Multi-level Feature Learning Research Articles

Holistically-Nested Edge Detection

We develop a new edge detection algorithm that addresses two important issues in this long-standing vision problem: (1) holistic image training and prediction; and (2) multi-scale and multi-level feature learning. Our proposed method, holistically-nested edge detection (HED), performs image-to-image prediction by means of a deep learning model that leverages fully convolutional neural networks and deeply-supervised nets. HED automatically learns rich hierarchical representations (guided by deep supervision on side responses) that are important in order to resolve the challenging ambiguity in edge and object boundary detection. We significantly advance the state-of-the-art on the BSDS500 dataset (ODS F-score of 0.790) and the NYU Depth dataset (ODS F-score of 0.746), and do so with an improved speed (0.4 s per image) that is orders of magnitude faster than some CNN-based edge detection algorithms developed before HED. We also observe encouraging results on other boundary detection benchmark datasets such as Multicue and PASCAL-Context.

Nonrigid 3D shape retrieval using deep auto-encoders

The soaring popularity of deep learning in a wide variety of fields ranging from computer vision and speech recognition to self-driving vehicles has sparked a flurry of research interest from both academia and industry. In this paper, we propose a deep learning approach to 3D shape retrieval using a multi-level feature learning paradigm. Low-level features are first extracted from a 3D shape using spectral graph wavelets. Then, mid-level features are generated via the bag-of-features model by employing locality-constrained linear coding as a feature coding method, in conjunction with the biharmonic distance and intrinsic spatial pyramid matching in a bid to effectively measure the spatial relationship between each pair of the bag-of-feature descriptors. Finally, high-level shape features are learned by applying a deep auto-encoder on mid-level features. Extensive experiments on SHREC-2014 and SHREC-2015 datasets demonstrate the much better performance of the proposed framework in comparison with state-of-the-art methods.

Deep shape-aware descriptor for nonrigid 3D object retrieval

Deep learning is a rapidly growing discipline that models high-level features in data as multilayered neural networks. In this paper, we propose a deep learning approach for 3D shape retrieval using a multi-level feature learning methodology. We first extract low-level features or local descriptors from a 3D shape using spectral graph wavelets. Then, we construct mid-level features from these local descriptors via the bag-of-features paradigm by employing locality-constrained linear coding as a feature coding method, together with the biharmonic distance as a measure of the spatial relationship between each pair of bag-of-feature descriptors. Finally, high-level shape features are learned via a deep auto-encoder, resulting in a deep shape-aware descriptor that is compact, geometrically informative and efficient to compute. The proposed 3D shape retrieval approach is evaluated on SHREC-2014 and SHREC-2015 datasets through extensive experiments, and the results show compelling superiority of our approach over the state-of-the-art methods.