I developed accurate 3D maps by implementing three local feature/matcher methods: LoFTR, DISK, and KeyNetAffNetHardNet. The competition aimed to reconstruct 3D scenes from multiple views, utilizing unstructured image collections available on the internet. This project has potential applications in photography, cultural heritage preservation, and various Google services, showcasing my expertise in Structure from Motion techniques and advancing to 3D modeling.
Leveraged AI to analyze players’ badminton training. By developing learned models of professional players’ shots and using perspective transformation and bird’s-eye view analysis, I revolutionized training by providing data-driven posture correction and precise performance assessment for amateur players based on distance, reachability, and other crucial metrics.
FitGen is a groundbreaking project that utilizes genetic algorithms and video pose analysis to create custom exercise plans based on individual body weight, height, exercise goals, and heart rate. By optimizing workout routines through real-time feedback on form, FitGen empowers users to achieve their fitness goals efficiently and safely by minimizing the risk of injury.
Developing a log file summarization tool using the Drain algorithm to efficiently extract key insights and patterns from complex log data, enhancing system monitoring and troubleshooting.
Creating a reinforcement learning project using Unity ML Agents, enabling the training of intelligent agents in virtual environments for applications in gaming, robotics, and simulations. Leveraging Unity's versatile platform to develop and fine-tune autonomous decision-making algorithms.
MAGNET is an advanced architecture to generate new data, producing images resembling those in the CelebA dataset. Distinguishing itself from traditional competitive MAP learning in GANs, MAGNET employs a dual-agent cooperative game approach, resulting in superior outcomes compared to the majority of GAN variants.
This project addresses the issue of overfitting in meta-learning by introducing an information-theoretic meta-regularizer, promoting the utilization of task-specific data for better generalization. Instead of relying on mutually-exclusive meta-training tasks, it directly maximizes mutual information using MINE, a scalable and trainable estimator, enhancing meta-learning algorithms' robustness to new tasks. The study's experimental evaluation showcases the effectiveness of this approach on non-mutually-exclusive tasks.
This project is dedicated to creating a comprehensive repository of machine learning models implemented using PyTorch. It provides a valuable resource for both beginners and experts, offering well-structured, PyTorch-based implementations of various machine learning algorithms and neural networks. Users can explore, experiment, and build upon these models for a wide range of applications, from computer vision to natural language processing.
This project leverages TensorFlow and dlib for accurate age and gender estimation from facial images. By employing a Convolutional Neural Network (CNN) for deep learning, it enables simultaneous estimation of age and gender for multiple faces within an image, enhancing its applicability for diverse scenarios requiring demographic analysis.