Intelligent annotation tool for multi-sensor visual recognition tasks

Abstract

This project aims to create a multi sensor annotation tool capable of annotating lidar and camera data. To reduce annotation time it will have the capability of creating annotations for the camera images by projecting the annotations from the 3D space into the image domain. The tool will possess semi automatic labelling features through the integration of object detection models such as Yolo. The tool will have the capability to select images using Active learning strategies to reduce the annotation effort and cost for fine-tuning the object detection model. Interpolation and tracking features will also be present which will track an annotated object across frames and different cameras and this will further reduce the time required for the annotations. The tool will also possess lidar segmentation models to label the lidar data. It will be able to read data provided by multiple sensors and process annotations to give calibration parameters for lidar as an output. Existing tools do not possess all of these features. This would be a novel tool especially for Indian data.