Towards Developing Better Object Detectors for Real-World Use

Deep visual models are fast surpassing human-level performance for various vision tasks, including object detection, increasing their use in day-to-day life applications. It is often the case that standard models that perform well when evaluated on the validation dataset—usually collected from the same source as the training dataset, often perform poorly on data different from that of the training data. Recent works also prove that adversarial examples can easily fool deep learning models and are primarily opaque. To address the issue of making object detectors more compatible for real-world use, we propose some steps to make them more reliable and robust for deployment. Proposed methods include the explanation method and data augmentation techniques. Data augmentation improves the performance and outcomes of machine learning models by generalizing them and explanation methods for getting new insights into black-box detectors. Such understanding can also help improve resistance to a wide range of adversarial attacks.