Aggregated Deep Convolutional Neural Networks for Multi-View 3D Object Retrieval

Extracting and aggregating discriminative image features is a key challenge for 3D multi-view object recognition and retrieval tasks. In this paper, we propose aggregated deep CNNs (ADCNN) model to address the limitations associated with projecting 3D models into multiple 2D images and their resulting high-dimensional representations. A deep learning network is developed to aggregate compact features of 3D objects using the activation kernels of convolutional layers directly. Two instances of the same CNN features extractor share the learning weights while they represent different object characteristics. Systematic experiments conducted on the benchmark dataset ModelNet40 demonstrate the efficacy of the proposed method in 3D object retrieval and a mAP accuracy of 91.1% is achieved, which shows its performance superiority over related state-of-the-art methods.