Scene Flow Prediction Based On Synthetic Data Generation With Simulated Scenario And Domain Adaptation

Artificial intelligence is in the early stage of cognition nowadays.Recent works about object scene flow prediction are mostly constrained to simple objects and scenes.Problems like shortages of data and the incapability of algorithms exist for scene flow prediction under complex scenes.The limits of sensors make it difficult to obtain and label scene flow data,and the algorithm was not designed from the perspective of common sense cognizing.This dissertation use synthetic data generation and domain adaptation method to provide data basics for research on scene flow prediction.With referring to human cognition,this dissertation proposed the concept of "object descriptor" and its decoder model(ODD)which implements predicting the scene flow under complex scenes.The research details of this dissertation are as follows:A method of synthetic data generation and domain adaptation is proposed,including the obtainment of digital assets,the object placement algorithm,domain transfer algorithm CUT,and the annotation system of virtual environment,which can render synthesis images with ground truth labels;Designed from the perspective of common sense cognizing,ModernCity dataset is proposed which provides RGB images and various labels like depth map,scene flow,and semantic segmentation,the dataset is the first big-scale,physical rules followed dataset of scene flow in simulated scenario,which provides data basics for research on scene flow prediction;Based on the process of inferencing physical world for human beings which includes "finding objectdistinguishing the type and property of object-predicting future statement of object",the concept of "object descriptor" is proposed and an object descriptor decoder model(ODD)is designed,and the ODD model extracts object descriptors for each object in the image through object detection algorithm and pre-trained neural network,and the descriptors are decoded,which partly solves the problem of physical common sense cognizing for computer.This dissertation lays a certain theoretical foundation for machines to have the ability of cognizing and inference the real world,as to provide technical solutions for complex tasks like self-driving and smart medical treatment.