We present a novel region based active learning method for semantic image segmentation, called MetaBox+. For acquisition, we train a meta regression model to estimate the segment-wise Intersection over Union (IoU) of each predicted segment of unlabeled images. This can be understood as an estimation of segment-wise prediction quality. Queried regions are supposed to minimize to competing targets, i.e., low predicted IoU values / segmentation quality and low estimated annotation costs. For estimating the latter we propose a simple but practical method for annotation cost estimation. We compare our method to entropy based methods, where we consider the entropy as uncertainty of the prediction. The comparison and analysis of the results provide insights into annotation costs as well as robustness and variance of the methods. Numerical experiments conducted with two different networks on the Cityscapes dataset clearly demonstrate a reduction of annotation effort compared to random acquisition. Noteworthily, we achieve 95%of the mean Intersection over Union (mIoU), using MetaBox+ compared to when training with the full dataset, with only 10.47% / 32.01% annotation effort for the two networks, respectively.
Deep neural networks (DNNs) have proven to be powerful tools for processing unstructured data. However for high-dimensional data, like images, they are inherently vulnerable to adversarial attacks. Small almost invisible perturbations added to the input can be used to fool DNNs. Various attacks, hardening methods and detection methods have been introduced in recent years. Notoriously, Carlini-Wagner (CW) type attacks computed by iterative minimization belong to those that are most difficult to detect. In this work, we demonstrate that such iterative minimization attacks can by used as detectors themselves. Thus, in some sense we show that one can fight fire with fire. This work also outlines a mathematical proof that under certain assumptions this detector provides asymptotically optimal separation of original and attacked images. In numerical experiments, we obtain AUROC values up to 99.73% for our detection method. This distinctly surpasses state of the art detection rates for CW attacks from the literature. We also give numerical evidence that our method is robust against the attacker's choice of the method of attack.
When deploying deep learning technology in self-driving cars, deep neural networks are constantly exposed to domain shifts. These include, e.g., changes in weather conditions, time of day, and long-term temporal shift. In this work we utilize a deep neural network trained on the Cityscapes dataset containing urban street scenes and infer images from a different dataset, the A2D2 dataset, containing also countryside and highway images. We present a novel pipeline for semantic segmenation that detects out-of-distribution (OOD) segments by means of the deep neural network's prediction and performs image retrieval after feature extraction and dimensionality reduction on image patches. In our experiments we demonstrate that the deployed OOD approach is suitable for detecting out-of-distribution concepts. Furthermore, we evaluate the image patch retrieval qualitatively as well as quantitatively by means of the semi-compatible A2D2 ground truth and obtain mAP values of up to 52.2%.
In semantic segmentation datasets, classes of high importance are oftentimes underrepresented, e.g., humans in street scenes. Neural networks are usually trained to reduce the overall number of errors, attaching identical loss to errors of all kinds. However, this is not necessarily aligned with human intuition. For instance, an overlooked pedestrian seems more severe than an incorrectly detected one. One possible remedy is to deploy different decision rules by introducing class priors which assigns larger weight to underrepresented classes. While reducing the false-negatives of the underrepresented class, at the same time this leads to a considerable increase of false-positive indications. In this work, we combine decision rules with methods for false-positive detection. We therefore fuse false-negative detection with uncertainty based false-positive meta classification. We present proof-of-concept results for CIFAR-10, and prove the efficiency of our method for the semantic segmentation of street scenes on the Cityscapes dataset based on predicted instances of the 'human' class. In the latter we employ an advanced false-positive detection method using uncertainty measures aggregated over instances. We thereby achieve improved trade-offs between false-negative and false-positive samples of the underrepresented classes.
In recent years, deep learning methods have outperformed other methods in image recognition. This has fostered imagination of potential application of deep learning technology including safety relevant applications like the interpretation of medical images or autonomous driving. The passage from assistance of a human decision maker to ever more automated systems however increases the need to properly handle the failure modes of deep learning modules. In this contribution, we review a set of techniques for the self-monitoring of machine-learning algorithms based on uncertainty quantification. In particular, we apply this to the task of semantic segmentation, where the machine learning algorithm decomposes an image according to semantic categories. We discuss false positive and false negative error modes at instance-level and review techniques for the detection of such errors that have been recently proposed by the authors. We also give an outlook on future research directions.
In the semantic segmentation of street scenes, the reliability of a prediction is of highest interest. The assessment of neural networks by means of uncertainties is a common ansatz to prevent safety issues. As in online applications like automated driving, a video stream of images is available, we present a time-dynamical approach to investigate uncertainties and assess the prediction quality of neural networks.To this end, we track segments over time and gather aggregated metrics per segment, e.g. mean dispersion metrics derived from the softmax output and segment sizes. Due to identifying segments over consecutive frames, we obtain time series of metrics from which we assess prediction quality. We do so by either classifying between intersection over union (IoU) = 0 and IoU > 0 (meta classification) or predicting the IoU directly (meta regression). In our tests, we analyze the influence of the length of the time series on the predictive power of metrics and study different models for meta classification and regression. We use two publicly available DeepLabv3+ networks as well as two street scene datasets, i.e., VIPER as a synthetic one and KITTI based on real data. We achieve classification accuracies of up to 81.20% and AUROC values of up to 88.68% for the task of meta classification. For meta regression we obtain $R^2$ values of up to 87.51%. We show that these results yield improvements compared to other approaches.
Neural networks for semantic segmentation can be seen as statistical models that provide for each pixel of one image a probability distribution on predefined classes. The predicted class is then usually obtained by the maximum a-posteriori probability (MAP) which is known as Bayes rule in decision theory. From decision theory we also know that the Bayes rule is optimal regarding the simple symmetric cost function. Therefore, it weights each type of confusion between two different classes equally, e.g., given images of urban street scenes there is no distinction in the cost function if the network confuses a person with a street or a building with a tree. Intuitively, there might be confusions of classes that are more important to avoid than others. In this work, we want to raise awareness of the possibility of explicitly defining confusion costs and the associated ethical difficulties if it comes down to providing numbers. We define two cost functions from different extreme perspectives, an egoistic and an altruistic one, and show how safety relevant quantities like precision / recall and (segment-wise) false positive / negative rate change when interpolating between MAP, egoistic and altruistic decision rules.
In the semantic segmentation of street scenes the reliability of the prediction and therefore uncertainty measures are of highest interest. We present a method that generates for each input image a hierarchy of nested crops around the image center and presents these, all re-scaled to the same size, to a neural network for semantic segmentation. The resulting softmax outputs are then post processed such that we can investigate mean and variance over all image crops as well as mean and variance of uncertainty heat maps obtained from pixel-wise uncertainty measures, like the entropy, applied to each crop's softmax output. In our tests, we use the publicly available DeepLabv3+ MobilenetV2 network (trained on the Cityscapes dataset) and demonstrate that the incorporation of crops improves the quality of the prediction and that we obtain more reliable uncertainty measures. These are then aggregated over predicted segments for either classifying between IoU=0 and IoU>0 (meta classification) or predicting the IoU via linear regression (meta regression). The latter yields reliable performance estimates for segmentation networks, in particular useful in the absence of ground truth. For the task of meta classification we obtain a classification accuracy of $81.93\%$ and an AUROC of $89.89\%$. For meta regression we obtain an $R^2$ value of $84.77\%$. These results yield significant improvements compared to other approaches.
As part of autonomous car driving systems, semantic segmentation is an essential component to obtain a full understanding of the car's environment. One difficulty, that occurs while training neural networks for this purpose, is class imbalance of training data. Consequently, a neural network trained on unbalanced data in combination with maximum a-posteriori classification may easily ignore classes that are rare in terms of their frequency in the dataset. However, these classes are often of highest interest. We approach such potential misclassifications by weighting the posterior class probabilities with the prior class probabilities which in our case are the inverse frequencies of the corresponding classes in the training dataset. More precisely, we adopt a localized method by computing the priors pixel-wise such that the impact can be analyzed at pixel level as well. In our experiments, we train one network from scratch using a proprietary dataset containing 20,000 annotated frames of video sequences recorded from street scenes. The evaluation on our test set shows an increase of average recall with regard to instances of pedestrians and info signs by $25\%$ and $23.4\%$, respectively. In addition, we significantly reduce the non-detection rate for instances of the same classes by $61\%$ and $38\%$.
We present a method that "meta" classifies whether segments (objects) predicted by a semantic segmentation neural network intersect with the ground truth. To this end, we employ measures of dispersion for predicted pixel-wise class probability distributions, like classification entropy, that yield heat maps of the input scene's size. We aggregate these dispersion measures segment-wise and derive metrics that are well-correlated with the segment-wise $\mathit{IoU}$ of prediction and ground truth. In our tests, we use two publicly available DeepLabv3+ networks (pre-trained on the Cityscapes data set) and analyze the predictive power of different metrics and different sets of metrics. To this end, we compute logistic LASSO regression fits for the task of classifying $\mathit{IoU}=0$ vs. $\mathit{IoU} > 0$ per segment and obtain classification rates of up to $81.91\%$ and AUROC values of up to $87.71\%$ without the incorporation of advanced techniques like Monte-Carlo dropout. We complement these tests with linear regression fits to predict the segment-wise $\mathit{IoU}$ and obtain prediction standard deviations of down to $0.130$ as well as $R^2$ values of up to $81.48\%$. We show that these results clearly outperform single-metric baseline approaches.