Convolutional Neural Networks for Predicting 2D and 3D Skeleton of Diverse Objects

Abstract

Computer vision, similar to the human visual system, helps computers with high–level intelligence to understand digital images or videos. Deep learning has made an impact on the rapidly developing field of computer vision. The years of work in the human visual system and machine vision have pointed out that common intermediate representations are always inherent in inference work and provide the necessary structure to improve generalization. Convolutional Neural Network (CNN) is one of the most commonly used scene interpretation approaches, including a black-box mapping. The proposed system is a credible framework to standardize intermediate concepts with supervision at deep CNN with the Gabor feature improves the generalization. The intermediate concepts are applied in a sequence to hidden layers for supervision and applied Gabor features for predicting 2D/3D object skeletons and object classification. This paper implemented 2D/3D key point localization and object classification of images on real image datasets such as IKEA, which achieves state-of-art performance.