Abstract1

Part model-based methods have been successfully  applied to object detection and scene classification and  have achieved state-of-the-art results. More recently the  "sparselets" work [1-3] were introduced to serve as a  universal set of shared basis learned from a large number  of part detectors, resulting in notable speedup. Inspired by  this framework, in this paper, we propose a novel scheme to  train more effective sparselets with a coarse-to-fine  framework. Specifically, we first train coarse sparselets to  exploit the redundancy existing among part detectors by  using an unsupervised single-hidden-layer auto-encoder.  Then, we simultaneously train fine sparselets and  activation vectors using a supervised single-hidden-layer  neural network, in which sparselets training and  discriminative activation vectors learning are jointly  embedded into a unified framework. In order to adequately  explore the discriminative information hidden in the part  detectors and to achieve sparsity, we propose to optimize a  new discriminative objective function by imposing L0-norm  sparsity constraint on the activation vectors. By using the  proposed framework, promising results for multi-class  object detection and scene classification are achieved on  PASCAL VOC 2007, MIT Scene-67, and UC Merced Land  Use datasets, compared with the existing sparselets  baseline methods