A Fine-Grained Filtered Viewpoint Informed Keypoint Prediction from 2D Images

Viewpoint informed keypoint prediction from 2D images is an essential task in computer vision, which captures the fine details of rigid objects, however, the cases of ambiguous viewpoint predicted by the convolutional neural network, especially for two peaks of high confidence viewpoint proposals, may specify a set of erroneous keypoints. To address the above issue, we present multiscale convolutional neural networks and propose a filter to ensure high confidence viewpoint informed, which provides a global perspective for keypoint prediction. Leveraging the global precedence, we combine multiscale local appearance based keypoint likelihood with filtered viewpoint conditioned likelihood to induce a considerable performance gain. Experimentally, we show that our framework outperforms state-of-the-art methods on PASCAL 3D benchmark.