Tim Fingscheidt

Marvin Klingner, Varun Ravi Kumar, Senthil Yogamani et al.

While deep neural networks (DNNs) achieve impressive performance on environment perception tasks, their sensitivity to adversarial perturbations limits their use in practical applications. In this paper, we (i) propose a novel adversarial perturbation detection scheme based on multi-task perception of complex vision tasks (i.e., depth estimation and semantic segmentation). Specifically, adversarial perturbations are detected by inconsistencies between extracted edges of the input image, the depth output, and the segmentation output. To further improve this technique, we (ii) develop a novel edge consistency loss between all three modalities, thereby improving their initial consistency which in turn supports our detection scheme. We verify our detection scheme's effectiveness by employing various known attacks and image noises. In addition, we (iii) develop a multi-task adversarial attack, aiming at fooling both tasks as well as our detection scheme. Experimental evaluation on the Cityscapes and KITTI datasets shows that under an assumption of a 5% false positive rate up to 100% of images are correctly detected as adversarially perturbed, depending on the strength of the perturbation. Code is available at https://github.com/ifnspaml/AdvAttackDet. A short video at https://youtu.be/KKa6gOyWmH4 provides qualitative results.

Marvin Klingner, Mouadh Ayache, Tim Fingscheidt

Environment perception in autonomous driving vehicles often heavily relies on deep neural networks (DNNs), which are subject to domain shifts, leading to a significantly decreased performance during DNN deployment. Usually, this problem is addressed by unsupervised domain adaptation (UDA) approaches trained either simultaneously on source and target domain datasets or even source-free only on target data in an offline fashion. In this work, we further expand a source-free UDA approach to a continual and therefore online-capable UDA on a single-image basis for semantic segmentation. Accordingly, our method only requires the pre-trained model from the supplier (trained in the source domain) and the current (unlabeled target domain) camera image. Our method Continual BatchNorm Adaptation (CBNA) modifies the source domain statistics in the batch normalization layers, using target domain images in an unsupervised fashion, which yields consistent performance improvements during inference. Thereby, in contrast to existing works, our approach can be applied to improve a DNN continuously on a single-image basis during deployment without access to source data, without algorithmic delay, and nearly without computational overhead. We show the consistent effectiveness of our method across a wide variety of source/target domain settings for semantic segmentation. Code is available at https://github.com/ifnspaml/CBNA.

Manuel Schwonberg, Joshua Niemeijer, Jan-Aike Termöhlen et al.

Deep neural networks (DNNs) have proven their capabilities in many areas in the past years, such as robotics, or automated driving, enabling technological breakthroughs. DNNs play a significant role in environment perception for the challenging application of automated driving and are employed for tasks such as detection, semantic segmentation, and sensor fusion. Despite this progress and tremendous research efforts, several issues still need to be addressed that limit the applicability of DNNs in automated driving. The bad generalization of DNNs to new, unseen domains is a major problem on the way to a safe, large-scale application, because manual annotation of new domains is costly, particularly for semantic segmentation. For this reason, methods are required to adapt DNNs to new domains without labeling effort. The task, which these methods aim to solve is termed unsupervised domain adaptation (UDA). While several different domain shifts can challenge DNNs, the shift between synthetic and real data is of particular importance for automated driving, as it allows the use of simulation environments for DNN training. In this work, we present an overview of the current state of the art in this field of research. We categorize and explain the different approaches for UDA. The number of considered publications is larger than any other survey on this topic. The scope of this survey goes far beyond the description of the UDA state-of-the-art. Based on our large data and knowledge base, we present a quantitative comparison of the approaches and use the observations to point out the latest trends in this field. In the following, we conduct a critical analysis of the state-of-the-art and highlight promising future research directions. With this survey, we aim to facilitate UDA research further and encourage scientists to exploit novel research directions to generalize DNNs better.

Jasmin Breitenstein, Tim Fingscheidt

Amodal perception terms the ability of humans to imagine the entire shapes of occluded objects. This gives humans an advantage to keep track of everything that is going on, especially in crowded situations. Typical perception functions, however, lack amodal perception abilities and are therefore at a disadvantage in situations with occlusions. Complex urban driving scenarios often experience many different types of occlusions and, therefore, amodal perception for automated vehicles is an important task to investigate. In this paper, we consider the task of amodal semantic segmentation and propose a generic way to generate datasets to train amodal semantic segmentation methods. We use this approach to generate an amodal Cityscapes dataset. Moreover, we propose and evaluate a method as baseline on Amodal Cityscapes, showing its applicability for amodal semantic segmentation in automotive environment perception. We provide the means to re-generate this dataset on github.

Timo Lohrenz, Björn Möller, Zhengyang Li et al.

The powerful modeling capabilities of all-attention-based transformer architectures often cause overfitting and - for natural language processing tasks - lead to an implicitly learned internal language model in the autoregressive transformer decoder complicating the integration of external language models. In this paper, we explore relaxed attention, a simple and easy-to-implement smoothing of the attention weights, yielding a two-fold improvement to the general transformer architecture: First, relaxed attention provides regularization when applied to the self-attention layers in the encoder. Second, we show that it naturally supports the integration of an external language model as it suppresses the implicitly learned internal language model by relaxing the cross attention in the decoder. We demonstrate the benefit of relaxed attention across several tasks with clear improvement in combination with recent benchmark approaches. Specifically, we exceed the former state-of-the-art performance of 26.90% word error rate on the largest public lip-reading LRS3 benchmark with a word error rate of 26.31%, as well as we achieve a top-performing BLEU score of 37.67 on the IWSLT14 (DE$\rightarrow$EN) machine translation task without external language models and virtually no additional model parameters. Code and models will be made publicly available.

Jan-Aike Termöhlen, Timo Bartels, Tim Fingscheidt

The task of semantic segmentation requires a model to assign semantic labels to each pixel of an image. However, the performance of such models degrades when deployed in an unseen domain with different data distributions compared to the training domain. We present a new augmentation-driven approach to domain generalization for semantic segmentation using a re-parameterized vision transformer (ReVT) with weight averaging of multiple models after training. We evaluate our approach on several benchmark datasets and achieve state-of-the-art mIoU performance of 47.3% (prior art: 46.3%) for small models and of 50.1% (prior art: 47.8%) for midsized models on commonly used benchmark datasets. At the same time, our method requires fewer parameters and reaches a higher frame rate than the best prior art. It is also easy to implement and, unlike network ensembles, does not add any computational complexity during inference.

Marvin Klingner, Konstantin Müller, Mona Mirzaie et al.

The emergence of data-driven machine learning (ML) has facilitated significant progress in many complicated tasks such as highly-automated driving. While much effort is put into improving the ML models and learning algorithms in such applications, little focus is put into how the training data and/or validation setting should be designed. In this paper we investigate the influence of several data design choices regarding training and validation of deep driving models trainable in an end-to-end fashion. Specifically, (i) we investigate how the amount of training data influences the final driving performance, and which performance limitations are induced through currently used mechanisms to generate training data. (ii) Further, we show by correlation analysis, which validation design enables the driving performance measured during validation to generalize well to unknown test environments. (iii) Finally, we investigate the effect of random seeding and non-determinism, giving insights which reported improvements can be deemed significant. Our evaluations using the popular CARLA simulator provide recommendations regarding data generation and driving route selection for an efficient future development of end-to-end driving models.

Jasmin Breitenstein, Franz Jünger, Andreas Bär et al.

In this work, we study amodal video instance segmentation for automated driving. Previous works perform amodal video instance segmentation relying on methods trained on entirely labeled video data with techniques borrowed from standard video instance segmentation. Such amodally labeled video data is difficult and expensive to obtain and the resulting methods suffer from a trade-off between instance segmentation and tracking performance. To largely solve this issue, we propose to study the application of foundation models for this task. More precisely, we exploit the extensive knowledge of the Segment Anything Model (SAM), while fine-tuning it to the amodal instance segmentation task. Given an initial video instance segmentation, we sample points from the visible masks to prompt our amodal SAM. We use a point memory to store those points. If a previously observed instance is not predicted in a following frame, we retrieve its most recent points from the point memory and use a point tracking method to follow those points to the current frame, together with the corresponding last amodal instance mask. This way, while basing our method on an amodal instance segmentation, we nevertheless obtain video-level amodal instance segmentation results. Our resulting S-AModal method achieves state-of-the-art results in amodal video instance segmentation while resolving the need for amodal video-based labels. Code for S-AModal is available at https://github.com/ifnspaml/S-AModal.

Zhengyang Li, Thomas Graave, Björn Möller et al.

In audiovisual automatic speech recognition (AV-ASR) systems, information fusion of visual features in a pre-trained ASR has been proven as a promising method to improve noise robustness. In this work, based on the prominent Whisper ASR, first, we propose a simple and effective visual fusion method -- use of visual features both in encoder and decoder (dual-use) -- to learn the audiovisual interactions in the encoder and to weigh modalities in the decoder. Second, we compare visual fusion methods in Whisper models of various sizes. Our proposed dual-use method shows consistent noise robustness improvement, e.g., a 35% relative improvement (WER: 4.41% vs. 6.83%) based on Whisper small, and a 57% relative improvement (WER: 4.07% vs. 9.53%) based on Whisper medium, compared to typical reference middle fusion in babble noise with a signal-to-noise ratio (SNR) of 0dB. Third, we conduct ablation studies examining the impact of various module designs and fusion options. Fine-tuned on 1929 hours of audiovisual data, our dual-use method using Whisper medium achieves 4.08% (MUSAN babble noise) and 4.43% (NoiseX babble noise) average WER across various SNRs, thereby establishing a new state-of-the-art in noisy conditions on the LRS3 AV-ASR benchmark. Our code is at https://github.com/ifnspaml/Dual-Use-AVASR

Joshua Niemeijer, Manuel Schwonberg, Jan-Aike Termöhlen et al.

When models, e.g., for semantic segmentation, are applied to images that are vastly different from training data, the performance will drop significantly. Domain adaptation methods try to overcome this issue, but need samples from the target domain. However, this might not always be feasible for various reasons and therefore domain generalization methods are useful as they do not require any target data. We present a new diffusion-based domain extension (DIDEX) method and employ a diffusion model to generate a pseudo-target domain with diverse text prompts. In contrast to existing methods, this allows to control the style and content of the generated images and to introduce a high diversity. In a second step, we train a generalizing model by adapting towards this pseudo-target domain. We outperform previous approaches by a large margin across various datasets and architectures without using any real data. For the generalization from GTA5, we improve state-of-the-art mIoU performance by 3.8% absolute on average and for SYNTHIA by 11.8% absolute, marking a big step for the generalization performance on these benchmarks. Code is available at https://github.com/JNiemeijer/DIDEX

Björn Möller, Zhengyang Li, Malte Stelzer et al.

Recent successful video generation systems that predict and create realistic automotive driving scenes from short video inputs assign tokenization, future state prediction (world model), and video decoding to dedicated models. These approaches often utilize large models that require significant training resources, offer limited insight into design choices, and lack publicly available code and datasets. In this work, we address these deficiencies and present OpenViGA, an open video generation system for automotive driving scenes. Our contributions are: Unlike several earlier works for video generation, such as GAIA-1, we provide a deep analysis of the three components of our system by separate quantitative and qualitative evaluation: Image tokenizer, world model, video decoder. Second, we purely build upon powerful pre-trained open source models from various domains, which we fine-tune by publicly available automotive data (BDD100K) on GPU hardware at academic scale. Third, we build a coherent video generation system by streamlining interfaces of our components. Fourth, due to public availability of the underlying models and data, we allow full reproducibility. Finally, we also publish our code and models on Github. For an image size of 256x256 at 4 fps we are able to predict realistic driving scene videos frame-by-frame with only one frame of algorithmic latency.

Björn Möller, Lucas Görnhardt, Tim Fingscheidt

Transformer architectures prominently lead single-image super-resolution (SISR) benchmarks, reconstructing high-resolution (HR) images from their low-resolution (LR) counterparts. Their strong representative power, however, comes with a higher demand for training data compared to convolutional neural networks (CNNs). For many real-world SR applications, the availability of high-quality HR training images is not given, sparking interest in LR-only training methods. The LR-only SISR benchmark mimics this condition by allowing only low-resolution (LR) images for model training. For a 4x super-resolution, this effectively reduces the amount of available training data to 6.25% of the HR image pixels, which puts the employment of a data-hungry transformer model into question. In this work, we are the first to utilize a lightweight vision transformer model with LR-only training methods addressing the unsupervised SISR LR-only benchmark. We adopt and configure a recent LR-only training method from microscopy image super-resolution to macroscopic real-world data, resulting in our multi-scale training method for bicubic degradation (MSTbic). Furthermore, we compare it with reference methods and prove its effectiveness both for a transformer and a CNN model. We evaluate on the classic SR benchmark datasets Set5, Set14, BSD100, Urban100, and Manga109, and show superior performance over state-of-the-art (so far: CNN-based) LR-only SISR methods. The code is available on GitHub: https://github.com/ifnspaml/SuperResolutionMultiscaleTraining.

Ziyi Xu, Samy Elshamy, Ziyue Zhao et al.

Estimating time-frequency domain masks for single-channel speech enhancement using deep learning methods has recently become a popular research field with promising results. In this paper, we propose a novel components loss (CL) for the training of neural networks for mask-based speech enhancement. During the training process, the proposed CL offers separate control over preservation of the speech component quality, suppression of the residual noise component, and preservation of a naturally sounding residual noise component. We illustrate the potential of the proposed CL by evaluating a standard convolutional neural network (CNN) for mask-based speech enhancement. The new CL obtains a better and more balanced performance in almost all employed instrumental quality metrics over the baseline losses, the latter comprising the conventional mean squared error (MSE) loss and also auditory-related loss functions, such as the perceptual evaluation of speech quality (PESQ) loss and the recently proposed perceptual weighting filter loss. Particularly, applying the CL offers better speech component quality, better overall enhanced speech perceptual quality, as well as a more naturally sounding residual noise. On average, an at least 0.1 points higher PESQ score on the enhanced speech is obtained while also obtaining a higher SNR improvement by more than 0.5 dB, for seen noise types. This improvement is stronger for unseen noise types, where an about 0.2 points higher PESQ score on the enhanced speech is obtained, while also the output SNR is ahead by more than 0.5 dB. The new proposed CL is easy to implement and code is provided at https://github.com/ifnspaml/Components-Loss.

Danish Nazir, Antoine Hanna-Asaad, Lucas Görnhardt et al.

Existing multi-view three-dimensional (3D) object detection approaches widely adopt large-scale pre-trained vision transformer (ViT)-based foundation models as backbones, being computationally complex. To address this problem, current state-of-the-art (SOTA) \texttt{ToC3D} for efficient multi-view ViT-based 3D object detection employs ego-motion-based relevant token selection. However, there are two key limitations: (1) The fixed layer-individual token selection ratios limit computational efficiency during both training and inference. (2) Full end-to-end retraining of the ViT backbone is required for the multi-view 3D object detection method. In this work, we propose an image token compensator combined with a token selection for ViT backbones to accelerate multi-view 3D object detection. Unlike \texttt{ToC3D}, our approach enables dynamic layer-wise token selection within the ViT backbone. Furthermore, we introduce a parameter-efficient fine-tuning strategy, which trains only the proposed modules, thereby reducing the number of fine-tuned parameters from more than $300$ million (M) to only $1.6$ M. Experiments on the large-scale NuScenes dataset across three multi-view 3D object detection approaches demonstrate that our proposed method decreases computational complexity (GFLOPs) by $48\%$ ... $55\%$, inference latency (on an \texttt{NVIDIA-GV100} GPU) by $9\%$ ... $25\%$, while still improving mean average precision by $1.0\%$ ... $2.8\%$ absolute and NuScenes detection score by $0.4\%$ ... $1.2\%$ absolute compared to so-far SOTA \texttt{ToC3D}.

Danish Nazir, Timo Bartels, Jan Piewek et al.

Distributed computing in the context of deep neural networks (DNNs) implies the execution of one part of the network on edge devices and the other part typically on a large-scale cloud platform. Conventional methods propose to employ a serial concatenation of a learned image and source encoder, the latter projecting the image encoder output (bottleneck features) into a quantized representation for bitrate-efficient transmission. In the cloud, a respective source decoder reprojects the quantized representation to the original feature representation, serving as an input for the downstream task decoder performing, e.g., semantic segmentation. In this work, we propose joint source and task decoding, as it allows for a smaller network size in the cloud. This further enables the scalability of such services in large numbers without requiring extensive computational load on the cloud per channel. We demonstrate the effectiveness of our method by achieving a distributed semantic segmentation SOTA over a wide range of bitrates on the mean intersection over union metric, while using only $9.8 \%$ ... $11.59 \%$ of cloud DNN parameters used in the previous SOTA on the COCO and Cityscapes datasets.

Danish Nazir, Gowtham Sai Inti, Timo Bartels et al.

Modern automotive systems leverage deep neural networks (DNNs) for semantic segmentation and operate in two key application areas: (1) In-car, where the DNN solely operates in the vehicle without strict constraints on the data rate. (2) Distributed, where one DNN part operates in the vehicle and the other part typically on a large-scale cloud platform with a particular constraint on transmission bitrate efficiency. Typically, both applications share an image and source encoder, while each uses distinct (joint) source and task decoders. Prior work utilized convolutional neural networks for joint source and task decoding but did not investigate transformer-based alternatives such as SegDeformer, which offer superior performance at the cost of higher computational complexity. In this work, we propose joint feature and task decoding for SegDeformer, thereby enabling lower computational complexity in both in-car and distributed applications, despite SegDeformer's computational demands. This improves scalability in the cloud while reducing in-car computational complexity. For the in-car application, we increased the frames per second (fps) by up to a factor of $11.7$ ($1.4$ fps to $16.5$ fps) on Cityscapes and by up to a factor of $3.5$ ($43.3$ fps to $154.3$ fps) on ADE20K, while being on-par w.r.t.\ the mean intersection over union (mIoU) of the transformer-based baseline that doesn't compress by a source codec. For the distributed application, we achieve state-of-the-art (SOTA) over a wide range of bitrates on the mIoU metric, while using only $0.14$\% ($0.04$\%) of cloud DNN parameters used in previous SOTA, reported on ADE20K (Cityscapes).

Patrick Blumenberg, Thomas Graave, Tim Fingscheidt

Large language models (LLMs) demand extensive memory capacity during both fine-tuning and inference. To enable memory-efficient fine-tuning, existing methods apply block-wise quantization techniques, such as NF4 and AF4, to the network weights. We show that these quantization techniques incur suboptimal quantization errors. Therefore, as a first novelty, we propose an optimization approach for block-wise quantization. Using this method, we design a family of quantizers named 4-bit block-wise optimal float (BOF4), which consistently reduces the quantization error compared to both baseline methods. We provide both a theoretical and a data-driven solution for the optimization process and prove their practical equivalence. Secondly, we propose a modification to the employed normalization method based on the signed absolute block maximum (BOF4-S), enabling further reduction of the quantization error and empirically achieving less degradation in language modeling performance. Thirdly, we explore additional variations of block-wise quantization methods applied to LLMs through an experimental study on the importance of accurately representing zero and large-amplitude weights on the one hand, and optimization towards various error metrics on the other hand. Lastly, we introduce a mixed-precision quantization strategy dubbed outlier-preserving quantization (OPQ) to address the distributional mismatch induced by outlier weights in block-wise quantization. By storing outlier weights in 16-bit precision (OPQ) while applying BOF4-S, we achieve top performance among 4-bit block-wise quantization techniques w.r.t. perplexity.

Jonas Löhdefink, Tim Fingscheidt

In recent years, semantic segmentation has taken benefit from various works in computer vision. Inspired by the very versatile CycleGAN architecture, we combine semantic segmentation with the concept of cycle consistency to enable a multitask training protocol. However, learning is largely prevented by the so-called steganography effect, which expresses itself as watermarks in the latent segmentation domain, making image reconstruction a too easy task. To combat this, we propose a noise injection, based either on quantization noise or on Gaussian noise addition to avoid this disadvantageous information flow in the cycle architecture. We find that noise injection significantly reduces the generation of watermarks and thus allows the recognition of highly relevant classes such as "traffic signs", which are hardly detected by the ERFNet baseline. We report mIoU and PSNR results on the Cityscapes dataset for semantic segmentation and image reconstruction, respectively. The proposed methodology allows to achieve an mIoU improvement on the Cityscapes validation set of 5.7% absolute over the same CycleGAN without noise injection, and still an absolute 4.9% over the ERFNet non-cyclic baseline.

Bile Peng, Jan-Aike Termöhlen, Cong Sun et al.

Reconfigurable intelligent surface (RIS) is an emerging technology for future wireless communication systems. In this work, we consider downlink spatial multiplexing enabled by the RIS for weighted sum-rate (WSR) maximization. In the literature, most solutions use alternating gradient-based optimization, which has moderate performance, high complexity, and limited scalability. We propose to apply a fully convolutional network (FCN) to solve this problem, which was originally designed for semantic segmentation of images. The rectangular shape of the RIS and the spatial correlation of channels with adjacent RIS antennas due to the short distance between them encourage us to apply it for the RIS configuration. We design a set of channel features that includes both cascaded channels via the RIS and the direct channel. In the base station (BS), the differentiable minimum mean squared error (MMSE) precoder is used for pretraining and the weighted minimum mean squared error (WMMSE) precoder is then applied for fine-tuning, which is nondifferentiable, more complex, but achieves a better performance. Evaluation results show that the proposed solution has higher performance and allows for a faster evaluation than the baselines. Hence it scales better to a large number of antennas, advancing the RIS one step closer to practical deployment.

Ziyi Xu, Maximilian Strake, Tim Fingscheidt

Speech enhancement employing deep neural networks (DNNs) for denoising are called deep noise suppression (DNS). During training, DNS methods are typically trained with mean squared error (MSE) type loss functions, which do not guarantee good perceptual quality. Perceptual evaluation of speech quality (PESQ) is a widely used metric for evaluating speech quality. However, the original PESQ algorithm is non-differentiable, and therefore cannot directly be used as optimization criterion for gradient-based learning. In this work, we propose an end-to-end non-intrusive PESQNet DNN to estimate the PESQ scores of the enhanced speech signal. Thus, by providing a reference-free perceptual loss, it serves as a mediator towards the DNS training, allowing to maximize the PESQ score of the enhanced speech signal. We illustrate the potential of our proposed PESQNet-mediated training on the basis of an already strong baseline DNS. As further novelty, we propose to train the DNS and the PESQNet alternatingly to keep the PESQNet up-to-date and perform well specifically for the DNS under training. Our proposed method is compared to the same DNS trained with MSE-based loss for joint denoising and dereverberation, and the Interspeech 2021 DNS Challenge baseline. Detailed analysis shows that the PESQNet mediation can further increase the DNS performance by about 0.1 PESQ points on synthetic test data and by 0.03 DNSMOS points on real test data, compared to training with the MSE-based loss. Our proposed method also outperforms the Challenge baseline by 0.2 PESQ points on synthetic test data and 0.1 DNSMOS points on real test data.

Daniel Bogdoll, Jasmin Breitenstein, Florian Heidecker et al.

Scaling the distribution of automated vehicles requires handling various unexpected and possibly dangerous situations, termed corner cases (CC). Since many modules of automated driving systems are based on machine learning (ML), CC are an essential part of the data for their development. However, there is only a limited amount of CC data in large-scale data collections, which makes them challenging in the context of ML. With a better understanding of CC, offline applications, e.g., dataset analysis, and online methods, e.g., improved performance of automated driving systems, can be improved. While there are knowledge-based descriptions and taxonomies for CC, there is little research on machine-interpretable descriptions. In this extended abstract, we will give a brief overview of the challenges and goals of such a description.

Timo Lohrenz, Patrick Schwarz, Zhengyang Li et al.

Recently, attention-based encoder-decoder (AED) models have shown high performance for end-to-end automatic speech recognition (ASR) across several tasks. Addressing overconfidence in such models, in this paper we introduce the concept of relaxed attention, which is a simple gradual injection of a uniform distribution to the encoder-decoder attention weights during training that is easily implemented with two lines of code. We investigate the effect of relaxed attention across different AED model architectures and two prominent ASR tasks, Wall Street Journal (WSJ) and Librispeech. We found that transformers trained with relaxed attention outperform the standard baseline models consistently during decoding with external language models. On WSJ, we set a new benchmark for transformer-based end-to-end speech recognition with a word error rate of 3.65%, outperforming state of the art (4.20%) by 13.1% relative, while introducing only a single hyperparameter.

Sebastian Houben, Stephanie Abrecht, Maram Akila et al.

The use of deep neural networks (DNNs) in safety-critical applications like mobile health and autonomous driving is challenging due to numerous model-inherent shortcomings. These shortcomings are diverse and range from a lack of generalization over insufficient interpretability to problems with malicious inputs. Cyber-physical systems employing DNNs are therefore likely to suffer from safety concerns. In recent years, a zoo of state-of-the-art techniques aiming to address these safety concerns has emerged. This work provides a structured and broad overview of them. We first identify categories of insufficiencies to then describe research activities aiming at their detection, quantification, or mitigation. Our paper addresses both machine learning experts and safety engineers: The former ones might profit from the broad range of machine learning topics covered and discussions on limitations of recent methods. The latter ones might gain insights into the specifics of modern ML methods. We moreover hope that our contribution fuels discussions on desiderata for ML systems and strategies on how to propel existing approaches accordingly.

Marvin Klingner, Andreas Bär, Marcel Mross et al.

In this work we address the task of observing the performance of a semantic segmentation deep neural network (DNN) during online operation, i.e., during inference, which is of high importance in safety-critical applications such as autonomous driving. Here, many high-level decisions rely on such DNNs, which are usually evaluated offline, while their performance in online operation remains unknown. To solve this problem, we propose an improved online performance prediction scheme, building on a recently proposed concept of predicting the primary semantic segmentation task's performance. This can be achieved by evaluating the auxiliary task of monocular depth estimation with a measurement supplied by a LiDAR sensor and a subsequent regression to the semantic segmentation performance. In particular, we propose (i) sequential training methods for both tasks in a multi-task training setup, (ii) to share the encoder as well as parts of the decoder between both task's networks for improved efficiency, and (iii) a temporal statistics aggregation method, which significantly reduces the performance prediction error at the cost of a small algorithmic latency. Evaluation on the KITTI dataset shows that all three aspects improve the performance prediction compared to previous approaches.

Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani et al.

A 360° perception of scene geometry is essential for automated driving, notably for parking and urban driving scenarios. Typically, it is achieved using surround-view fisheye cameras, focusing on the near-field area around the vehicle. The majority of current depth estimation approaches focus on employing just a single camera, which cannot be straightforwardly generalized to multiple cameras. The depth estimation model must be tested on a variety of cameras equipped to millions of cars with varying camera geometries. Even within a single car, intrinsics vary due to manufacturing tolerances. Deep learning models are sensitive to these changes, and it is practically infeasible to train and test on each camera variant. As a result, we present novel camera-geometry adaptive multi-scale convolutions which utilize the camera parameters as a conditional input, enabling the model to generalize to previously unseen fisheye cameras. Additionally, we improve the distance estimation by pairwise and patchwise vector-based self-attention encoder networks. We evaluate our approach on the Fisheye WoodScape surround-view dataset, significantly improving over previous approaches. We also show a generalization of our approach across different camera viewing angles and perform extensive experiments to support our contributions. To enable comparison with other approaches, we evaluate the front camera data on the KITTI dataset (pinhole camera images) and achieve state-of-the-art performance among self-supervised monocular methods. An overview video with qualitative results is provided at https://youtu.be/bmX0UcU9wtA. Baseline code and dataset will be made public.

Timo Lohrenz, Zhengyang Li, Tim Fingscheidt

Stream fusion, also known as system combination, is a common technique in automatic speech recognition for traditional hybrid hidden Markov model approaches, yet mostly unexplored for modern deep neural network end-to-end model architectures. Here, we investigate various fusion techniques for the all-attention-based encoder-decoder architecture known as the transformer, striving to achieve optimal fusion by investigating different fusion levels in an example single-microphone setting with fusion of standard magnitude and phase features. We introduce a novel multi-encoder learning method that performs a weighted combination of two encoder-decoder multi-head attention outputs only during training. Employing then only the magnitude feature encoder in inference, we are able to show consistent improvement on Wall Street Journal (WSJ) with language model and on Librispeech, without increase in runtime or parameters. Combining two such multi-encoder trained models by a simple late fusion in inference, we achieve state-of-the-art performance for transformer-based models on WSJ with a significant WER reduction of 19% relative compared to the current benchmark approach.

Ernst Seidel, Jan Franzen, Maximilian Strake et al.

In recent years, deep neural networks (DNNs) were studied as an alternative to traditional acoustic echo cancellation (AEC) algorithms. The proposed models achieved remarkable performance for the separate tasks of AEC and residual echo suppression (RES). A promising network topology is a fully convolutional recurrent network (FCRN) structure, which has already proven its performance on both noise suppression and AEC tasks, individually. However, the combination of AEC, postfiltering, and noise suppression to a single network typically leads to a noticeable decline in the quality of the near-end speech component due to the lack of a separate loss for echo estimation. In this paper, we propose a two-stage model (Y$^2$-Net) which consists of two FCRNs, each with two inputs and one output (Y-Net). The first stage (AEC) yields an echo estimate, which - as a novelty for a DNN AEC model - is further used by the second stage to perform RES and noise suppression. While the subjective listening test of the Interspeech 2021 AEC Challenge mostly yielded results close to the baseline, the proposed method scored an average improvement of 0.46 points over the baseline on the blind testset in double-talk on the instrumental metric DECMOS, provided by the challenge organizers.

Florian Heidecker, Jasmin Breitenstein, Kevin Rösch et al.

Systems and functions that rely on machine learning (ML) are the basis of highly automated driving. An essential task of such ML models is to reliably detect and interpret unusual, new, and potentially dangerous situations. The detection of those situations, which we refer to as corner cases, is highly relevant for successfully developing, applying, and validating automotive perception functions in future vehicles where multiple sensor modalities will be used. A complication for the development of corner case detectors is the lack of consistent definitions, terms, and corner case descriptions, especially when taking into account various automotive sensors. In this work, we provide an application-driven view of corner cases in highly automated driving. To achieve this goal, we first consider existing definitions from the general outlier, novelty, anomaly, and out-of-distribution detection to show relations and differences to corner cases. Moreover, we extend an existing camera-focused systematization of corner cases by adding RADAR (radio detection and ranging) and LiDAR (light detection and ranging) sensors. For this, we describe an exemplary toolchain for data acquisition and processing, highlighting the interfaces of the corner case detection. We also define a novel level of corner cases, the method layer corner cases, which appear due to uncertainty inherent in the methodology or the data distribution.

Jasmin Breitenstein, Jan-Aike Termöhlen, Daniel Lipinski et al.

Automated driving has become a major topic of interest not only in the active research community but also in mainstream media reports. Visual perception of such intelligent vehicles has experienced large progress in the last decade thanks to advances in deep learning techniques but some challenges still remain. One such challenge is the detection of corner cases. They are unexpected and unknown situations that occur while driving. Conventional visual perception methods are often not able to detect them because corner cases have not been witnessed during training. Hence, their detection is highly safety-critical, and detection methods can be applied to vast amounts of collected data to select suitable training data. A reliable detection of corner cases will not only further automate the data selection procedure and increase safety in autonomous driving but can thereby also affect the public acceptance of the new technology in a positive manner. In this work, we continue a previous systematization of corner cases on different levels by an extended set of examples for each level. Moreover, we group detection approaches into different categories and link them with the corner case levels. Hence, we give directions to showcase specific corner cases and basic guidelines on how to technically detect them.

Andreas Bär, Jonas Löhdefink, Nikhil Kapoor et al.

Enabling autonomous driving (AD) can be considered one of the biggest challenges in today's technology. AD is a complex task accomplished by several functionalities, with environment perception being one of its core functions. Environment perception is usually performed by combining the semantic information captured by several sensors, i.e., lidar or camera. The semantic information from the respective sensor can be extracted by using convolutional neural networks (CNNs) for dense prediction. In the past, CNNs constantly showed state-of-the-art performance on several vision-related tasks, such as semantic segmentation of traffic scenes using nothing but the red-green-blue (RGB) images provided by a camera. Although CNNs obtain state-of-the-art performance on clean images, almost imperceptible changes to the input, referred to as adversarial perturbations, may lead to fatal deception. The goal of this article is to illuminate the vulnerability aspects of CNNs used for semantic segmentation with respect to adversarial attacks, and share insights into some of the existing known adversarial defense strategies. We aim to clarify the advantages and disadvantages associated with applying CNNs for environment perception in AD to serve as a motivation for future research in this field.

Nikhil Kapoor, Andreas Bär, Serin Varghese et al.

Despite recent advancements, deep neural networks are not robust against adversarial perturbations. Many of the proposed adversarial defense approaches use computationally expensive training mechanisms that do not scale to complex real-world tasks such as semantic segmentation, and offer only marginal improvements. In addition, fundamental questions on the nature of adversarial perturbations and their relation to the network architecture are largely understudied. In this work, we study the adversarial problem from a frequency domain perspective. More specifically, we analyze discrete Fourier transform (DFT) spectra of several adversarial images and report two major findings: First, there exists a strong connection between a model architecture and the nature of adversarial perturbations that can be observed and addressed in the frequency domain. Second, the observed frequency patterns are largely image- and attack-type independent, which is important for the practical impact of any defense making use of such patterns. Motivated by these findings, we additionally propose an adversarial defense method based on the well-known Wiener filters that captures and suppresses adversarial frequencies in a data-driven manner. Our proposed method not only generalizes across unseen attacks but also beats five existing state-of-the-art methods across two models in a variety of attack settings.

Nikhil Kapoor, Chun Yuan, Jonas Löhdefink et al.

Deep neural networks are often not robust to semantically-irrelevant changes in the input. In this work we address the issue of robustness of state-of-the-art deep convolutional neural networks (CNNs) against commonly occurring distortions in the input such as photometric changes, or the addition of blur and noise. These changes in the input are often accounted for during training in the form of data augmentation. We have two major contributions: First, we propose a new regularization loss called feature-map augmentation (FMA) loss which can be used during finetuning to make a model robust to several distortions in the input. Second, we propose a new combined augmentations (CA) finetuning strategy, that results in a single model that is robust to several augmentation types at the same time in a data-efficient manner. We use the CA strategy to improve an existing state-of-the-art method called stability training (ST). Using CA, on an image classification task with distorted images, we achieve an accuracy improvement of on average 8.94% with FMA and 8.86% with ST absolute on CIFAR-10 and 8.04% with FMA and 8.27% with ST absolute on ImageNet, compared to 1.98% and 2.12%, respectively, with the well known data augmentation method, while keeping the clean baseline performance.

Marvin Klingner, Jan-Aike Termöhlen, Jacob Ritterbach et al.

In this paper we present a solution to the task of "unsupervised domain adaptation (UDA) of a given pre-trained semantic segmentation model without relying on any source domain representations". Previous UDA approaches for semantic segmentation either employed simultaneous training of the model in the source and target domains, or they relied on an additional network, replaying source domain knowledge to the model during adaptation. In contrast, we present our novel Unsupervised BatchNorm Adaptation (UBNA) method, which adapts a given pre-trained model to an unseen target domain without using -- beyond the existing model parameters from pre-training -- any source domain representations (neither data, nor networks) and which can also be applied in an online setting or using just a few unlabeled images from the target domain in a few-shot manner. Specifically, we partially adapt the normalization layer statistics to the target domain using an exponentially decaying momentum factor, thereby mixing the statistics from both domains. By evaluation on standard UDA benchmarks for semantic segmentation we show that this is superior to a model without adaptation and to baseline approaches using statistics from the target domain only. Compared to standard UDA approaches we report a trade-off between performance and usage of source domain representations.

Atiye Sadat Hashemi, Andreas Bär, Saeed Mozaffari et al.

Deep neural networks tend to be vulnerable to adversarial perturbations, which by adding to a natural image can fool a respective model with high confidence. Recently, the existence of image-agnostic perturbations, also known as universal adversarial perturbations (UAPs), were discovered. However, existing UAPs still lack a sufficiently high fooling rate, when being applied to an unknown target model. In this paper, we propose a novel deep learning technique for generating more transferable UAPs. We utilize a perturbation generator and some given pretrained networks so-called source models to generate UAPs using the ImageNet dataset. Due to the similar feature representation of various model architectures in the first layer, we propose a loss formulation that focuses on the adversarial energy only in the respective first layer of the source models. This supports the transferability of our generated UAPs to any other target model. We further empirically analyze our generated UAPs and demonstrate that these perturbations generalize very well towards different target models. Surpassing the current state of the art in both, fooling rate and model-transferability, we can show the superiority of our proposed approach. Using our generated non-targeted UAPs, we obtain an average fooling rate of 93.36% on the source models (state of the art: 82.16%). Generating our UAPs on the deep ResNet-152, we obtain about a 12% absolute fooling rate advantage vs. cutting-edge methods on VGG-16 and VGG-19 target models.

Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani et al.

State-of-the-art self-supervised learning approaches for monocular depth estimation usually suffer from scale ambiguity. They do not generalize well when applied on distance estimation for complex projection models such as in fisheye and omnidirectional cameras. This paper introduces a novel multi-task learning strategy to improve self-supervised monocular distance estimation on fisheye and pinhole camera images. Our contribution to this work is threefold: Firstly, we introduce a novel distance estimation network architecture using a self-attention based encoder coupled with robust semantic feature guidance to the decoder that can be trained in a one-stage fashion. Secondly, we integrate a generalized robust loss function, which improves performance significantly while removing the need for hyperparameter tuning with the reprojection loss. Finally, we reduce the artifacts caused by dynamic objects violating static world assumptions using a semantic masking strategy. We significantly improve upon the RMSE of previous work on fisheye by 25% reduction in RMSE. As there is little work on fisheye cameras, we evaluated the proposed method on KITTI using a pinhole model. We achieved state-of-the-art performance among self-supervised methods without requiring an external scale estimation.

Antonia Breuer, Jan-Aike Termöhlen, Silviu Homoceanu et al.

Analyzing and predicting the traffic scene around the ego vehicle has been one of the key challenges in autonomous driving. Datasets including the trajectories of all road users present in a scene, as well as the underlying road topology are invaluable to analyze the behavior of the different traffic participants. The interaction between the various traffic participants is especially high in intersection types that are not regulated by traffic lights, the most common one being the roundabout. We introduce the openDD dataset, including 84,774 accurately tracked trajectories and HD map data of seven different roundabouts. The openDD dataset is annotated using images taken by a drone in 501 separate flights, totalling in over 62 hours of trajectory data. As of today, openDD is by far the largest publicly available trajectory dataset recorded from a drone perspective, while comparable datasets span 17 hours at most. The data is available, for both commercial and noncommercial use, at: http://www.l3pilot.eu/openDD.

Marvin Klingner, Jan-Aike Termöhlen, Jonas Mikolajczyk et al.

Self-supervised monocular depth estimation presents a powerful method to obtain 3D scene information from single camera images, which is trainable on arbitrary image sequences without requiring depth labels, e.g., from a LiDAR sensor. In this work we present a new self-supervised semantically-guided depth estimation (SGDepth) method to deal with moving dynamic-class (DC) objects, such as moving cars and pedestrians, which violate the static-world assumptions typically made during training of such models. Specifically, we propose (i) mutually beneficial cross-domain training of (supervised) semantic segmentation and self-supervised depth estimation with task-specific network heads, (ii) a semantic masking scheme providing guidance to prevent moving DC objects from contaminating the photometric loss, and (iii) a detection method for frames with non-moving DC objects, from which the depth of DC objects can be learned. We demonstrate the performance of our method on several benchmarks, in particular on the Eigen split, where we exceed all baselines without test-time refinement.

Jonas Löhdefink, Justin Fehrling, Marvin Klingner et al.

Autonomous driving requires self awareness of its perception functions. Technically spoken, this can be realized by observers, which monitor the performance indicators of various perception modules. In this work we choose, exemplarily, a semantic segmentation to be monitored, and propose an autoencoder, trained in a self-supervised fashion on the very same training data as the semantic segmentation to be monitored. While the autoencoder's image reconstruction performance (PSNR) during online inference shows already a good predictive power w.r.t. semantic segmentation performance, we propose a novel domain mismatch metric DM as the earth mover's distance between a pre-stored PSNR distribution on training (source) data, and an online-acquired PSNR distribution on any inference (target) data. We are able to show by experiments that the DM metric has a strong rank order correlation with the semantic segmentation within its functional scope. We also propose a training domain-dependent threshold for the DM metric to define this functional scope.

Marvin Klingner, Andreas Bär, Philipp Donn et al.

While neural networks trained for semantic segmentation are essential for perception in autonomous driving, most current algorithms assume a fixed number of classes, presenting a major limitation when developing new autonomous driving systems with the need of additional classes. In this paper we present a technique implementing class-incremental learning for semantic segmentation without using the labeled data the model was initially trained on. Previous approaches still either rely on labels for both old and new classes, or fail to properly distinguish between them. We show how to overcome these problems with a novel class-incremental learning technique, which nonetheless requires labels only for the new classes. Specifically, (i) we introduce a new loss function that neither relies on old data nor on old labels, (ii) we show how new classes can be integrated in a modular fashion into pretrained semantic segmentation models, and finally (iii) we re-implement previous approaches in a unified setting to compare them to ours. We evaluate our method on the Cityscapes dataset, where we exceed the mIoU performance of all baselines by 3.5% absolute reaching a result, which is only 2.2% absolute below the upper performance limit of single-stage training, relying on all data and labels simultaneously.

Marvin Klingner, Andreas Bär, Tim Fingscheidt

While current approaches for neural network training often aim at improving performance, less focus is put on training methods aiming at robustness towards varying noise conditions or directed attacks by adversarial examples. In this paper, we propose to improve robustness by a multi-task training, which extends supervised semantic segmentation by a self-supervised monocular depth estimation on unlabeled videos. This additional task is only performed during training to improve the semantic segmentation model's robustness at test time under several input perturbations. Moreover, we even find that our joint training approach also improves the performance of the model on the original (supervised) semantic segmentation task. Our evaluation exhibits a particular novelty in that it allows to mutually compare the effect of input noises and adversarial attacks on the robustness of the semantic segmentation. We show the effectiveness of our method on the Cityscapes dataset, where our multi-task training approach consistently outperforms the single-task semantic segmentation baseline in terms of both robustness vs. noise and in terms of adversarial attacks, without the need for depth labels in training.

Ziyue Zhao, Samy Elshamy, Tim Fingscheidt

Single-channel speech enhancement with deep neural networks (DNNs) has shown promising performance and is thus intensively being studied. In this paper, instead of applying the mean squared error (MSE) as the loss function during DNN training for speech enhancement, we design a perceptual weighting filter loss motivated by the weighting filter as it is employed in analysis-by-synthesis speech coding, e.g., in code-excited linear prediction (CELP). The experimental results show that the proposed simple loss function improves the speech enhancement performance compared to a reference DNN with MSE loss in terms of perceptual quality and noise attenuation. The proposed loss function can be advantageously applied to an existing DNN-based speech enhancement system, without modification of the DNN topology for speech enhancement. The source code for the proposed approach is made available.

Jan-Aike Bolte, Andreas Bär, Daniel Lipinski et al.

The progress in autonomous driving is also due to the increased availability of vast amounts of training data for the underlying machine learning approaches. Machine learning systems are generally known to lack robustness, e.g., if the training data did rarely or not at all cover critical situations. The challenging task of corner case detection in video, which is also somehow related to unusual event or anomaly detection, aims at detecting these unusual situations, which could become critical, and to communicate this to the autonomous driving system (online use case). Such a system, however, could be also used in offline mode to screen vast amounts of data and select only the relevant situations for storing and (re)training machine learning algorithms. So far, the approaches for corner case detection have been limited to videos recorded from a fixed camera, mostly for security surveillance. In this paper, we provide a formal definition of a corner case and propose a system framework for both the online and the offline use case that can handle video signals from front cameras of a naturally moving vehicle and can output a corner case score.

Jonas Löhdefink, Andreas Bär, Nico M. Schmidt et al.

The high amount of sensors required for autonomous driving poses enormous challenges on the capacity of automotive bus systems. There is a need to understand tradeoffs between bitrate and perception performance. In this paper, we compare the image compression standards JPEG, JPEG2000, and WebP to a modern encoder/decoder image compression approach based on generative adversarial networks (GANs). We evaluate both the pure compression performance using typical metrics such as peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and others, but also the performance of a subsequent perception function, namely a semantic segmentation (characterized by the mean intersection over union (mIoU) measure). Not surprisingly, for all investigated compression methods, a higher bitrate means better results in all investigated quality metrics. Interestingly, however, we show that the semantic segmentation mIoU of the GAN autoencoder in the highly relevant low-bitrate regime (at 0.0625 bit/pixel) is better by 3.9% absolute than JPEG2000, although the latter still is considerably better in terms of PSNR (5.91 dB difference). This effect can greatly be enlarged by training the semantic segmentation model with images originating from the decoder, so that the mIoU using the segmentation model trained by GAN reconstructions exceeds the use of the model trained with original images by almost 20% absolute. We conclude that distributed perception in future autonomous driving will most probably not provide a solution to the automotive bus capacity bottleneck by using standard compression schemes such as JPEG2000, but requires modern coding approaches, with the GAN encoder/decoder method being a promising candidate.

Ziyue Zhao, Huijun Liu, Tim Fingscheidt

Enhancing coded speech suffering from far-end acoustic background noise, quantization noise, and potentially transmission errors, is a challenging task. In this work we propose two postprocessing approaches applying convolutional neural networks (CNNs) either in the time domain or the cepstral domain to enhance the coded speech without any modification of the codecs. The time domain approach follows an end-to-end fashion, while the cepstral domain approach uses analysis-synthesis with cepstral domain features. The proposed postprocessors in both domains are evaluated for various narrowband and wideband speech codecs in a wide range of conditions. The proposed postprocessor improves speech quality (PESQ) by up to 0.25 MOS-LQO points for G.711, 0.30 points for G.726, 0.82 points for G.722, and 0.26 points for adaptive multirate wideband codec (AMR-WB). In a subjective CCR listening test, the proposed postprocessor on G.711-coded speech exceeds the speech quality of an ITU-T-standardized postfilter by 0.36 CMOS points, and obtains a clear preference of 1.77 CMOS points compared to legacy G.711, even better than uncoded speech with statistical significance. The source code for the cepstral domain approach to enhance G.711-coded speech is made available.