Ravi Kiran Basuthkar

Edge Model:

• Examined the shortcomings of detectron model and developed an efficient image segmentation model for car component identification.
• Using post-training quantization model size is minimized to 1.5 MB, enabling deployment on Android devices.
• This Simplified model assists users with pre-inspection video capture by classifying car pose and providing guidance (Increased user base by 5%).

Frame Extraction:

• Streamlined the auto pre-inspection process by means of a frame extraction bot using a SQS queue
• This service efficiently extracts frames showcasing a car in all necessary poses from pre-inspection videos, routing them for further processing.
• The model is jointly used to segment out components of car, classify the pose of car, guide the customer by showing the directions and calculate the bounding box using regression.

Odometer Recognition:

• Developed and deployed a Transformer (Donut) model for robust Optical Character Recognition (OCR).
• Achieved significant improvements in accuracy: odometer reading (58% to 82%) and registration number (92%).

Fraud Detection:

• Led a project focused on vehicle fraud detection through background manipulation identification.
• Initially implemented image processing techniques (SIFT, Superglue) achieving 46% accuracy.
• Then Redesigned an autoencoder based approach, significantly improving accuracy to 78%.

In-house Face Recognition:

• Working on security surveillance project which involves automating attendance and authorization using face recognition by leveraging Kafka queue for processing data from real time streams.

Face Recognition on Edge and Web:

• Created a secure employee access system using a custom face recognition pipeline which included data acquisition, model training and deployment on Raspberry Pi 4 with door access control. Was able to achieve 3 FPS in processing frames on edge device.
• Established a comprehensive employee portal web application for registration, data management (attendance, timestamps), and backend administration.

Knowledge Distillation and Pruning:

• Employed knowledge distillation and pruning techniques to achieve efficient model deployment.
• Transferred knowledge from larger models (ResNet-50) to smaller, faster models (ResNet-18, ResNet-10) while maintaining accuracy.
• Network pruning resulted in a 4x reduction in model size and 3x improvement in inference speed.

Pedestrian Detection and Tracking:

• Developed a custom pedestrian detection model for video surveillance applications by using the TinyYOLOv3 architecture as backbone.
• Trained from scratch with by optimizing anchor boxes for the dataset, and integrated an Intersection over Union (IOU) tracker for robust person tracking.