Rakesh Reddy Kondeti

- Increased 2 AP points by researching and analyzing the potential benefits of class prototyping and feature fusion in dense meta-detectors, when compared to traditional methods. - Improved the performance by 3 AP points by implementing the SOTA Swin Transformer-based backbone into the meta-learning-based one-stage few-shot object detection pipeline, which is the first of its kind. - Achieved a state-of-the-art result with an impressive AP increase from 14.4 to 19.4 by conducting experiments on 10-shot COCO

- Conducted literature review of object detection & meta-learning research papers, identifying key trends & patterns in the field. - Trained and fine-tuned the YoloX model in the Detectron2 framework, based on meta-learning few-shot context.

- Collaborated with the Learn2Trust project team to integrate the software plugin into their AI course curriculum for medical students, to upload images and produce masks using Streamlit.

- Implemented U-net & Pix2Pix GAN networks to generate depth images from monocular RGB bronchoscopy images. - Improved model performance by 4% increase in SSIM metric, by combining SSIM loss and mean gradient error with the L1 loss. - Achieved sharper depth images with an SSIM metric of around 97%, surpassing the previous model from 93%.

Artificial Intelligence, Machine Learning, Deep Learning, Computer Vision, Reinforcement Learning