GCOOD: A Generic Coupled Out-of-Distribution Detector for Robust Classification - Inria - Institut national de recherche en sciences et technologies du numérique Access content directly
Conference Papers Year : 2021

GCOOD: A Generic Coupled Out-of-Distribution Detector for Robust Classification


Neural networks have achieved high degrees of accuracy in classification tasks. However, when an out-of-distribution (OOD) sample (i.e., entries from unknown classes) is submitted to the classification process, the result is the association of the sample to one or more of the trained classes with different degrees of confidence. If any of these confidence values are more significant than the user-defined threshold, the network will mislabel the sample, affecting the model credibility. The definition of the acceptance threshold itself is a sensitive issue in the face of the classifier’s overconfidence. This paper presents the Generic Coupled OOD Detector (GCOOD), a novel Convolutional Neural Network (CNN) tailored to detect whether an entry submitted to a trained classification model is an OOD sample for that model. From the analysis of the Softmax output of any classifier, our approach can indicate whether the resulting classification should be considered or not as a sample of some of the trained classes. To train our CNN, we had to develop a novel training strategy based on Voronoi diagrams of the location of representative entries in the latent space of the classification model and graph coloring. We evaluated our approach using ResNet, VGG, DenseNet, and SqueezeNet classifiers with images from the CIFAR-10 dataset.
No file

Dates and versions

hal-04396403 , version 1 (15-01-2024)





Rogério Ferreira de Moraes, Raphael dos S. Evangelista, Leandro Fernandes, Luis Martí. GCOOD: A Generic Coupled Out-of-Distribution Detector for Robust Classification. 2021 34th SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI), Oct 2021, Gramado, Rio Grande do Sul, Brazil. pp.169-174, ⟨10.1109/SIBGRAPI54419.2021.00062⟩. ⟨hal-04396403⟩


8 View
0 Download



Gmail Facebook X LinkedIn More