An important open problem in computer vision is open-set recognition: determining whether input data is from known or unknown classes. Two dominant approaches exist for open-set discrimination: learning a binary classifier from outlier data and using a GAN discriminator as a likelihood function. Both have limitations: binary classifiers overfit to available outlier data, while GAN discriminators are unstable and unreliable. OpenGAN addresses these through three innovations: (1) carefully selecting GAN discriminators on real outlier data, (2) augmenting open training examples with adversarially synthesized fake data, and (3) building discriminators over features from K-way networks rather than pixels.

Real-world recognition systems must handle open-set conditions: inputs may belong to classes not seen during training. This is critical for safety applications such as autonomous driving, where "contemporary benchmarks such as Cityscapes focus on K classes of interest for evaluation, ignoring a sizeable set of 'other' pixels that include vulnerable objects like wheelchairs and strollers." Two conceptual approaches exist: learning binary open-vs-closed classifiers from outlier data (which generalizes poorly to diverse test data) and using GAN discriminators as likelihood functions (which performs poorly due to unstable GAN training). OpenGAN combines these insights by operating on learned feature representations rather than raw pixels.

Open-set recognition methods typically train K-way classifiers then exploit them for detecting unknown classes through density estimation, uncertainty modeling, and reconstruction errors. OpenMax replaces the softmax layer with a calibrated open-set probability. Prior work using GANs for open-set recognition generates fake data to augment training sets and relies on reconstruction error, but the discriminator itself has not been successfully used as a likelihood function due to training instability. Outlier exposure methods show that simple binary classifiers work "surprisingly well" but overfit to available outlier data, failing to generalize across datasets.

The OpenGAN objective combines real and synthetic open data: max_D min_G E[log D(x)] + lambda_o * E[log(1-D(x_bar))] + lambda_G * E[log(1-D(G(z)))], where D(x) represents the probability of closed-set membership, x are closed-set features, x_bar are real outlier features, and G(z) generates synthetic outlier features. The generator G learns to produce features that fool the discriminator into classifying them as open-set, while the discriminator D learns to distinguish closed-set features from both real and synthesized outliers. This adversarial interplay creates a more robust decision boundary than training on real outliers alone.

The critical design choice is building discriminators over "off-the-shelf features computed by closed-world K-way networks" rather than raw pixels. Using a pre-trained ResNet-18 or HRNet to extract low-dimensional feature vectors (512-dim or 720-dim), the GAN operates in a compact feature space where the discriminator architecture is a simple MLP with batch normalization and LeakyReLU (~2MB). This design choice yields dramatic improvements: on Cityscapes segmentation, feature-based OpenGAN achieves 0.885 AUROC vs. only 0.549 for pixel-based variants. The features already encode class-discriminative information learned during closed-world training, giving the discriminator a strong starting representation.

A crucial finding is that model selection requires a validation set of real outlier data. The authors demonstrate that longer GAN training does not consistently improve open-set performance due to training instability. The solution is "open validation": using a small held-out set of real outlier examples to select the discriminator checkpoint that achieves the best open-vs-closed classification accuracy. This is critical because "synthesized data are insufficient for model selection" — the generator may collapse or drift, producing samples that no longer represent the true open distribution. The requirement for a small outlier validation set is a practical limitation, but the authors show that even a few hundred examples suffice.

Experiments span three setups. Setup I (single dataset): On MNIST, SVHN, CIFAR, and TinyImageNet, OpenGAN achieves AUROC of 0.999, 0.988, 0.973, and 0.907 respectively, substantially outperforming prior methods. Setup II (cross-dataset): Using TinyImageNet as closed-set with diverse open-sets, OpenGAN achieves 0.984 average AUROC, far surpassing binary classifiers (0.918) which overfit to training outliers. Setup III (semantic segmentation): On Cityscapes with 19 closed classes, feature-based OpenGAN achieves 0.885 AUROC with full data, vastly exceeding pixel-based (0.549) and entropy-based (0.697) methods. The model is also extremely lightweight: ~2MB discriminator vs. 250MB HRNet backbone.

OpenGAN demonstrates that GAN discriminators can achieve state-of-the-art open-set discrimination once properly selected using a validation set of outlier examples. The key insight is that operating in learned feature space, combining real and synthesized outliers, and using open validation for model selection together overcome the long-standing limitations of both binary classifiers and GAN-based approaches. The method is lightweight, practical, and significantly outperforms prior methods across classification and semantic segmentation tasks. This work challenges the assumption that discriminators cannot serve as effective likelihood functions, opening new avenues for GAN-based open-set recognition.