Bag of feature with discriminative module for non-rigid shape retrieval

Abstract

Non-grid shape retrieval has been an important and challenging task that usually requires high-level and discriminative shape features. However, most of the existing descriptors are hand-crafted, which fail to distinguish and focus on important features from different regions and scales for given query shape. To tackle this issue, this paper presents a novel shape retrieval method, namely BoF-DM (Bag of feature with Discriminative Module), which is specially designed to highlight discriminative features from the multi-scale context structures. Specifically, a shape decomposing method is designed by combining the KPs (key points) and ring graph, result in local to global feature learning. To obtain high-level shape features, the spatial context of each contour fragment is learned by the stacked lightweight network in an unsupervised manner. In people's cognition, discriminative parts, such as heads or legs, play an essential role for identifying animal shape classes. Inspired by this phenomenon, we propose a TF-IDF (term frequency-inverse document frequency) with entropy-based Discriminative Module, which fully considers the attention issue of inter-class difference and intra-class similarity. Finally, decision-making is done by metric similarity. The experiments demonstrate that the proposed method achieves the high BER (Bull's Eye Retrieval) rates of 93.63%, 100%, 100%, and 99.00% on the considered MPEG-7, Kimia's 99, Kimia's 216, and Articulated shape datasets respectively. Furthermore, our method also obtains high performance under noisy and articulated conditions.