Hello, I'm Duc Hai

I am a Research Resident at VinAI Research working on 3D computer vision and generative modeling.

At VinAI, under the supervision of Dr. Rang Nguyen, I tried to make 3D scene understanding more affordable (by using less supervision) and more practical (by being more robust to real-world data). I am also interested in generative modeling, especially in the context of 3D data. I have had the pleasure to collaborate with Prof. Son Hua, Dr. Phong Nguyen and Dr. Khoi Nguyen .

CV Scholar Github Twitter

News

  • Feb 2025: SharpDepth is accepted at CVPR 2025! Check out our project page for more details.
  • Dec 2024: I am actively looking for PhD/MS opportunities in 3D computer vision and geometry processing.
  • Dec 2024: SharpDepth, our paper on increasing high-frequency details of metric depth estimators, is available on arXiv 2024. We will release the code soon.
  • Dec 2024: VFG-SSC, our paper on semi-supervised 3D semantic scene completion, has been accepted to AAAI 2025. We will release the code soon.

Publications

SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation

SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation

Duc-Hai Pham*, Tung Do*, Phong Nguyen, Binh-Son Hua, Khoi Nguyen, Rang Nguyen
The Conference on Computer Vision and Pattern Recognition (CVPR 2025)

We present SharpDepth, a diffusion-based depth model for refining metric depth estimators, e.g., UniDepth, without relying on ground-truth depth data. Our method can recover sharp details in thin structures and improve overall point cloud quality.

Project Page PDF arXiv
Semi-supervised 3D Semantic Scene Completion with 2D Vision Foundation Model Guidance

Semi-supervised 3D Semantic Scene Completion with 2D Vision Foundation Model Guidance

Duc-Hai Pham, Tuan Ho, Dung Nguyen Duc, Anh Pham, Phong Nguyen, Khoi Nguyen, Rang Nguyen
The 39th Annual AAAI Conference on Artificial Intelligence (AAAI 2025)

We propose a semi-supervised framework for 3D semantic occupancy prediction that reduces reliance on costly labeled data by leveraging 2D foundation models, achieving up to 85% of full-supervision performance with only 10% labeled data.

Project Page PDF arXiv