Duc Vu Email & Phone Number

Vietnam

• Synthesized 30,000 additional augmented infrared facial images using optical flow, generative models and 3D Face Rotation, enabling facial landmarks detection of drivers with 15% increased accuracy for driver monitoring system.• Implemented and trained a state-of-the-art facial landmarks detection model, decreasing overall RMSE by 10% and eyes’ RMSE by 20% compared to baseline, resulting in 12% improved driver drowsiness detection accuracy.• Published a public face dataset exclusively focusing on extreme head pose angles containing 450,000 frames, resulting in a 10-20% decrease in error rates for current state-of-the-art face generation and reenactment techniques.• Introduced a fused one-step diffusion model by combining two efficient training strategies and a novel CLIP loss, resulting in a state-of-the-art Fréchet Inception Score of 8.14.

Cầu Giấy, Hanoi, Vietnam

• Developed a data cleaning pipeline to crawl and categorized 30 million images in company’s database based on human attributes such as age, emotion, and location.• Provided a pipeline for data cleansing to cope with noisy and imbalanced data and formulated a multi-task model for age and emotion classification using 3.5 million images from database.• Deployed and evaluated a multi-task model that helped enhance customer experience, cleaned around 4 million unlabeled images (5% of total images), and re-categorized age and emotion groups of 13.5 million stock images (15% of total images).