A Scene-specific Deformable Part-based Model For Object Detection

Detecting and localizing specific objects is one of the fundamental challenges in computer vision, and significant strides have been made in the past few decades. However, objects vary significantly in appearance, viewpoint, scale, illumination, texture, etc., which poses a great challenge to object detection. Most state-of-the-art object detectors follow the sliding-window paradigm, i.e., extracting features from an input window and deciding whether there exists an object in it.While most existing models focus on detection in static images, we investigate the static video surveillance scenario. Two typical challenges have been widely recognized for object detection in the static video surveillance. Firstly, objects appear at a wide range of scales in real world. It is difficult to detect small-scale objects because their visual information is essentially insufficient, and detectors with rigid template sizes can hardly recognize objects that are smaller than the template. Secondly, the detectors are likely to confuse target objects with background or semantically similar objects. As a result, standard low-level detectors are prone to false alerts and poor localization of objects.To address the performance gap, we propose a probabilistic graphical model that integrates a local generic object detector and scene-specific contextual features. The local object detector is a Deformable Part-based Model(DPM) represented in a multi-resolution structure to detect objects over various scales. Two scene-specific contextual features, namely ground plane estimation and object position estimation, are used to reduce false positives and enhance the overall detection accuracy. Experimental results on the public datasets CAVIAR and LISA demonstrate that our model surpasses the conventional DPM. In addition, our model can be easily adapted to a new scenario without a re-training process.