Zero-Shot Object Detection for Indoor Robots
Abdalwhab Abdalwhab, Huaping Liu · 2019
Object detection is one of the most crucial tasks for robotic systems. Therefore, a plethora of researches have explored the traditional object detection, where the model must be trained on instances of all objects of interest. However, this is not a very realistic setting because of the huge number of classes in real-world settings and the high cost and time associated with annotating data. Thus, it is almost impossible to train a model in all possible classes. This led to the introduction of the concept of zero-shot object detection, in which a model is trained on some seen classes and then tries to use the obtained knowledge to be able to detect novel unseen classes. Despite the fact that, this setting is more realistic, but very few researchers have explored it. Furthermore, to the best of our knowledge, no previous work has explored it specifically for indoor robots. In fact, indoor environments can be more challenging, since the number of objects of interests is way larger compared to other applications like self-driving cars for instance. In this work, we explore zero shot object detection for indoor robots by embedding deep features and class labels to a shared semantic space. We use the famous SUN RGB-D dataset [1] and our own collected dataset for training and evaluation, and we propose two novel splits for SUN RGB-D dataset especially for zero-shot object detection.