RBE Practicum Presentation
Rahul Allam
Establish ML pipelines for custom 2D/3D Object Detection tasks
Friday, April 28th
10:30 AM - 11:30 AM
Zoom Link: https://wpi.zoom.us/j/3742245696
Abstract: Establishing efficient, scalable and customizable ML pipelines that can produce EDGE deployable applications for various perception/vision tasks has become more important than ever in the robotics field. Specifically, object detection is essential for enabling robots to perceive and interact with the environment. During my internship, I have worked specifically on both leveraging and establishing ML pipeline for 2D and 3D object detection. In regard to 3D object detection, customer specific dataset has been used for training and delivering an EDGE deployable 3D detection model
for detection of the object of interest, followed by establishing a 3D object detection pipeline that can be easily customized for detecting any object of interest in a point cloud dataset. In regard to 2D object detection, I built a GStreamer pipeline to integrate a customer camera into NVIDIA DeepStream (GStreamer-based SDK), followed by leveraging both NVIDIA DeepStream and NVIDIA TAO Toolkit to train, evaluate on custom dataset and publicly available datasets (KITTI dataset), and finally, deploy the model on NVIDIA Jetson AGX Orin
Practicum​ Advisors:
Professor Greg Lewin, Worcester Polytechnic Institute (WPI)
Professor Emmanuel Agu, Worcester Polytechnic Institute (WPI)