Bengaluru, Karnataka, India
https://geekysethi.github.io/
Face Detection • Architected face detection on edge devices, end to end. • Enabled face detection with mask • Improved mAP score from 54% to 76%. • Deployed model on Qualcomm SNPE, which is serving on 20K trucks/cars worldwide. Head Pose Estimation • Developed pose estimation for faces with mask. • Use of contrastive loss brought down the MAE from 15.7 to 9.4. • Architected the uncertainty-based pose estimation algorithm to reduce the falses. • Uncertainty-based pose estimation reduced the falses by 15%. Other Responsibilities • Collaborating with IIIT Delhi for federated learning project on Edge Devices. • Took interviews of more than 20 candidates in the machine learning domain.
• Architected the whole problem from data collection to model training. • Developed a new small-size dataset of malnourished and healthy children using images from the internet with help of doctors. • Developed deep learning-based algorithm to work on small size dataset for malnourishment detection. • Designed a 3D printing model of cradle to collect image dataset of newborn babies at the hospital.
• Developed the new algorithm for camera calibration in omnistereo cameras end to end. • We considered each camera as virtual camera which is present at a tangent to the imaginary circle of the field of view. • This algorithm can be adapted to any field of view of the camera.
◦ Worked on pedestrian tracking using mean shift tracking and background subtraction algorithms. ◦ Object detection using HOG features and Random forest. ◦ Complete the literature survey of object detection using non-deep learning methods.
◦ Camera calibration of stereo camera setup. ◦ Developed 3D maps of Anterior Segment of the eye.