Towards open domain-specific recognition using Quad-Channel Self-Attention Reciprocal Point Learning and Autoencoder

Abstract

Open-Set recognition (OSR) emphasizes its ability 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. OSR operates under the assumption that unknown classes will be present during testing, and it identifies a single distribution for these unknowns. Recognizing unknowns both within and outside the domain of interest can enhance future learning efforts. The rejected unknown samples within the domain of interest can be used to refine deep learning models further. We introduced a new challenge within OSR called Open Domain-Specific Recognition (ODSR). This approach formalizes the risk in the open domain-specific space to address the recognition of two distinct unknown distributions, i.e., in-domain (ID) and out-of-domain (OOD) unknowns. To address this, we proposed an initial baseline that employs Quad-channel Self-attention Reciprocal Point Learning (QSRPL) to mitigate open-space risk. Additionally, we used an Autoencoder to handle open domain-specific space risk. We harnessed the knowledge embedded in pre-trained models and optimized the open-set hyperparameters before benchmarking against other methods. We also explored how different pre-trained models influence open-set recognition performance. To validate our method, we tested our model across various domains, including garbage, vehicles, household items, and pets. Experimental results indicate that our approach is effective at rejecting unseen classes while maintaining closed-set accuracy. Furthermore, the Autoencoder shows potential in addressing open domain-specific space risks for future development. The choice of pre-trained models significantly affects the performance of open-set recognition in rejecting unknowns. The source code of the project is available at https://github.com/gusti-alfarisy/QSRPL.