Harsh Agarwal

The primary focus is to embed hierarchical class labels for accurate object classification of automotive RADAR 3D point cloud. • Researched the impact of multiple embedding strategies in both Euclidean & Hyperbolic Space, leveraging the intrinsic semantic hierarchy • Integrated entropy based multi-level network modifications enabling adaptive collaboration & information sharing Supervisors: - Dr. Martin Elmer and Mr. Raphael Kolk, Bosch - Prof. Philipp Slusallek and Dr. Christian Müller, DFKI

• Implemented Confidence Estimation Network over trained model, to estimate the prediction confidence for improved post processing • Developed novel location stacking algorithm with a ring buffer for the production code

Next Generation RADAR • Developed a novel DeepLoc Network that performs object type classification on point-cloud from automotive RADAR • Created metrics for efficient testing of classification model for the Automatic Emergency Braking use-case Synthetic Traffic Sign Generator • Developed the pipeline to deal with data-imbalance problem with hundreds of classes for traffic sign image classifier • Used conventional techniques like affine transformation, background blending for a deterministic generation

• Optimized and translated Classification and Segmentation models for faster computation on Jetson TX2 using TensorRT in the half-precision mode • Reduced model's final inference time by about 70% with 0.5% change in accuracy • Worked with YOLO v3 model for head detection in sports dataset