Biologically Plausible Detection of Amorphous Objects in the Wild

Abstract

The problem of amorphous object detection is investigated. A dataset of amorphous objects, Panda bears, with no defined shape or distinctive edge configurations is introduced. A biologically plausible amorphous object detector, based on discriminant saliency templates, is then proposed. The detector is based on the principles of discriminant saliency, and implemented with a hierarchical architecture of two layers. The first computes a feature-based top-down saliency measure tuned for object detection. The second relies on a similar saliency measure, but based on saliency templates, selected from the responses of the first layer. This architecture is shown to have a number of interesting properties for amorphous object detection, including the ability to detect objects characterized by the absence of features, and an interpretation as discriminant blob detection. Extensive experimental evaluation shows that it substantially outperforms state-of-the-art approaches for non-amorphous object detection, such as deformable parts models, sparse coded pyramid matching, detection based on the bag-of-features architecture, and the Viola and Jones approach. This brings into question some currently popular beliefs about object detection, which are discussed.