Distance Invariant RGB-D Object Recognition Using DSMS System

Abstract

In computer vision, object recognition has gained a lot of attention due to its numerous practical usage. For real-world applications, it is necessary to consider conditions like object images are captured from multiple viewpoints, change in illumination and different distance locations of objects from the camera for better recognition. In this work, a new CVPR34K RGB-D dataset is proposed consisting of RGB-D images which are acquired from different distance location from the camera. A distance invariant RGB-D object recognition system is introduced using Depth Estimation, Scale data with Unit Depth and Multimodal Convolutional neural network with SVM (DSMS). The proposed DSMS system is divided into three parts. First, the Depth Estimation is introduced to detect distance location of acquired RGB-D object image. The second stage consists of several preprocessing operation to normalize input RGB-D data with respect to a reference distance. The final stage is to learn features from normalized RGB and depth images and performed RGB-D object recognition. The experimental results show that the DSMS method achieves comparable performance to state-of-the-art methods on the RGB-D object dataset. Effectiveness of our method is clearly observed for the cases when distance location RGB-D object image is changed in proposed CVPR34K Dataset.