RGB-D Object Recognition via Incorporating Latent Data Structure and Prior Knowledge

摘要: For the task of RGB-D object recognition, it is important to identify suitable representations images, which can boost performance recognition. In this work, we propose a novel representation learning method for images by jointly incorporating underlying data structure and prior knowledge data. Specifically, convolutional neural networks (CNN) are employed learn image exploiting structure. To handle problem limited RGB depth multi-level hierarchies features trained on ImageNet from CNN transferred rich generic feature while labeled leveraged. On other hand, deep auto-encoders (DAE) exploit knowledge, overcome expensive computational cost optimization in encoding. The expected obtained integrating two types representations. verify effectiveness proposed method, thoroughly conduct extensive experiments publicly available datasets. encouraging experimental results compared with state-of-the-art approaches demonstrate advantages method.