Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization

CERTIFICATE OF ACHIEVEMENT 

Neural Networks and Deep Learning

Control of Mobile Robots

Bachelor of Technology

Benchmarking open source Visual SLAM and Visual Odometry approaches.
∗ Compiling libraries on linux as well as windows platform
∗ Testing out the approaches on publicly available dataset as well as custom dataset
∗ Carrying out performance analysis: accuracy and robustness of localisation as well as compute requirement
∗ Camera calibration

Intern - Computer Vision SLAM

Primarily working on the development of SPATIO- an AR tracking engine which provided real-time camera tracking data based on a single camera feed. Currently SPATIO is used on live sports events such as IPL,Khelo India to enable AR on drone camera feeds broadcasted live to millions of viewers.

∗ Working on software architecture and design of SPATIO: back-end tracking algorithm integration as well as UI/UX
∗ Designed and implemented efficient AR object placement workflow
∗ Integration of tracking data with different front-end engine such as Unreal Engine, Viz Artist
∗ Writing scalable, efficient and bug free code
∗ Handling version control and software release
Briefly worked on ARCore and ARkit to integrate tracking data from Android and iOS devices with broadcast cameras for augmented reality

Augmented Reality Software Engineer

Workshop on aero modelling 

- Worked on hierarchical localization, deep features extraction and matching (super point, superglue) and build a real time camera localization pipeline in C++,libtorch on windows OS.

- Led the R&D of deep frame interpolation algorithms, improving PSNR and LPIPS scores by 25% for sports slow-motion video quality

Research And Development Engineer

Member Of Technical Team

◦ Developed and implemented an AI-based vertical broadcast production pipeline, significantly reducing production costs by converting 16:9 to 9:16 format using spatiotemporal AI models focusing on salient regions
 ◦ Co-designed and optimized real-time system architecture, ensuring a 99.9% uptime during live broadcast
 ◦ Successfully integrated multi-model architecture, boosting the overall accuracy of the pipeline  by 15% at the same time proving more control on the output

Research And Development Engineer II

CV/AI Engineer. Built real-time vertical production AI for Cricket WC'24/Football. C++, SLAM, Real-Time AI

Know about Manoj Sharma’s work experience, education, bootcamps, certifications, and much more. Manoj Sharma