Michael Paulitsch

Syed Qutub, Florian Geissler, Yang Peng et al.

Object detection neural network models need to perform reliably in highly dynamic and safety-critical environments like automated driving or robotics. Therefore, it is paramount to verify the robustness of the detection under unexpected hardware faults like soft errors that can impact a systems perception module. Standard metrics based on average precision produce model vulnerability estimates at the object level rather than at an image level. As we show in this paper, this does not provide an intuitive or representative indicator of the safety-related impact of silent data corruption caused by bit flips in the underlying memory but can lead to an over- or underestimation of typical fault-induced hazards. With an eye towards safety-related real-time applications, we propose a new metric IVMOD (Image-wise Vulnerability Metric for Object Detection) to quantify vulnerability based on an incorrect image-wise object detection due to false positive (FPs) or false negative (FNs) objects, combined with a severity analysis. The evaluation of several representative object detection models shows that even a single bit flip can lead to a severe silent data corruption event with potentially critical safety implications, with e.g., up to (much greater than) 100 FPs generated, or up to approx. 90% of true positives (TPs) are lost in an image. Furthermore, with a single stuck-at-1 fault, an entire sequence of images can be affected, causing temporally persistent ghost detections that can be mistaken for actual objects (covering up to approx. 83% of the image). Furthermore, actual objects in the scene are continuously missed (up to approx. 64% of TPs are lost). Our work establishes a detailed understanding of the safety-related vulnerability of such critical workloads against hardware faults.

Neslihan Kose, Ranganath Krishnan, Akash Dhamasia et al.

Reliable uncertainty quantification in deep neural networks is very crucial in safety-critical applications such as automated driving for trustworthy and informed decision-making. Assessing the quality of uncertainty estimates is challenging as ground truth for uncertainty estimates is not available. Ideally, in a well-calibrated model, uncertainty estimates should perfectly correlate with model error. We propose a novel error aligned uncertainty optimization method and introduce a trainable loss function to guide the models to yield good quality uncertainty estimates aligning with the model error. Our approach targets continuous structured prediction and regression tasks, and is evaluated on multiple datasets including a large-scale vehicle motion prediction task involving real-world distributional shifts. We demonstrate that our method improves average displacement error by 1.69% and 4.69%, and the uncertainty correlation with model error by 17.22% and 19.13% as quantified by Pearson correlation coefficient on two state-of-the-art baselines.

Ralf Graafe, Qutub Syed Sha, Florian Geissler et al.

Transient or permanent faults in hardware can render the output of Neural Networks (NN) incorrect without user-specific traces of the error, i.e. silent data errors (SDE). On the other hand, modern NNs also possess an inherent redundancy that can tolerate specific faults. To establish a safety case, it is necessary to distinguish and quantify both types of corruptions. To study the effects of hardware (HW) faults on software (SW) in general and NN models in particular, several fault injection (FI) methods have been established in recent years. Current FI methods focus on the methodology of injecting faults but often fall short of accounting for large-scale FI tests, where many fault locations based on a particular fault model need to be analyzed in a short time. Results need to be concise, repeatable, and comparable. To address these requirements and enable fault injection as the default component in a machine learning development cycle, we introduce a novel fault injection framework called PyTorchALFI (Application Level Fault Injection for PyTorch) based on PyTorchFI. PyTorchALFI provides an efficient way to define randomly generated and reusable sets of faults to inject into PyTorch models, defines complex test scenarios, enhances data sets, and generates test KPIs while tightly coupling fault-free, faulty, and modified NN. In this paper, we provide details about the definition of test scenarios, software architecture, and several examples of how to use the new framework to apply iterative changes in fault location and number, compare different model modifications, and analyze test results.

Syed Sha Qutub, Neslihan Kose, Rafael Rosales et al.

This paper introduces the Budding Ensemble Architecture (BEA), a novel reduced ensemble architecture for anchor-based object detection models. Object detection models are crucial in vision-based tasks, particularly in autonomous systems. They should provide precise bounding box detections while also calibrating their predicted confidence scores, leading to higher-quality uncertainty estimates. However, current models may make erroneous decisions due to false positives receiving high scores or true positives being discarded due to low scores. BEA aims to address these issues. The proposed loss functions in BEA improve the confidence score calibration and lower the uncertainty error, which results in a better distinction of true and false positives and, eventually, higher accuracy of the object detection models. Both Base-YOLOv3 and SSD models were enhanced using the BEA method and its proposed loss functions. The BEA on Base-YOLOv3 trained on the KITTI dataset results in a 6% and 3.7% increase in mAP and AP50, respectively. Utilizing a well-balanced uncertainty estimation threshold to discard samples in real-time even leads to a 9.6% higher AP50 than its base model. This is attributed to a 40% increase in the area under the AP50-based retention curve used to measure the quality of calibration of confidence scores. Furthermore, BEA-YOLOV3 trained on KITTI provides superior out-of-distribution detection on Citypersons, BDD100K, and COCO datasets compared to the ensembles and vanilla models of YOLOv3 and Gaussian-YOLOv3.

Rafael Rosales, Pablo Munoz, Michael Paulitsch

In this paper, we investigate the relationship between diversity metrics, accuracy, and resiliency to natural image corruptions of Deep Learning (DL) image classifier ensembles. We investigate the potential of an attribution-based diversity metric to improve the known accuracy-diversity trade-off of the typical prediction-based diversity. Our motivation is based on analytical studies of design diversity that have shown that a reduction of common failure modes is possible if diversity of design choices is achieved. Using ResNet50 as a comparison baseline, we evaluate the resiliency of multiple individual DL model architectures against dataset distribution shifts corresponding to natural image corruptions. We compare ensembles created with diverse model architectures trained either independently or through a Neural Architecture Search technique and evaluate the correlation of prediction-based and attribution-based diversity to the final ensemble accuracy. We evaluate a set of diversity enforcement heuristics based on negative correlation learning to assess the final ensemble resilience to natural image corruptions and inspect the resulting prediction, activation, and attribution diversity. Our key observations are: 1) model architecture is more important for resiliency than model size or model accuracy, 2) attribution-based diversity is less negatively correlated to the ensemble accuracy than prediction-based diversity, 3) a balanced loss function of individual and ensemble accuracy creates more resilient ensembles for image natural corruptions, 4) architecture diversity produces more diversity in all explored diversity metrics: predictions, attributions, and activations.

Rafael Rosales, Peter Popov, Michael Paulitsch

Ensembling is a successful technique to improve the performance of machine learning (ML) models. Conf-Ensemble is an adaptation to Boosting to create ensembles based on model confidence instead of model errors to better classify difficult edge-cases. The key idea is to create successive model experts for samples that were difficult (not necessarily incorrectly classified) by the preceding model. This technique has been shown to provide better results than boosting in binary-classification with a small feature space (~80 features). In this paper, we evaluate the Conf-Ensemble approach in the much more complex task of image classification with the ImageNet dataset (224x224x3 features with 1000 classes). Image classification is an important benchmark for AI-based perception and thus it helps to assess if this method can be used in safety-critical applications using ML ensembles. Our experiments indicate that in a complex multi-label classification task, the expected benefit of specialization on complex input samples cannot be achieved with a small sample set, i.e., a good classifier seems to rely on very complex feature analysis that cannot be well trained on just a limited subset of "difficult samples". We propose an improvement to Conf-Ensemble to increase the number of samples fed to successive ensemble members, and a three-member Conf-Ensemble using this improvement was able to surpass a single model in accuracy, although the amount is not significant. Our findings shed light on the limits of the approach and the non-triviality of harnessing big data.

Florian Geissler, Syed Qutub, Michael Paulitsch et al.

We present a highly compact run-time monitoring approach for deep computer vision networks that extracts selected knowledge from only a few (down to merely two) hidden layers, yet can efficiently detect silent data corruption originating from both hardware memory and input faults. Building on the insight that critical faults typically manifest as peak or bulk shifts in the activation distribution of the affected network layers, we use strategically placed quantile markers to make accurate estimates about the anomaly of the current inference as a whole. Importantly, the detector component itself is kept algorithmically transparent to render the categorization of regular and abnormal behavior interpretable to a human. Our technique achieves up to ~96% precision and ~98% recall of detection. Compared to state-of-the-art anomaly detection techniques, this approach requires minimal compute overhead (as little as 0.3% with respect to non-supervised inference time) and contributes to the explainability of the model.

Kai Yuan, Christoph Bauinger, Xiangyi Zhang et al.

This paper presents a SYCL implementation of Multi-Layer Perceptrons (MLPs), which targets and is optimized for the Intel Data Center GPU Max 1550. To increase the performance, our implementation minimizes the slow global memory accesses by maximizing the data reuse within the general register file and the shared local memory by fusing the operations in each layer of the MLP. We show with a simple roofline model that this results in a significant increase in the arithmetic intensity, leading to improved performance, especially for inference. We compare our approach to a similar CUDA implementation for MLPs and show that our implementation on the Intel Data Center GPU outperforms the CUDA implementation on Nvidia's H100 GPU by a factor up to 2.84 in inference and 1.75 in training. The paper also showcases the efficiency of our SYCL implementation in three significant areas: Image Compression, Neural Radiance Fields, and Physics-Informed Machine Learning. In all cases, our implementation outperforms the off-the-shelf Intel Extension for PyTorch (IPEX) implementation on the same Intel GPU by up to a factor of 30 and the CUDA PyTorch version on Nvidia's H100 GPU by up to a factor 19. The code can be found at https://github.com/intel/tiny-dpcpp-nn.

Nina Wiedemann, Sainan Liu, Quentin Leboutet et al.

Following rapid advancements in text and image generation, research has increasingly shifted towards 3D generation. Unlike the well-established pixel-based representation in images, 3D representations remain diverse and fragmented, encompassing a wide variety of approaches such as voxel grids, neural radiance fields, signed distance functions, point clouds, or octrees, each offering distinct advantages and limitations. In this work, we present a unified evaluation framework designed to assess the performance of 3D representations in reconstruction and generation. We compare these representations based on multiple criteria: quality, computational efficiency, and generalization performance. Beyond standard model benchmarking, our experiments aim to derive best practices over all steps involved in the 3D generation pipeline, including preprocessing, mesh reconstruction, compression with autoencoders, and generation. Our findings highlight that reconstruction errors significantly impact overall performance, underscoring the need to evaluate generation and reconstruction jointly. We provide insights that can inform the selection of suitable 3D models for various applications, facilitating the development of more robust and application-specific solutions in 3D generation. The code for our framework is available at https://github.com/isl-org/unifi3d.

Zhipeng Cai, Matthias Mueller, Reiner Birkl et al.

In the current era of generative AI breakthroughs, generating panoramic scenes from a single input image remains a key challenge. Most existing methods use diffusion-based iterative or simultaneous multi-view inpainting. However, the lack of global scene layout priors leads to subpar outputs with duplicated objects (e.g., multiple beds in a bedroom) or requires time-consuming human text inputs for each view. We propose L-MAGIC, a novel method leveraging large language models for guidance while diffusing multiple coherent views of 360 degree panoramic scenes. L-MAGIC harnesses pre-trained diffusion and language models without fine-tuning, ensuring zero-shot performance. The output quality is further enhanced by super-resolution and multi-view fusion techniques. Extensive experiments demonstrate that the resulting panoramic scenes feature better scene layouts and perspective view rendering quality compared to related works, with >70% preference in human evaluations. Combined with conditional diffusion models, L-MAGIC can accept various input modalities, including but not limited to text, depth maps, sketches, and colored scripts. Applying depth estimation further enables 3D point cloud generation and dynamic scene exploration with fluid camera motion. Code is available at https://github.com/IntelLabs/MMPano. The video presentation is available at https://youtu.be/XDMNEzH4-Ec?list=PLG9Zyvu7iBa0-a7ccNLO8LjcVRAoMn57s.

Gabriela Ben Melech Stan, Diana Wofk, Estelle Aflalo et al.

Latent diffusion models have proven to be state-of-the-art in the creation and manipulation of visual outputs. However, as far as we know, the generation of depth maps jointly with RGB is still limited. We introduce LDM3D-VR, a suite of diffusion models targeting virtual reality development that includes LDM3D-pano and LDM3D-SR. These models enable the generation of panoramic RGBD based on textual prompts and the upscaling of low-resolution inputs to high-resolution RGBD, respectively. Our models are fine-tuned from existing pretrained models on datasets containing panoramic/high-resolution RGB images, depth maps and captions. Both models are evaluated in comparison to existing related methods.

Nina Wiedemann, Quentin Leboutet, Michael Paulitsch et al.

Optimizing GPU kernels presents a significantly greater challenge for large language models (LLMs) than standard code generation tasks, as it requires understanding hardware architecture, parallel optimization strategies, and performance profiling outputs. Most existing LLM-based approaches to kernel generation rely on simple prompting and feedback loops, incorporating hardware awareness only indirectly through profiling feedback. We introduce KernelFoundry, an evolutionary framework that efficiently explores the GPU kernel design space through three key mechanisms: (1) MAP-Elites quality-diversity search with kernel-specific behavioral dimensions to sustain exploration across diverse optimization strategies; (2) meta-prompt evolution, which co-evolves prompts with kernels to uncover task-specific optimization strategies, and (3) template-based parameter optimization to tune kernels to inputs and hardware. We evaluate this framework on KernelBench, robust-kbench, and custom tasks, generating SYCL kernels as a cross-platform GPU programming model and CUDA kernels for comparison to prior work. Our approach consistently outperforms the baseline methods, achieving an average speedup of 2.3x on KernelBench for SYCL. Moreover, KernelFoundry is implemented as a distributed framework with remote access to diverse hardware, enabling rapid benchmarking and featuring a flexible user input layer that supports kernel generation for a wide range of real-world use cases beyond benchmarking.

Yujin Chen, Yinyu Nie, Benjamin Ummenhofer et al.

We present Mesh2NeRF, an approach to derive ground-truth radiance fields from textured meshes for 3D generation tasks. Many 3D generative approaches represent 3D scenes as radiance fields for training. Their ground-truth radiance fields are usually fitted from multi-view renderings from a large-scale synthetic 3D dataset, which often results in artifacts due to occlusions or under-fitting issues. In Mesh2NeRF, we propose an analytic solution to directly obtain ground-truth radiance fields from 3D meshes, characterizing the density field with an occupancy function featuring a defined surface thickness, and determining view-dependent color through a reflection function considering both the mesh and environment lighting. Mesh2NeRF extracts accurate radiance fields which provides direct supervision for training generative NeRFs and single scene representation. We validate the effectiveness of Mesh2NeRF across various tasks, achieving a noteworthy 3.12dB improvement in PSNR for view synthesis in single scene representation on the ABO dataset, a 0.69 PSNR enhancement in the single-view conditional generation of ShapeNet Cars, and notably improved mesh extraction from NeRF in the unconditional generation of Objaverse Mugs.

Junda Cheng, Zhipeng Cai, Zhaoxing Zhang et al.

Visual odometry (VO) aims to estimate camera poses from visual inputs -- a fundamental building block for many applications such as VR/AR and robotics. This work focuses on monocular RGB VO where the input is a monocular RGB video without IMU or 3D sensors. Existing approaches lack robustness under this challenging scenario and fail to generalize to unseen data (especially outdoors); they also cannot recover metric-scale poses. We propose Robust Metric Visual Odometry (RoMeO), a novel method that resolves these issues leveraging priors from pre-trained depth models. RoMeO incorporates both monocular metric depth and multi-view stereo (MVS) models to recover metric-scale, simplify correspondence search, provide better initialization and regularize optimization. Effective strategies are proposed to inject noise during training and adaptively filter noisy depth priors, which ensure the robustness of RoMeO on in-the-wild data. As shown in Fig.1, RoMeO advances the state-of-the-art (SOTA) by a large margin across 6 diverse datasets covering both indoor and outdoor scenes. Compared to the current SOTA DPVO, RoMeO reduces the relative (align the trajectory scale with GT) and absolute trajectory errors both by >50%. The performance gain also transfers to the full SLAM pipeline (with global BA & loop closure). Code will be released upon acceptance.

Yujin Chen, Yinyu Nie, Benjamin Ummenhofer et al.

We present PBR-SR, a novel method for physically based rendering (PBR) texture super resolution (SR). It outputs high-resolution, high-quality PBR textures from low-resolution (LR) PBR input in a zero-shot manner. PBR-SR leverages an off-the-shelf super-resolution model trained on natural images, and iteratively minimizes the deviations between super-resolution priors and differentiable renderings. These enhancements are then back-projected into the PBR map space in a differentiable manner to produce refined, high-resolution textures. To mitigate view inconsistencies and lighting sensitivity, which is common in view-based super-resolution, our method applies 2D prior constraints across multi-view renderings, iteratively refining the shared, upscaled textures. In parallel, we incorporate identity constraints directly in the PBR texture domain to ensure the upscaled textures remain faithful to the LR input. PBR-SR operates without any additional training or data requirements, relying entirely on pretrained image priors. We demonstrate that our approach produces high-fidelity PBR textures for both artist-designed and AI-generated meshes, outperforming both direct SR models application and prior texture optimization methods. Our results show high-quality outputs in both PBR and rendering evaluations, supporting advanced applications such as relighting.

Zijun Li, Zhipeng Cai, Bochun Yang et al.

Visual localization is a fundamental machine learning problem. Absolute Pose Regression (APR) trains a scene-dependent model to efficiently map an input image to the camera pose in a pre-defined scene. However, many applications have continually changing environments, where inference data at novel poses or scene conditions (weather, geometry) appear after deployment. Training APR on a fixed dataset leads to overfitting, making it fail catastrophically on challenging novel data. This work proposes Continual Domain Expansion (ConDo), which continually collects unlabeled inference data to update the deployed APR. Instead of applying standard unsupervised domain adaptation methods which are ineffective for APR, ConDo effectively learns from unlabeled data by distilling knowledge from scene-agnostic localization methods. By sampling data uniformly from historical and newly collected data, ConDo can effectively expand the generalization domain of APR. Large-scale benchmarks with various scene types are constructed to evaluate models under practical (long-term) data changes. ConDo consistently and significantly outperforms baselines across architectures, scene types, and data changes. On challenging scenes (Fig.1), it reduces the localization error by >7x (14.8m vs 1.7m). Analysis shows the robustness of ConDo against compute budgets, replay buffer sizes and teacher prediction noise. Comparing to model re-training, ConDo achieves similar performance up to 25x faster.

Hongchi Xia, Chih-Hao Lin, Hao-Yu Hsu et al.

Digitizing the physical world into accurate simulation-ready virtual environments offers significant opportunities in a variety of fields such as augmented and virtual reality, gaming, and robotics. However, current 3D reconstruction and scene-understanding methods commonly fall short in one or more critical aspects, such as geometry completeness, object interactivity, physical plausibility, photorealistic rendering, or realistic physical properties for reliable dynamic simulation. To address these limitations, we introduce HoloScene, a novel interactive 3D reconstruction framework that simultaneously achieves these requirements. HoloScene leverages a comprehensive interactive scene-graph representation, encoding object geometry, appearance, and physical properties alongside hierarchical and inter-object relationships. Reconstruction is formulated as an energy-based optimization problem, integrating observational data, physical constraints, and generative priors into a unified, coherent objective. Optimization is efficiently performed via a hybrid approach combining sampling-based exploration with gradient-based refinement. The resulting digital twins exhibit complete and precise geometry, physical stability, and realistic rendering from novel viewpoints. Evaluations conducted on multiple benchmark datasets demonstrate superior performance, while practical use-cases in interactive gaming and real-time digital-twin manipulation illustrate HoloScene's broad applicability and effectiveness. Project page: https://xiahongchi.github.io/HoloScene.

Qutub Syed Sha, Michael Paulitsch, Karthik Pattabiraman et al.

As transformer-based object detection models progress, their impact in critical sectors like autonomous vehicles and aviation is expected to grow. Soft errors causing bit flips during inference have significantly impacted DNN performance, altering predictions. Traditional range restriction solutions for CNNs fall short for transformers. This study introduces the Global Clipper and Global Hybrid Clipper, effective mitigation strategies specifically designed for transformer-based models. It significantly enhances their resilience to soft errors and reduces faulty inferences to ~ 0\%. We also detail extensive testing across over 64 scenarios involving two transformer models (DINO-DETR and Lite-DETR) and two CNN models (YOLOv3 and SSD) using three datasets, totalling approximately 3.3 million inferences, to assess model robustness comprehensively. Moreover, the paper explores unique aspects of attention blocks in transformers and their operational differences from CNNs.

Qutub Syed, Michael Paulitsch, Korbinian Hagn et al.

We introduce Situation Monitor, a novel zero-shot Out-of-Distribution (OOD) detection approach for transformer-based object detection models to enhance reliability in safety-critical machine learning applications such as autonomous driving. The Situation Monitor utilizes the Diversity-based Budding Ensemble Architecture (DBEA) and increases the OOD performance by integrating a diversity loss into the training process on top of the budding ensemble architecture, detecting Far-OOD samples and minimizing false positives on Near-OOD samples. Moreover, utilizing the resulting DBEA increases the model's OOD performance and improves the calibration of confidence scores, particularly concerning the intersection over union of the detected objects. The DBEA model achieves these advancements with a 14% reduction in trainable parameters compared to the vanilla model. This signifies a substantial improvement in efficiency without compromising the model's ability to detect OOD instances and calibrate the confidence scores accurately.

Cornelius Buerkle, Florian Geissler, Michael Paulitsch et al.

While the most visible part of the safety verification process of automated vehicles concerns the planning and control system, it is often overlooked that safety of the latter crucially depends on the fault-tolerance of the preceding environment perception. Modern perception systems feature complex and often machine-learning-based components with various failure modes that can jeopardize the overall safety. At the same time, a verification by for example redundant execution is not always feasible due to resource constraints. In this paper, we address the need for feasible and efficient perception monitors and propose a lightweight approach that helps to protect the integrity of the perception system while keeping the additional compute overhead minimal. In contrast to existing solutions, the monitor is realized by a well-balanced combination of sensor checks -- here using LiDAR information -- and plausibility checks on the object motion history. It is designed to detect relevant errors in the distance and velocity of objects in the environment of the automated vehicle. In conjunction with an appropriate planning system, such a monitor can help to make safe automated driving feasible.

Florian Geissler, Syed Qutub, Sayanta Roychowdhury et al.

Convolutional neural networks (CNNs) have become an established part of numerous safety-critical computer vision applications, including human robot interactions and automated driving. Real-world implementations will need to guarantee their robustness against hardware soft errors corrupting the underlying platform memory. Based on the previously observed efficacy of activation clipping techniques, we build a prototypical safety case for classifier CNNs by demonstrating that range supervision represents a highly reliable fault detector and mitigator with respect to relevant bit flips, adopting an eight-exponent floating point data representation. We further explore novel, non-uniform range restriction methods that effectively suppress the probability of silent data corruptions and uncorrectable errors. As a safety-relevant end-to-end use case, we showcase the benefit of our approach in a vehicle classification scenario, using ResNet-50 and the traffic camera data set MIOVision. The quantitative evidence provided in this work can be leveraged to inspire further and possibly more complex CNN safety arguments.