Towards Open Domain-Specific Recognition using Quad-Channel
Self-Attention Reciprocal Point Learning and Autoencoder
Abstract
Open-Set Recognition (OSR) has been emphasizing its capability to reject
unknown classes and maintain closed-set performance simultaneously. The
primary objective of OSR is to minimize the risk of unknown classes
being predicted as one of the known classes. The OSR assumes that
unknown classes are present during testing and identifies only one
distribution of the unknowns. On the other hand, recognizing unknowns in
the domain of interest and outside of the domain of interest will
benefit future learning endeavors. These rejected unknown samples in the
domain of interest could be leveraged for the further development of
deep learning models. We introduce and formalize the open
domain-specific space risk to mitigate the recognition of two unknown
distributions. To achieve the solution, we propose an initial baseline
using quad-channel self-attention reciprocal point learning to mitigate
open-space risk and autoencoder to mitigate open domain-specific space
risk. We utilize the knowledge in pre-trained models and tune the
open-set hyperparameter first before comparing it with other methods as
a fair analysis. We also investigate the effect of different pre-trained
models in achieving open-set performance. To validate our approach, we
tested the model with various domains: garbage, vehicle, household
items, and pets. The experimental results demonstrate that our method is
robust in rejecting unseen classes while keeping the closed-set accuracy
performance at bay. Furthermore, the autoencoder could potentially
mitigate open domain-specific space risk in the future, and the
determination of pre-trained models highly contributes to the open-set
performance.