Angular Penalty for Few-Shot Incremental 3D Object Learning

Bingtao Ma, Yang Cong

Abstract: 3D object recognition has garnered notable success in both academic and industrial contexts. However, the majority of existing 3D object recognition approaches are tailor-made for static scenarios that have ample availability of training instances. Therefore, we target a more arduous and pragmatic task: few-shot incremental 3D object learning (FSI3DL), which aims to learn the new 3D objects in an incremental manner, yet new classes only have a few training instances. However, this task presents two challenges: overfitting to few-shot, and catastrophic forgetting on previous classes due to the absence of previous training instances. To mitigate these challenges, we introduce an Angular Penalty method to consume irregular mesh directly and reserve embedding space for incoming new classes. Specifically, we design a graph convolution network that can take advantage of the better representation ability of mesh and overcome its irregularity. Moreover, we use an angular penalty loss to increase the similarity between inra-class instances and reduce the similarity between instances from different classes. This can leave space in embedding space for incoming new classes. Experiments on representative 3D object classification datasets demonstrate the better efficacy of our method.