Experiments on Zero-Shot Object Detection for Indoor Robots

Zero-shot object detection is a new very challenging concept in computer vision. While zero-shot object detection is a more practical setting than traditional object detection, not many researchers have explored it, and not for specific applications like indoor robots. In this work, experiments to evaluate zero-shot object detection were conducted on images collected by a real indoor robot whereas the model was trained on a larger dataset (SUN RGB-D [1]). This work aims to evaluate this more challenging setup resulting from the difference in distribution between training and testing images.In this work, the image visual features were extracted and embedded with class labels to a common semantic space. Then, the network was trained to learn an aligning function that maps the image embedded visual features to the embedded class name, to be capable of predicting novel classes that were not presented at training time.Despite that, the currently achieved results are not very high, but this research opens the doors for further exploration of this challenging setup.