Object Recognition and 3D Pose Estimation Using Improved VGG16 Deep Neural Network in Cluttered Scenes

Abstract

3D pose Estimation and object detection are important tasks for robot-environment interaction. Impressive progress has been made in this field over the past decade. So far, the problem is still challenging because cluttered scenes usually have a negative influence on the recognition process. On the other hand, lack of samples in the training stage also limits the application of the learning-based algorithm. In this work, an improved VGG16 deep neural network pipeline which can extract better feature descriptors is proposed to implement object recognition as well as 3D pose estimation. In addition, we also described a method for quick image data synthesis, which can generate large amount of eligible training data in a short period of time. Experimental results demonstrate the effectiveness and better performance of the proposed method by comparing with classical deep neural networks.