Glocal Energy-based Learning for Few-Shot Open-Set Recognition
This addresses the practical problem of few-shot learning with unknown classes, but appears incremental as it builds on existing energy-based and hybrid approaches.
The paper tackles few-shot open-set recognition by proposing a glocal energy-based hybrid model that combines class-wise and pixel-wise features to detect unknown classes, achieving superior performance on three standard datasets.
Few-shot open-set recognition (FSOR) is a challenging task of great practical value. It aims to categorize a sample to one of the pre-defined, closed-set classes illustrated by few examples while being able to reject the sample from unknown classes. In this work, we approach the FSOR task by proposing a novel energy-based hybrid model. The model is composed of two branches, where a classification branch learns a metric to classify a sample to one of closed-set classes and the energy branch explicitly estimates the open-set probability. To achieve holistic detection of open-set samples, our model leverages both class-wise and pixel-wise features to learn a glocal energy-based score, in which a global energy score is learned using the class-wise features, while a local energy score is learned using the pixel-wise features. The model is enforced to assign large energy scores to samples that are deviated from the few-shot examples in either the class-wise features or the pixel-wise features, and to assign small energy scores otherwise. Experiments on three standard FSOR datasets show the superior performance of our model.