CVPR Daily - Tuesday

Generalized few-shot object detection is a powerful technique that can improve the performance of object detection models. For example, in self- driving cars , the algorithm is trained to detect a wide range of objects and environments. However, in the real world, there is not enough training data to model every object and environment the car may encounter, so users may meet instances that the model has not been trained on and needs to adapt for . After the adaptation, the model must retain the ability to precisely detect the original pre-trained objects. Long-tail object detection is another relevant scenario. For example, in the cat species, some object classes, such as domestic cats, are particularly prevalent. In contrast, others are rare, such as tigers. In these cases where collecting enough training data for all classes may not be possible, generalized few-shot object detection can help. Jiawei tells us this work has not been without its challenges, which have covered two aspects: the technical and the conceptual. “ On the technical side, how can we ensure the well-separated classifier weights after those experimental observations? How can we ensure a perfect geometry of the expected feature distribution? To have a good object detector for the base and novel classes, we must carefully design the feature space’s geometry”, he poses. “ On the conceptual side, because this TFA framework is popular and no one has ever studied it, there is little related literature. ” 13 DAILY CVPR Tuesday DiGeo: Discriminative Geometry-Aware Learning ...