (e.g., helmet vs. no_helmet). Strict quality control procedures were applied, including multi-level review, cross-validation among labelers, and consistency checks using predefined annotation guidelines. The final dataset met high accuracy and consistency standards suitable for training deep learning detection models.

The PPE (Personal Protective Equipment) project focused on labeling safety compliance equipment in industrial environments. The dataset consisted of six distinct classes: helmet, vest, shoe, no_helmet, no_vest, and no_shoe. The main objective was to accurately identify whether workers were properly equipped with the required safety gear or not. The data labeling tasks involved annotating images using bounding boxes to classify each detected object into one of the six PPE categories. Special attention was given to precise object boundaries and correct differentiation between positive and negative cases (e.g., helmet vs. no_helmet). Strict quality control procedures were applied, including multi-level review, cross-validation among labelers, and consistency checks using predefined annotation guidelines. The final dataset met high accuracy and consistency standards suitable for training deep learning detection models.

2025 - 2025

Education

Work History