Detecting Out-of-distribution Objects
Using Neuron Activation Patterns

Heyndrycks et al. [15] proposed a baseline OOD samples detection method. It only involves setting a threshold on the winning class softmax probability (maximum softmax probability, MSP). In object detection, it is already necessary to set a softmax probability threshold for non-maximum-suppression (NMS) to filter redundant background predictions. Therefore, setting another one to separate OOD objects can not be particularly useful in practice. Nevertheless, it still provides a decent baseline.

Initially, Helmholtz Energy [21] had been successfully utilized in OOD detection in image classification. It is a very easy-to-use and universal way to estimate semantic uncertainty. It does not require any detector customization or loss handcrafting. Internally, object detectors perform the classification of proposals resulting in a classification vector for which one can compute the energy score.

Virtual outlier synthesis (VOS) [7] is the first work focusing on OOD identification in the object detection task. VOS allows OOD detection by synthesizing virtual outliers in feature (latent) space, thereby regularizing the model’s decision boundary during training. VOS samples the virtual outliers from the low-likelihood region in the feature space and uses them as input for an unknown-aware training objective. The contrastive loss function shapes the uncertainty (Helmholtz energy) space between known data and synthesized outlier data. VOS is not architecture-agnostic because the outlier synthesis process occurs in feature space bound to the fully-connected layers of Faster R-CNN ROI head. The authors of VOS proposed an evaluation scheme that requires an OOD dataset i.e. a set of images that do not contain ID categories. Any output bounding box generated for these images is considered an "OOD object", while any bounding box generated for images from the ID test dataset is deemed an "ID object". We find this approach is over-simplified because model predictions are not evaluated, so any "OOD object" or "ID object" can actually be a background prediction.

Gaussian mixture models (GMM) [23] approach also introduces a change in the default loss function, focusing on classification loss. The authors add an anchor loss term to facilitate learning a structured logit space. Next, they fit class-specific GMMs to the logit space with a validation dataset. For any test sample, uncertainty is estimated as a log-likelihood of belonging to any one of the known GMMs.

The authors of OpenDet [11] follow the intuition that known-class objects tend to be clustered to form high-density regions in the latent space, while unknown objects are distributed in low-density regions. Consequently, they propose identifying unknown objects by separating high- and low-density regions in the latent space using a contrastive loss. On top of this, they provide another loss function component responsible for learning predictive uncertainty directly as a softmax probability without the logit of ground-truth class. The authors of OpenDet provide the implementation for both Faster R-CNN and RetinaNet. However, the method does not apply to anchor-free architectures because IOU-based sampling of proposals for contrastive learning is not feasible.

The OWOD [17] approach was developed before OpenDet. Both ideas share the part of contrastive clustering in latent space. The authors of OWOD also propose an unknown-aware RPN head and the Helmholtz energy-based unknown identification. They model the energy distribution of the known and unknown energy values with a set of shifted Weibull distributions. Similarly to GMM, they fit the distributions using a validation dataset. Due to the RPN requirement, Faster R-CNN is the only compatible architecture.

Previous work by Choi et. al. [5] introduced Gaussian YOLOv3. The authors redesign YOLO’s loss function to model localization uncertainty directly. Each coordinate of a bounding box is represented by a pair of Gaussian parameters (mean and variance) instead of just one value. There is no ground truth for the variance of coordinates, but Gaussian negative-log-likelihood loss conveniently requires ground truth for the mean only, which allows one to learn variance without providing the ground truth. We managed to extend this approach to Faster R-CNN and RetinaNet architectures. In FCOS, however, a "localization centerness" branch is designed to perform a very similar role as Gaussian localization uncertainty; trying to combine them both led to poor results.

The popular Monte-Carlo Dropout (MCD) [9] method has already been adapted to the object detection task [22, 24]. The most straightforward adaptation technique was proposed in ref. [24], in which the authors suggested averaging both classification and regression vectors over the number of Monte-Carlo forward passes for each proposal separately. The variance of these vectors for each proposal is the final uncertainty measure.

Faster R-CNN-based Object Localization Network (OLN) [18] learns to detect objects focusing on localization instead of foreground-background classification. The authors argued that focusing on learning "objectness cues", for the price of obtaining class-agnostic outputs is the best way to achieve cross-dataset generalization. In practice, their method requires replacing the default classification losses in both Faster R-CNN stages with a pair of centerness and IOU losses.

As an alternative baseline focused on objectness for Faster R-CNN, we attempt to establish a threshold for the RPN objectness score of a bounding box proposal. This score is later converted into the final model output. Although the RPN objectness score is not used in the ROI head, we aim to determine its potential usefulness for the OOD detection tasks.

NAPs have been previously used for uncertainty estimation in image classifiers [4, 25]. Image processing neural networks (NNs) use multiple layers that transform input samples sequentially into the desired form. Nonlinear activation functions present in each hidden layer of a NN significantly contribute to the network’s ability to approximate complex, multidimensional relationships. Of all activation functions, ReLU is nowadays the most common choice for CV problems. All ReLU-activated layers produce matrices of either positive or zeroed values, which correspond respectively to active and inactive layers’ units. An activation pattern is obtained by assigning a true to positive units and false to zeros. Thus, for a given network layer, NAP is a binary interpretation of which neurons or convolution units of the layer were activated during the processing of an image.

Algorithm 1 describes how NAPTRON estimates the uncertainty of predicted bounding boxes. First, for every training image, we perform an object detection inference, simultaneously extracting binary NAPs corresponding to output bounding boxes. The patterns are extracted from a pre-selected layer. Then, for every true positive prediction, we store the extracted pattern in a memory structure. Each object class has a dedicated memory instance. For a given test sample we perform the inference and NAP extraction again. Next, for each inferred bounding box, we find the Hamming distance between its NAP and the nearest pattern out of those stored in the memory structure corresponding to the predicted label. The Hamming distance to the nearest known NAP is the NAPTRON uncertainty estimate.

The authors of [25] provided an efficient implementation of finding the minimal Hamming distance to known NAPs, which makes real-time uncertainty estimation possible. In this implementation, the Hamming distances between the test NAP and all known NAPs are computed concurrently. The minimal distance is the output of the uncertainty estimation algorithm.

In every state-of-the-art object detector, one can distinguish a detection backbone part (e.g. ResNet50) that extracts features of an input image and a detection head that performs classification and regression of proposal bounding boxes (priors) based on the extracted features. Typically, the classification and regression are computed in two separate subnetworks that work in parallel. The extractNAPS function extracts binary NAPs of the $l$-th ReLU-activated layer of the classification branch in the detection head of object detector $\theta$. For Faster R-CNN, the ROI-head consists of fully connected (FC) layers. The operation is straightforward because there is no ambiguity in the link of hidden layers’ activations to the final output. Assuming $P$ proposals generated for a single image by the RPN-head, each layer processes $[P,In]$-shaped feature maps into $[P,Out]$-shaped feature maps, where $In$ is the number of neurons of the previous layer and $Out$ is the number of neurons of the current layer. Thus, for each layer and each proposed bounding box, we can easily distinguish an activation pattern of length $Out$.

Nevertheless, detection heads of single-stage architectures such as RetinaNet and FCOS, use convolutional layers instead of FC layers. Each hidden layer takes as an input $[W,H,C]$-shaped feature maps and processes them without changing their dimensions. The final layer changes the number of channels of the matrix from $C$ to $K*A$, where $W$ and $H$ are the width and height of processed feature maps, respectively, $K$ is the number of classes, and $A$ is the number of priors with centers in each of $W*H$ locations of the feature maps. In our approach, all the $A$ possible bounding boxes predicted in the $(w,h)$ location are associated with a $C$-long activation vector located at $(w,h)$ of the chosen hidden layer’s output. In other words, any $(w,h)$-centered output bounding box is attributed to the activation values located in the same coordinates of the hidden feature maps.

Results in Tables 1, 2, and 3 suggest that tuning training samples’ IOU and softmax threshold parameters is not worthwhile. Maintaining the standard 0.5 IOU threshold, the lowest (0.0) softmax threshold - thus letting as many training samples as possible be stored in the known patterns database - allows an obtaining of the optimal results. Moreover, shifts in these parameters do not introduce significant performance variation. As for the binarization percentile threshold, abstaining from binarization seems to be the best choice. We find that the higher the threshold, the worse the results we obtained.

On the other hand, introducing the averaging Hamming distance (instead of the previously proposed finding of the distance to the closest binary pattern) brings a slight improvement - especially if Layer 0 is chosen. The choice of the layer is the most important parameter to tune. We recommend extracting patterns from the penultimate layer of the detection head.

In the next experiment, we compared the performance of the proposed algorithm with the state-of-the-art OOD object detectors:

• •

  confidence score from Faster R-CNN (Standard)

• •

  objectness score from Region Proposal Network in Faster R-CNN (RPN)

• •

  Energy [21]

• •

  Virtual Outlier Synthesis (VOS) [7]

• •

  Gaussian Mixture Models (GMM) [23]

• •

  OpenDet [11]

• •

  Open World Object Detection (OWOD) [17]

• •

  Gaussian YOLOv3 (Gasussian) [5]

• •

  Monte-Carlo Dropout (MCD) [24]

• •

  Object Localization Network (OLN) [18]

All methods were compared using the officially published code of each algorithm with recommended configuration parameters, except the MCD method, which we implemented ourselves. It is important to notice that some of the methods were only designed for certain DNN architectures (e.g., Faster R-CNN); and we applied the methods only for the dedicated models.

Additionally, the results showed that the open-sets methods (OpenDet, GMM, and OWOD) are not the most effective OOD detectors. They fail to outperform the baseline detector (Standard) consistently. GMM manages to do so in 3 or 4 (depending on the metric) out of 6 cases, OpenDet in 1 out of 4, and OWOD in 1 or 2 out of 2.

The rest of the methods perform very poorly in our challenging experimental setup. Gaussian, OLN, RPN, and MCD are unable to beat the baseline detector (Standard).

All the figures confirm that the quality of all the OOD methods is sensitive to underlying object detectors’ NMS threshold variations. However, the ranking of OOD detectors does not change much for different NMS values since most of the methods’ curves share a common trend. Typically, performance metrics reach their upper limit around 0.5 and decline sharply when the softmax probability threshold is between 0.8 and 0.99. These sharp shifts of metrics’ levels for high thresholds occur because object detectors hardly ever predict OOD objects with such a high probability, so OOD metrics are being computed for scarce data. This issue does not matter much in practice because, in real-world applications of object detectors, the NMS threshold is usually set somewhere between 0.5 and 0.8 - thus filtering most of the FPs and letting through most of the TPs.

Many of the state-of-the-art OOD detection methods require altering default components of models to obtain improved uncertainty scores. It is expected that these alterations should not impose any damage on regular object detection quality. Therefore, we performed additional experiments to evaluate the influence of the customization of standard object detectors. In other words, how modification of the original architecture affects the performance of the object detector.

Object detector performance heavily depends on a predefined, application-dependent confidence score threshold. The authors of [10] observed that decreasing the NMS confidence score threshold and consequently massively increasing the number of false positive detections increases mAP. This phenomenon undermines the validity of using mAP as the primary quality metric. Therefore, as a way to gauge the impact of the OOD methods on object detection quality, we conduct a visual collation of TPR vs FP curves (see Figures 4 and 5). All the plots were horizontally limited to 100 000 FPs.

Drawing conclusions about the performance can hardly be completed merely by looking at the curves, so we compute the area under each curve (AUC), limiting the FP number to $2N$ where $N$ is the number of all known ground truth objects in the test dataset. Next, we compare each method’s AUC with the standard detector’s AUC - Table 5 shows the results. Positive $\Delta$AUC values signify a higher (better) curve, while negative $\Delta$AUC indicates inferior performance.

OLN is class-agnostic (unlike all other models) and, for that reason, is subject to a more relaxed evaluation regime. The class-agnostic true positive prediction required to the correct localization of an object, without the need for accurate label assignment.

Figures 4 and 5 and Table 5 show that altering the default configurations of object detectors usually harms their performance. Except for the FCOS PASCAL-VOC $\rightarrow$ COCO that provides relatively poor results, standard detectors are the most effective. Despite more relaxed rules, OLN fails to reach the performance levels of other detectors. GMM’s anchor loss proves to be detrimental, especially for RetinaNet architecture. Gaussian negative log-likelihood loss bears a negative effect on object detectors too. For a fixed number of FPs, Gaussian Faster R-CNN generates hundreds of TPs less than the standard Faster R-CNN. RetinaNet OpenDet obtains significantly worse precision in the PASCAL-VOC $\rightarrow$ COCO scenario, but in the remaining scenarios works very well. The AUC of VOS, OWOD, and MC-Dropout detectors is, on average, lower than the baseline, but the differences are smaller than for other methods.

Ideally, an OOD detector provides an independent uncertainty score, yet complementary to the objects detector softmax score. By design, the softmax score is solely responsible for distinguishing background from foreground objects, and the uncertainty score separates known objects from unknown ones.

The relationship between both the scores is nontrivial and simple statistic coefficients, such as the Pearson correlation coefficient, do not explain it sufficiently. Therefore, we provide visual 2D characteristics for the best-performing OOD detectors. A perfect characteristic would depict a cloud of blue points (OOD objects) vertically separable from a cloud of green points (TPs), and a cloud of red points (FPs) horizontally separable from the TPs. Figure 6 shows the NAPTRON uncertainty and softmax score relationship. We can observe that the blue OOD triangles are placed, in general, higher than the green TPs, especially for the most certain samples on the right side of the plot.

Overconfident OOD and FPs are the users’ nightmare but our approach enables us to filter at least some of them. Examples of the samples identified as OOD by the proposed NAPTRON are in Figures 7 and 8. For all the examined algorithms, both scores are imperfect. In many cases, FPs have a high softmax score, whereas TPs have a lower score. Analogously, a low uncertainty score may be attributed to an OOD object. We observe that every type of uncertainty score is correlated with the softmax confidence score of an underlying object detector. This outcome is intuitive because we expect the TPs to have low uncertainty and high softmax probability. Thankfully, object detectors tend to assign lower probabilities to OOD objects, and OOD detectors assign higher uncertainty to FPs even though they are not explicitly meant to do so. These circumstances provide an opportunity to use softmax probability and one or more types of uncertainty scores combined in a manner that would boost both regular object detection and OOD detection performance. Finding an optimal way to combine multiple scores would require another set of experiments; this is beyond the scope of this work.