Dung Vo Pham Tuan

• Pretrained a multilingual vision-language backbone (Vietnamese/English/Chinese) for Text-based Person Re-identification on a large-scale dataset (36 million image-text pairs) using 4 NVIDIA A100 GPUs.
• Fine-tuned a model for Text-based Person Re-identification and collaborated with the Deployment Team for serving it.
• Kept up with advancements in Large Language Models (LLMs), Multimodal Large Language Models (MLLMs) and inference frameworks to accelerate data augmentation and increase the diversity of the pretraining dataset.
• Extended the original English-only pretraining dataset by adding Chinese and Vietnamese captioning annotations. Conducted experiments to demonstrate that multilingual pretraining improves zero-shot retrieval performance of the backbone in each language by over 1.2% Rank-1.
• Augmented the pretraining dataset with large-scale open-vocabulary detection/segmentation annotations and attribute tagging annotations in 1-2 days through the utilization of SOTA models, then enhanced pretraining by integrating with new additional tasks, improving zero-shot retrieval performance above 1.0%.

Click to view more

• Proposed the company change from fixed-attribute category person re-identification to Vietnamese Text-based Person Re-identification that is practical for Vietnamese.
• Making this feature the company's flagship AI product compared to other opponents, attracting attention from government agencies.
• Constructed the first Vietnamese pretraining dataset and Vietnamese benchmark dataset for this task, improving fine-tuning efficiency and model generalization.
• Developed a Vietnamese Vision-Language backbone based on ALBEF, leveraging SOTA Vietnamese backbone PhoBERT/Videberta for text encoder and HAP/SOLIDER (human-centric image encoder) for vision encoder.

Click to view more

• Conducted research on Transformer-based architectures for Object Detection and Multiple object Tracking.
• Applied Trackformers (Facebook AI Research, CVPR 2023) for tracking pedestrians and vehicles at the campus of Ho Chi Minh University of Technology.
• Preprocessed realistic surveillance video data and evaluated multiple efficient data annotation tools.
• Optimized Trackformers by modifying loss functions and architecture to extend from single-class (human-only) to multi-class tracking and mitigate class imbalance in realistic training dataset.

Click to view more

• Conducted academic research on Text-based Person Re-identification under the supervision of the Lab Head, who is also the founder of Tris Company. Currently co-authoring a research paper for submission to a high-impact journal.
• Developed a state-of-the-art model, achieving a 2.8% Rank-1 accuracy improvement on benchmark datasets over recent SOTA models, making it the highlighted AI product of the lab and a benchmark for future research.