3D object retrieval with stacked local convolutional autoencoder

• It proposes a novel 3D model retrieval method with stacked local convolutional autoencoder (SLCAE).
• The greedy layerwise strategy is applied to train the SLCAE.
• A gradient descent method is used for training each layer.
• The experiments are conducted on three publicly available 3D object datasets, and the results demonstrate that the proposed method can greatly improve 3D object retrieval performance, compared with several state-of-the-art methods.

The success of object recognition and retrieval is largely determined by data representation. A good feature descriptor can detect the high-level abstraction of objects, which contains much discriminative information. In this paper, a novel 3D object retrieval method is proposed based on stacked local convolutional autoencoder (SLCAE). In this approach, the greedy layerwise strategy is applied to train SLCAE, and gradient descent method is used for training each layer. After the processing of training, the representations of input data can be obtained, regarded as the features of 3D objects. The experiments are conducted on three publicly available 3D object datasets, and the results demonstrate that the proposed method can greatly improve 3D object retrieval performance, compared with several state-of-the-art methods.