Object detectors are typically trained once and for all on a fixed set of classes. However, this closed-world assumption is unrealistic in practice, as new classes will inevitably emerge after the detector is deployed in the wild. In this work, we look at ways to extend a detector trained for a set of base classes so it can i) spot the presence of novel classes, and ii) automatically enrich its repertoire to be able to detect those newly discovered classes together with the base ones. We propose PANDAS, a method for novel class discovery and detection. It discovers clusters representing novel classes from unlabeled data, and represents old and new classes with prototypes. During inference, a distance-based classifier uses these prototypes to assign a label to each detected object instance. The simplicity of our method makes it widely applicable. We experimentally demonstrate the effectiveness of PANDAS on the VOC 2012 and COCO-to-LVIS benchmarks. It performs favorably against the state of the art for this task while being computationally more affordable.

Object detection has been an important task in the computer vision domain in recent decades. The goal is to detect instances of objects, such as humans, cars, etc., in digital images. Hundreds of methods have been developed to answer a single question: What objects are where? Traditional methods tried to answer this question by extracting hand-crafted features like edges and corners within the image. Most of these approaches used a sliding-window approach, meaning that they kept checking small parts of the image in different scales to see if any of these parts contained the object they were looking for.