Online Object Tracking, Learning, and Parsing with And-Or Graphs

Abstract

This paper presents a framework for simultaneously tracking, learning and parsing objects with a hierarchical and compositional And-Or graph (AOG) representation. The AOG is discriminatively learned online to account for the appearance (e.g., lighting and partial occlusion) and structural (e.g., different poses and viewpoints) variations of the object itself, as well as the distractors (e.g., similar objects) in the scene background. In tracking, the state of the object (i.e., bounding box) is inferred by parsing with the current AOG using a spatial-temporal dynamic programming (DP) algorithm. When the AOG grows big for handling objects with large variations in long-term tracking, we propose a bottom-up/top-down scheduling scheme for efficient inference, which performs focused inference with the most stable and discriminative small sub-AOG. During online learning, the AOG is re-learned iteratively with two steps: (i) Identifying the false positives and false negatives of the current AOG in a new frame by exploiting the spatial and temporal constraints observed in the trajectory; (ii) Updating the structure of the AOG, and re-estimating the parameters based on the augmented training dataset. In experiments, the proposed method outperforms state-of-theart tracking algorithms on a recent public tracking benchmark with 50 testing videos and 30 publicly available trackers evaluated [34].

CNNTracker: Online discriminative object tracking via deep convolutional neural network

“…The typical classifiers include SVM [22,23], boosting [6,24], random forest [25], Hough forest [26], structural learning [7,8,27], sparse coding [28][29][30][31][32][33], discriminative feature learning [34], multiple instance learning [35], co-training technique [36], tracking-learning-detection [37], weakly supervised learning [38,39], and-or graphs [14], coupled 2-layer model [40], etc. However, most of these classifiers are limited by their shallow or linear nature structures while object appearance variations are complex, highly nonlinear, and time-varying.…”

“…Feature representation: up to now, a variety of well-known features have been introduced for object tracking, including color histograms [1] or attributes [2], subspace-based features [3,4], Haar-like features [5][6][7][8], LBP [9][10][11], HoG [12][13][14], SFIT [15,16], SURF [17], covariance matrix [18,19], 3D-DCT [20], shape features [21], and combining several complementary cues, etc. While these predefined and handcrafted features have achieved great success for some specific data and tasks, they are low-level features and not tuned for the tracked object.…”

Visual Tracking by Sampling in Part Space

“…Subsequently, it attracts popularity and numerous improvements to DPM have been presented [15]- [17]. In visual object tracking, part-based models have also been proposed to deal with target deformation and partial occlusions [12], [20], [44], [48]. Yao et al [12] presented a part-based tracking method with online latent structured learning.…”

“…In [44], a part-based tracking method with cascaded regression was proposed, which exploits the spatial constraints between parts to learn the intrinsic shape of an object. Lu et al [20] proposed an online tracking-learning-parsing framework that utilizes an and-or graph to capture the construction of objects. Fig.…”

“…Numerous local appearance models have been proposed in recent years and have achieved promising results. Existing methods can be roughly categorized into the following classes: sparse representation based methods [7]- [9], [22], segmentation based methods [5], [26], [29], pooling methods [10], [11] and part-based methods [12], [20], [44].…”