Building Context-Aware Object Detectors: Tying Objects and Context in a loop

Recent object detection research has witnessed a surge in context-driven methods that advocate the use of context to improve detection performance at run time. By priming the set of possible locations and scales for finding an object, context helps the detector to focus its attention to a much smaller search space. However these methods continue to use the detectors trained in the conventional setting i.e. learn a classifier discriminating positive bounding boxes (containing an object instance) and negative image patches taken from ALL possible locations/scales. This paper takes a step ahead and proposes to use contextual information even while training an object detector. As context primes the detector’s focus towards context-plausible image regions at test time, it would benefit by even training the detector only using examples sampled from the contextplausible regions. This update not only makes the training process computationally efficient but also naturally provides those ‘hard’ negative instances to the detector that it would find potentially confusing at run time. We demonstrate our hypothesis on the task of pedestrian detection using 3D camera viewpoint constraints as context.