Out-of-Distribution Detection with Semantic Mismatch under Masking

In general, OOD detection methods can be categorized into: classification-based, density-based and distance-based methods [46].

Classification-based methods derive OOD scores based on the output of DNNs, as shown in Fig. 1(a). Maximum Softmax Probability (MSP) [16] simply uses the maximum softmax probability as the indicator of In-D data. ODIN[27] is applies a temperature scaling to the softmax value for OOD detection. Follow-up works include methods based on the output of DNNs [29, 43, 28, 44], the gradient of DNNs [20] and data generation or augmentation[42, 41]. Although simple to implement, most of them alter the training process of the original classifier, thereby reducing the classification accuracy for In-D samples.

Density-based methods usually apply some probabilistic models for the distribution of In-D samples and regard test data in low-density regions as OOD [26, 37, 3], as shown in Fig. 1(c). Some methods in this category also resort to generative models [36] to learn the distribution of data. However, recent research found that the learned density model may assign high likelihood value to some OODs, since the obtained likelihood could be dominated by low-level features such as location and variance instead of the high-level semantics, which is related to the specific network architecture and data used for learning [31, 3].

Distance-based methods consider that OODs should be relatively far away from In-Ds. They either calculate the centroids of In-D classes in the feature space [50, 19] (Fig. 1(b)) or reconstruct the input itself (Section. 2.2, Fig. 1(i)) for OOD detection. However, for high-level semantic features, their assumption for distance disparity may not hold, and high reconstruction quality cannot ensure In-Ds. In this paper, we use conditional synthesis on masked images to highlight the semantics difference in the image space.

Reconstruction-based methods, which fall into the category of distance-based methods, are closely related to the proposed MoodCat technique. These methods are based on the assumption that In-D data can be well-reconstructed from a trained generative model, but OOD cannot as they are not seen during training (see Fig. 1(i)). Previous reconstruction-based detectors generally distinguish OOD samples by comparing pixel-level quality “degradation” of the reconstruction for given input [6, 39]. However, without prior-knowledge about OOD samples, there is no guarantee for such quality degradation. In contrast, MoodCat tries to synthesize In-D images instead of reconstructing the inputs, which is in line with the objective of the generative model.

The framework of MoodCat (Fig. 1(ii)) is inspired by [47], which detects adversarial examples (AE) by generating synthesized images conditioned on the output of the misled classifier. AE detection is quite different from OOD detection because adversarial examples are In-D samples with imperceptible perturbations. The classification label itself is sufficient to train the generative model to differentiate AEs and benign samples. This is not the case for OOD samples, which motivates the proposed MoodCat solution for OOD detection, as detailed in Section 3.

Recently, some researchers propose to involve data from other datasets to simulate OOD samples for model training. Representative “OOD-aware” techniques include Outlier Exposure (OE) [17], Maximum Classifier Discrepancy (MCD) [49], and Unsupervised Dual Grouping (UDG) [45]. OE relies on large-scale purified OOD samples, whereas MCD and UDG only need extra unlabeled data, which contains both In-D and OOD data. However, all of them are classification-based methods, where external data are used to train a modified classifier model. Following the same unlabeled extra data setting of MCD and UDG, we present that including extra training data (In-D, OOD mixture) into the training process can further improve MoodCat’s performance. Since our MoodCat works independently with the original classifiers, MoodCat will not degrade the accuracy of the original classifiers. We provide the detailed description in Section 3.6 and experimental results in Section 4.

A similar problem to OOD detection is the so-called Open Set Recognition (OSR) problem, which aims to distinguish the known and unknown classes[35, 46]. Several existing works [35, 11, 10, 32] targeted on OSR also employ generative models, whereas differ from MoodCat significantly. Specifically, OSRCI [32] uses Generative Adversarial Network (GAN) as data augmentation to train the classifier; C2AE [35] and CVAECapOSR [11] are conditional VAE/Autoencoder-based detectors. C2AE identifies outliers based on reconstruction errors, and requires K-time inference to give the final decision. In contrast, MoodCat infers once and makes the decision based on semantic contradiction. CVAECapOSR use Conditional VAE (CVAE) to model the distribution of In-D samples and detect outliers at latent space (i.e., without using the generator), whereas MoodCat detects at image space.

This work considers the scenario where we have an In-D data trained classifier ($\mathcal{C}$), which needs to be deployed in the wild. Consequently, the classifier will be threatened by OOD samples. We aim at building an OOD detection method, which can identify the OOD samples effectively without compromising the classification accuracy of the classifier. Our model is assumed to have access to the predicted label, $y$, but we do not modify any part of the classifier, including but may not limit to the architecture and the trained weights. As a result, our method can be a plug-and-play detector that easily cooperates with classifiers.

As pointed out by [46], OOD samples ($x_{o}\in\mathcal{O}$) are defined by label-shifted samples or samples with non-overlapping labels w.r.t the training data, or In-Ds ($x_{in}\in\mathcal{I}$). Hence, the semantics of any OOD sample contradicts with any In-D sample. This is the observation that motivates us to design a framework for OOD detection by spotlighting their semantic discrepancy.

Fig. 2 depicts the overview of our method. The proposed Masked Out-of-Distribution Catcher (MoodCat) contains three stages: randomly masking, generative synthesis and scoring. Specifically, we first randomly mask the input image $x$ as $x_{m}=\mathbf{M}(x)$, where $\mathbf{M}(\cdot)$ indicates the randomly masking operation. Then, we apply a generative model, $\mathbf{G}$, to synthesize a new image, $x^{\prime}$, by taking $x_{m}$ as the template and conditioning on the label $y$. Finally, we apply an anomalous scoring model to judge the discrepancy between the input and its synthesis.

Through masking, $x_{m}$ will partially lose its original semantic meaning, and thus leave more space for $\mathbf{G}$ to synthesize new content. With $y$ as the condition, the newly synthesized content should be consistent with the semantic meaning indicated by $y$. Note that, we use the ground truth label $y$ to train the generative model and use the output from the classifier when inference, i.e., $y=\mathcal{C}(x)$.

Here, we analyze different situations with In-D or OOD samples. On the one hand, if an In-D sample $x_{in}$ comes, the predicted label $y$ matches $x_{in}$’s intrinsic semantic meaning appropriately. Although the input image, $x_{in\cdot m}$, is partially masked, there should be some visual clue related to its semantic meaning, e.g., wings of a bird or paws of a dog. As a result, $\mathbf{G}$ can synthesize $x^{\prime}_{in}$ quite faithful to $x_{in}$. As exemplified in the upper half of Fig. 2, the synthesis of the “stop” sign can be very close to the original input. On the other hand, when it comes to an OOD sample $x_{o}$, the predicted label provided by the classifier is irreverent to $x_{o}$’s semantic meaning. Even with $x_{o}$ as a template, the generative model will try its best to synthesize contents related to the semantic label. As a result, the mismatch of semantic meaning between input and label can be spotlighted by the discrepancy between input and its synthesis. As the example shows in the lower part in Fig. 2, if an OOD sample (car) is wrongly predicted as a “stop” sign. The synthesis will be highly related to the “stop” sign rather than the original image.

Through such conditional synthesis, we can spotlight the discrepancy caused by OOD samples. Thus OODs can be easily distinguished by comparing the pair of input with its synthesis, $(x,x^{\prime})$.

In MoodCat, the generative model uses the input image as a template and synthesizes an image with the same semantic meaning as the given label. A high-quality synthesis can better highlight the contradiction. However, due to the intrinsic contradiction between the input image and the label for OOD, too much information from the input image can degrade the quality of generation. Therefore, we propose to apply masking on $x$ to remove some redundant information while leaving more space for the generative model to synthesize.

The use of masking follows the key motivation of MoodCat in OOD detection, which applies generation for synthesis rather than reconstruction. Previous reconstruction-based methods (e.g., [6, 39]) tend to reconstruct the image based on pixel-level dependency. In practice, the assumption that an OOD sample cannot be reconstructed well may not hold since they do not consider any semantics. However, our generative model aims at semantic synthesis. The masking mechanism can cooperate with the predicted label from the classifier to spotlight the contradiction caused by OOD.

The contribution of randomly masking is twofold: 1) masking the input image can encourage the generative model to better depict the semantic meaning of the given label on the synthesis, especially to an OOD sample; 2) masking, as a typical data augmentation method, can encourage the encoder to summarize the features of the input from a holistic perspective, thus improve the quality of synthesis, especially when synthesizing with In-Ds as templates. Obviously, the above two aspects both contribute to apart the behavior of In-D and OOD. As a result, a large discrepancy lies in the OOD sample and its synthesis.

The Generative model is responsible for generating a synthesis by taking both the masked input $x_{m}$ and the pre-assigned semantic label $y$ into consideration. As shown in Fig. 2, we select the Encoder ($\mathbf{E}$) and Decoder ($\mathbf{D}$) architecture as the generative model, i.e. $\mathbf{G}=\mathbf{E}\cdot\mathbf{D}$. This architecture is inspired by [47]. The encoder $\mathbf{E}$ acts as a feature extractor (as shown by the gray part in Fig. 2). By taking the masked image $x_{m}$ as input, $\mathbf{E}$ is expected to capture necessary low-level features and encoder them as a latent vector, $z=\mathbf{E}(x_{m})$. As done by VAE [21], we use the KL Divergence to regulate the latent vector $z$, which can be formulated as Eq. (1).

where $\mathcal{N}(\mu,\varSigma)$ indicates the Gaussian distribution with respect to $\mu$ and $\mathcal{\varSigma}$. We use the reparameterization trick from VAE on the latent variable $z$ during training, $z=\mu(x_{m})+\varSigma({x_{m}})\cdot\epsilon$, where $\epsilon\thicksim\mathcal{N}(0,1)$.

The decoder, $\mathbf{D}$, is trained to generate a synthesis $x^{\prime}=\mathbf{D}(z,y)$. The given semantic label $y$ is used to control the semantic meaning of the synthesis, while $z$ is used to provide low-level features from the template image $x$. This synthetic target is fulfilled through the class-conditional batch normalization layer [5]. This layer is usually used in conditional image generation [30, 52]. Since the normalization is determined by the given semantic label $y$, the semantic meaning of the synthesis can be highly dependent on it. As a result, if the semantic meaning of $x$ is consistent with $y$ (in the case of In-D samples), the synthesis can be highly close to $x$. However, if input an OOD sample, the semantic contradiction between the input image and the label will lead the synthesis to be far away from the input image, thus spotlighting the contradiction.

We implement $\mathbf{D}$ based on the generator architecture proposed in [2]. We apply the classic $\ell_{1}$, $\ell_{2}$ and $\mathcal{SSIM}$ [38] as part of loss items to constrain that $x^{\prime}$ resembles $x$. Furthermore, we adopt the U-net based discriminator [40] to operate an adversarial loss on the training process to further improve the quality of the synthetic image. Compared with the vanilla discriminator, this U-net based one can additionally provide a per-pixel real/fake map to locate the fake parts in the image. Therefore, the generative model can be trained to focus on both local and global features with more realistic details. Due to space limitations, we detailed the training process and corresponding objective functions of $\mathbf{G}$ in Appendix.

As analyzed in the former sections, MoodCat can generate high-quality syntheses in terms of similarity for In-D samples, but not for OOD samples. To distinguish OODs from In-Ds, we develop an anomalous scoring model. The proposed anomalous scoring model is built on two types of scorers: one is the conditional binary classifier, and the other is Image Quality Assessment models (IQA) [53]. Both can provide assessments of the syntheses by referring to input images.

Recently, researchers have shown that including OOD data into the training process can improve the performance of OOD detection [17, 49]. However, they may rely on well labeled data that need to manually identify In-D and OOD elements ¹¹ 1 Note that, images from another dataset are not necessarily to be OOD w.r.t semantic meaning [45]., e.g.,  [17]. On the contrary, unlabeled data can be easily collected from various sources with little cost. Actually, MoodCat can be further improved with unlabeled data.

The Conditional Binary Classifier introduced in Section 3.5 is trained by treating In-D data with synthesis of mismatched semantic labels as negative pairs. This procedure can be directly combined with external data. Note that, our negative pair only require that the semantic label used for synthesis is different from that of input image. To eliminate possible In-D samples from unlabeled data, we can apply a classifier trained with In-D data to generate pseudo label for unlabeled data. Then, during training, when we randomly sample the mismatched semantic label, we uniformly sample from all possible labels other than the pseudo label. In this way, the sampled semantic label is ensured to be mismatched for both OOD and some possible In-D data from the unlabeled data. Therefore, all of them can be utilized correctly to improve the performance of MoodCat.

Benchmarks. We evaluate MoodCat on the most recent semantic OOD detection benchmarks, SC-OOD benchmarks [45]. SC-OOD benchmarks provide extensive semantic-level OOD detection settings for evaluation. Specifically, from SC-OOD, images from different datasets are filtered to ensure that only those containing different semantic meanings are considered as OOD samples. SC-OOD focus on the semantic difference between samples, thus being more practical for the real-world model deployment than other previous OOD benchmarks[16, 27, 17], which are built by setting one dataset as In-D and all others as OOD.

We employ FPR@TPR95%, AUROC, AUPR, and Classification Accuracy as the evaluation metrics following [45, 29]. In this paper, unless otherwise specified, we denote the In-D as Positive (P), and the OOD as Negative (N).

We evaluate MoodCat with two settings: 1) MoodCat trained with In-D dataset only; 2) MoodCat trained with external unlabeled data (Section 3.6). We report the results in Table 7 and Table 8, respectively. Experiments are performed with ResNet18 [14] classifier²² 2 For more results under other classifier architectures (WRN28 [51], DenseNet [18]), please refer to Appendix. for fair comparison.

As discussed in Section 3.6, MoodCat can be equipped with external OOD data for better detection ability. In Table 7, we report the performance for MoodCat using Tiny-ImageNet as external unlabeled training data in gray. These results evidence the effectiveness of the training strategy with external data for MoodCat, where significant performance improvements can be found. Due to space limitation, for this setting, we only compare results with baselines on Cifar-100 (see below) and leave that on Cifar-10 benchmarks in Appendix.

Table 8 shows the results on Cifar-100 benchmarks. As can be seen, though taking advantage of external data, MCD, OE and UDG suffer from the complicated Cifar-100 dataset. By contrast, MoodCat’s performance on Cifar-100 is comparable with that on Cifar-10 against same OOD shown in Table 7. This is due to MoodCat’s detection mechanism, which relying on the semantic mismatch. Even when the In-D dataset becomes complicated, the semantic mismatch in OODs is still obvious. For methods detect OODs based on DNN-extracted features, e.g. MCD, OE, UDG, they may suffer from poor decision boundaries as the number of classes increases. Similarly as in the case on Cifar-10, to show the effectiveness of the proposed training strategy with external data, we also report the performance for MoodCat without external data in gray. The improvement can be seen on all metrics across all OOD settings. For space limitation, we leave the comparison on Cifar-100 benchmarks without external data in Appendix.

The experimental protocol in OSR is to randomly selecting $K$ classes from a specific $n$-class dataset as “known” classes and the left $n-K$ as “unknown” classes. To make a fair comparison, we retrain MoodCat under the experimental settings in CVAECapOSR [11], where only 4 or 6 classes from Cifar-10 are used for training. We report AUROC scores in Table 3, and the results for other methods are from CVAECapOSR [11]. As can be seen, MoodCat outperforms these methods in all three settings, especially in the original CIFAR setting, where MoodCat outperforms the second best method about 6%.

In this section, we first analyze the effectiveness of every anomalous scorer and how scorers cooperate to achieve the final decent performance. Then we conduct experiments on masking, and give insights on masking’s effectiveness.