Key Projects & Achievements
Monocular 3D Estimation and Deployment
- Designed data samplers and orchestrated end-to-end workflows using Airflow for critical object selection
- Built and trained deep learning models for monocular 3D estimation on AWS using PyTorch; deployed models on embedded systems with TensorRT and C++
- Achieved \(\leq 3\%\) relative error in distance estimation up to 400 meters and improved size estimation by 10%
Stereo 3D Perception and Annotation
- Led the design and deployment of an auto-annotation pipeline that generated over 500,000 stereo 3D samples
- Developed advanced stereo 3D models using 2D priors; achieved 90%+ projection accuracy in challenging terrain (hilltop/downhill)
- Created real-time onboard APIs in C++ integrating TensorRT-optimized models
Large-scale Data Pipelines & Benchmarking
- Built and maintained scalable pipelines to generate over 5 million labeled 3D ground truth data points (mono, stereo, single-frame), efficiently stored and queried using SQL
- Designed and launched a benchmarking framework for long-range depth estimation (beyond 250 meters), enabling rigorous model evaluation and experimentation
- Unified disparate tools into a centralized platform, improving collaboration across research teams and standardizing performance assessment for all 3D perception models