Zero-Shot Object Detection

Owing to the large number of real-world applications of Object Detection namely, robotics, self-driving cars, medical imaging, surveillance, etc, Object Detection has seen many successes over some benchmarks datasets such as PASCAL VOC, Imagenet and MS COCO. However, drone datasets are much more difficult where the challenges of Object Detection are compounded. We demonstrate the effectiveness of state-of-the-art Object Detectors on VisDrone (2018) dataset (which is a drone dataset) and explore improvements on the best performing detector (Faster R-CNN). We also note that in real-world high level vision tasks which require Object Detection for numerous categories, dependence on a large amount of annotations can act as an obstacle to the task. Zero-Shot Object Detection (ZSD) - where training examples are not available for target classes - aids in overcoming this problem as it provides semantic scalability to detecting objects. In this thesis, we propose a novel multimodal approach for ZSD where we combine predictions obtained in different search spaces with potent discriminative capabilities. We learn individual projections in semantic and visual spaces, extract useful information from joint space, predict similarity scores in the individual spaces and combine them. We present state-of-the-art results on two popular datasets, PASCAL VOC and MS COCO. We also demonstrate how our approach alleviates a problem inherent to Zero-Shot Recognition (ZSR) - called hubness - thereby resulting in a performance superior to previously proposed methods.