Hi, I am Vighnesh Budharapu

• Benchmarked YOLOv8, YOLOv12, and Faster R-CNN on a self-labeled 9K-frame dataset with targets spanning ~5 px at 1080p, ablating input resolution (4K and 1080p) against model capacity to isolate small-object recall drivers; mAP50 of 0.94, 0.87, and 0.92 • Quantified cross-color generalization via hue-augmentation ablation, training exclusively on red-ball data; mAP50 fell from 0.863 in-domain to 0.111 on unseen white-ball footage • Developed a training-free pipeline for ego-moving umpire-mounted cameras using MOG2 background subtraction, HSV segmentation, and geometry-driven ROI masking, sustaining 30 FPS on CPU at accuracy parity with the trained detectors • Projected detections to field coordinates through homography from pitch-line correspondences at 98% accuracy; combined Kalman filtering over free flight with a particle filter across bounce and occlusion to maintain track continuity • Deployed YOLOv8 and classical CV pipelines as a Dockerized FastAPI service on AWS EC2 with Nginx reverse proxy, SSL, and GitHub Actions CI/CD; automated build, push to ECR, and zero-touch deployment on every commit