Shivaam Dave

Worked on development/research of AI-based complimentoring platform for utilizing CCTV video data. Data Acquisition: Used Google Cloud Bucket Storage for data acquisition. Developed an calculus-based algorithm to sample robust frames from large database under time constrained requirement. Data Cleaning: Developed custom image level object detection duplicate removal algorithm using perceptual hashing algorithm and DBSCAN algorithm. Model Selection and Training: Brainstormed test strategy for complained detection module. Developed complained detection module using deepsort tracker and yolov5 model. Tracker was used to maximize the local of complained detection without compromise in precision. Used GCP compute instance for yolov5 model training. Evaluation: Created and developed evaluation framework based on business problem scenario of complained detection module.

Worked on optimization of existing pose estimation approach of mirror AR app-A guided home exercise app for Android devices. Benchmarked inference performance for various state-of-the-art pose estimation models on target Android device emulator at optimal select best model with limited resource requirements. Compared selected approach with the existing using PCK metrics and analyzed the error in it by GRAD-CAM technique. Prioritized must to have vs nice to have features based upon insights. Explored existing solutions and developed a successful POC under senior engineer supervision leading to significant improvement for must have features.

Developed custom lab coat detection model for lab hygiene using YOLOv1 model. Developed demo POC for human activity recognition using 3D ResNet from dept store team.