Non-rectangular Part Discovery for Object Detection

Abstract

The deformable part-based model (DPM) is commonly used for object detection and many efforts have been made to improve the model. However, much less work has been done to discover parts for DPM. Most DPM-based methods adopt the greedy search approach proposed in [2] to initialize a predefined number of parts of rectangular shapes, which may not be optimal for some object categories. Moreover, object structures are not well exploited by the approach. In [4], a three-layer spatial pyramid structure is used to simplify the initialization of parts. An And-Or tree model [3] is proposed to select discriminative part configurations by a dynamic programming algorithm. Although the method can determine part sizes automatically, part shapes are still restricted to rectangles. To address the limitations of these methods, we propose a novel datadriven approach to discover non-rectangular parts by exploiting object structures. Figure 1 shows rectangular and non-rectangular parts obtained by the greedy search approach and our approach, respectively. Generally, the parts obtained by our approach can better cover object regions.

Part-based deformable object detection with a single sketch

“…The basic assumption in most of these cases is that an object can be represented in terms of a collection of local templates that deform and articulate with respect to one another, while maintaining a good amount of independence of structural variations within each of these templates. In contrast to the established methods [32,48], which attempt to learn the optimized part model of an object from a large training set, we aim to minimize this effort by proposing an automatic part-decomposition approach inspired by a normalized cut based method by Gopalan et al [43]. As a first step, given a closed sketch, 1 we follow an area-based approach of Ling and Jacobs [49], as also used by Gopalan et al [43], to roughly identify some significant concave points (also called nonconvex points, identified based on a convexity ratio [43] of 0.8) in the sketch.…”

Discriminative Middle-Level Parts Mining for Object Detection

“…The models are trained with 6 roots and 8 parts per root. To validate the advantages of the proposed method, we compare it with three related methods: DPM-release 5 [27], and-or tree (AOT) models [22] and non-rectangular part discovery (NPD) [29]. DPM performs greedy part search for part model initialization.…”

“…Although deeper hierarchies can capture finer information in the tree model, background regions or suboptimal parts may be selected. Zhou et al [29] discover non-rectangular parts by exploiting object structures. Song et al [22] propose an and-or tree model to learn part configurations.…”

“…And-or tree (AOT) models [22] configurations by a dynamic programming algorithm. Nonrectangular part discovery [29] is proposed for DPM by exploiting object structures. In general, existing DPM based methods directly derive parts from the root model.…”

“…DPM handles appearance variations and has achieved competitive performance on the PASCAL VOC benchmark datasets. Many efforts [12], [13], [18]- [20], [22], [29] have been made to improve DPM recently. However, the performance of DPM is still limited by inferior part search.…”