3D scene understanding is a critical yet challenging task in autonomous driving due to the irregularity and sparsity of LiDAR data, as well as the computational demands of processing large-scale point clouds. Recent methods leverage range-view representations to enhance efficiency, but they often adopt higher azimuth resolutions to mitigate information loss during spherical projection, where only the closest point is retained for each 2D grid. However, processing wide panoramic range-view images remains inefficient and may introduce additional distortions. Our empirical analysis shows that training with multiple range images, obtained from splitting the full point cloud, improves both segmentation accuracy and computational efficiency. However, this approach also poses new challenges of exacerbated class imbalance and increase in projection artifacts. To address these, we introduce FLARES, a novel training paradigm that incorporates two tailored data augmentation techniques and a specialized post-processing method designed for multi-range settings. Extensive experiments demonstrate that FLARES is highly generalizable across different architectures, yielding 2.1%–7.9% mIoU improvements on SemanticKITTI and 1.8%–3.9% mIoU on nuScenes, while delivering over 40% speed-up in inference.1

Vision Transformers (ViTs) have emerged as the state-of-the-art models in various Computer Vision (CV) tasks, but their high computational and resource demands pose significant challenges. While Mixture-of-Experts (MoE) can make these models more efficient, they often require costly retraining or even training from scratch. Recent developments aim to reduce these computational costs by leveraging pretrained networks. These have been shown to produce sparse activation patterns in the Multi-Layer Perceptrons (MLPs) of the encoder blocks, allowing for conditional activation of only relevant subnetworks for each sample.Building on this idea, we propose a new method to construct MoE variants from pretrained models. Our approach extracts expert subnetworks from the model’s MLP layers post-training in two phases. First, we cluster output activations to identify distinct activation patterns. In the second phase, we use these clusters to extract the corresponding subnetworks responsible for producing them. On ImageNet-1k recognition tasks, we demonstrate that these extracted experts can perform surprisingly well out of the box and require only minimal fine-tuning to regain 98% of the original performance, all while reducing MACs and model size, by up to 36% and 32% respectively.

Convolutional neural networks (CNNs) are crucial for computer vision tasks on resource-constrained devices. Quantization effectively compresses these models, reducing storage size and energy cost. However, in modern depthwise-separable architectures, the computational cost is distributed unevenly across its components, with pointwise operations being the most expensive. By applying a general quantization scheme to this imbalanced cost distribution, existing quantization approaches fail to fully exploit potential efficiency gains. To this end, we introduce PROM, a straightforward approach for quantizing modern depthwise-separable convolutional networks by selectively using two distinct bit-widths. Specifically, pointwise convolutions are quantized to ternary weights, while the remaining modules use 8-bit weights, which is achieved through a simple quantization-aware training procedure. Additionally, by quantizing activations to 8-bit, our method transforms pointwise convolutions with ternary weights into int8 additions, which enjoy broad support across hardware platforms and effectively eliminates the need for expensive multiplications. Applying PROM to MobileNetV2 reduces the model’s energy cost by more than an order of magnitude (23.9×) and its storage size by 2.7× compared to the float16 baseline while retaining similar classification performance on ImageNet. Our method advances the Pareto frontier for energy consumption vs. top-1 accuracy for quantized convolutional models on ImageNet. PROM addresses the challenges of quantizing depthwise-separable convolutional networks to both ternary and 8-bit weights, offering a simple way to reduce energy cost and storage size.

Increasingly expensive training of ever larger models such as Vision Transfomers motivate reusing the vast library of already trained state-of-the-art networks. However, their latency, high computational costs and memory demands pose significant challenges for deployment, especially on resource-constrained hardware. While structured pruning methods can reduce these factors, they often require costly retraining, sometimes for up to hundreds of epochs, or even training from scratch to recover the lost accuracy resulting from the structural modifications. Maintaining the provided performance of trained models after structured pruning and thereby avoiding extensive retraining remains a challenge. To solve this, we introduce Variance-Based Pruning, a simple and structured one-shot pruning technique for efficiently compressing networks, with minimal finetuning. Our approach first gathers activation statistics, which are used to select neurons for pruning. Simultaneously the mean activations are integrated back into the model to preserve a high degree of performance. On ImageNet-1k recognition tasks, we demonstrate that directly after pruning DeiT-Base retains over 70% of its original performance and requires only 10 epochs of fine-tuning to regain 99% of the original accuracy while simultaneously reducing MACs by 35% and model size by 36%, thus speeding up the model by 1.44x. The code is available at: https://github.com/boschresearch/variance-based-pruning

We present Frame-Averaging Kernel-Point Convolution (FA-KPConv), a neural network architecture built on top of the well-known KPConv, a widely adopted backbone for 3D point cloud analysis. Even though invariance and/or equivariance to Euclidean transformations are required for many common tasks, KPConv-based networks can only approximately achieve such properties when training on large datasets or with significant data augmentations. Using Frame Averaging, we allow to flexibly customize point cloud neural networks built with KPConv layers, by making them exactly invariant and/or equivariant to translations, rotations and/or reflections of the input point clouds. By simply wrapping around an existing KPConv-based network, FA-KPConv embeds geometrical prior knowledge into it while preserving the number of learnable parameters and not compromising any input information. We showcase the benefit of such an introduced bias for point cloud classification and point cloud registration, especially in challenging cases such as scarce training data or randomly rotated test data. We plan to open-source our implementation upon publication.

In automated driving, predicting trajectories of surrounding vehicles supports reasoning about scene dynamics and enables safe planning for the ego vehicle. However, existing models handle predictions as an instantaneous task of forecasting future trajectories based on observed information. As time proceeds, the next prediction is made independently of the previous one, which means that the model cannot correct its errors during inference and will repeat them. To alleviate this problem and better leverage temporal data, we propose a novel retrospection technique. Through training on closed-loop rollouts the model learns to use aggregated feedback. Given new observations it reflects on previous predictions and analyzes its errors to improve the quality of subsequent predictions. Thus, the model can learn to correct systematic errors during inference. Comprehensive experiments on nuScenes and Argoverse demonstrate a considerable decrease in minimum Average Displacement Error of up to 31.9% compared to the state-of-the-art baseline without retrospection. We further showcase the robustness of our technique by demonstrating a better handling of out-of-distribution scenarios with undetected road-users.

Monocular 3D lane detection has become a fundamental problem in the context of autonomous driving, which comprises the tasks of finding the road surface and locating lane markings. One major challenge lies in a flexible but robust line representation capable of modeling complex lane structures, while still avoiding unpredictable behavior. While previous methods rely on fully data-driven approaches, we instead introduce a novel approach LaneCPP that uses a continuous 3D lane detection model leveraging physical prior knowledge about the lane structure and road geometry. While our sophisticated lane model is capable of modeling complex road structures, it also shows robust behavior since physical constraints are incorporated by means of a regularization scheme that can be analytically applied to our parametric representation. Moreover, we incorporate prior knowledge about the road geometry into the 3D feature space by modeling geometry-aware spatial features, guiding the network to learn an internal road surface representation. In our experiments, we show the benefits of our contributions and prove the meaningfulness of using priors to make 3D lane detection more robust. The results show that LaneCPP achieves state-of-the-art performance in terms of F-Score and geometric errors.

Planning the trajectory of the controlled ego vehicle is a key challenge in automated driving. As for human drivers, predicting the motions of surrounding vehicles is important to plan the own actions. Recent motion prediction methods utilize equivariant neural networks to exploit geometric symmetries in the scene. However, no existing method combines motion prediction and trajectory planning in a joint step while guaranteeing equivariance under roto-translations of the input space. We address this gap by proposing a lightweight equivariant planning model that generates multi-modal joint predictions for all vehicles and selects one mode as the ego plan. The equivariant network design improves sample efficiency, guarantees output stability, and reduces model parameters. We further propose equivariant route attraction to guide the ego vehicle along a high-level route provided by an off-the-shelf GPS navigation system. This module creates a momentum from embedded vehicle positions toward the route in latent space while keeping the equivariance property. Route attraction enables goal-oriented behavior without forcing the vehicle to stick to the exact route. We conduct experiments on the challenging nuScenes dataset to investigate the capability of our planner. The results show that the planned trajectory is stable under roto-translations of the input scene which demonstrates the equivariance of our model. Despite using only a small split of the dataset for training, our method improves L2 distance at 3 s by 20.6 % and surpasses the state of the art.

We address the problem of improving the performance and in particular the sample complexity of deep neural networks by enforcing and guaranteeing invariances to symmetry transformations rather than learning them from data. Group-equivariant convolutions are a popular approach to obtain equivariant representations. The desired corresponding invariance is then imposed using pooling operations. For rotations, it has been shown that using invariant integration instead of pooling further improves the sample complexity. In this contribution, we first expand invariant integration beyond rotations to flips and scale transformations. We then address the problem of incorporating multiple desired invariances into a single network. For this purpose, we propose a multi-stream architecture, where each stream is invariant to a different transformation such that the network can simultaneously benefit from multiple invariances. We demonstrate our approach with successful experiments on Scaled-MNIST, SVHN, CIFAR-10 and STL-10.

Deep neural networks tend to make overconfident predictions and often require additional detectors for misclassifi-cations, particularly for safety-critical applications. Existing detection methods usually only focus on adversarial attacks or out-of-distribution samples as reasons for false predictions. However, generalization errors occur due to diverse reasons often related to poorly learning relevant invariances. We therefore propose GIT, a holistic approach for the detection of generalization errors that combines the usage of gradient information and invariance transformations. The invariance transformations are designed to shift misclassified samples back into the generalization area of the neural network, while the gradient information measures the contradiction between the initial prediction and the corre-sponding inherent computations of the neural network using the transformed sample. Our experiments demonstrate the superior performance of GIT compared to the state-of-the-art on a variety of network architectures, problem setups and perturbation types.

Monocular 3D traffic line detection jointly tackles the detection of lane markings and regression of their 3D location. The greatest challenge is the exact estimation of various line shapes in the world, which highly depends on the chosen representation. While anchor-based and grid-based line representations have been proposed, all suffer from the same limitation, the necessity of discretizing the 3D space. To address this limitation, we present an anchor-free parametric lane representation, which defines traffic lines as continuous curves in 3D space. Choosing splines as our representation, we show their superiority over polynomials of different degrees that were proposed in previous 2D lane detection approaches. Our continuous representation allows us to model even complex lane shapes at any position in the 3D space, while implicitly enforcing smoothness constraints. Our model is validated on a synthetic 3D lane dataset including a variety of scenes in terms of complexity of road shape and illumination. We outperform the state-of-the-art in nearly all geometric performance metrics and achieve a great leap in the detection rate. In contrast to discrete representations, our parametric model requires no post-processing achieving highest processing speed. Additionally, we provide a thorough analysis over different parametric representations for 3D lane detection. The code and trained models are available on our project website https://3d-splinenet.github.io/.

Unstructured pruning is well suited to reduce the memory footprint of convolutional neural networks (CNNs), both at training and inference time. CNNs contain parameters arranged in K x K filters. Standard unstructured pruning (SP) reduces the memory footprint of CNNs by setting filter elements to zero, thereby specifying a fixed subspace that constrains the filter. Especially if pruning is applied before or during training, this induces a strong bias. To overcome this, we introduce interspace pruning (IP), a general tool to improve existing pruning methods. It uses filters represented in a dynamic interspace by linear combinations of an underlying adaptive filter basis (FB). For IP, FB coefficients are set to zero while un-pruned coefficients and FBs are trained jointly. In this work, we provide mathematical evidence for IP's superior performance and demonstrate that IP outperforms SP on all tested state-of-the-art unstructured pruning methods. Especially in challenging situations, like pruning for ImageNet or pruning to high sparsity, IP greatly exceeds SP with equal runtime and parameter costs. Finally, we show that advances of IP are due to improved trainability and superior generalization ability.

State-of-the-art deep neural network (DNN) pruning techniques, applied one-shot before training starts, evaluate sparse architectures with the help of a single criterion-called pruning score. Pruning weights based on a solitary score works well for some architectures and pruning rates but may also fail for other ones. As a common baseline for pruning scores, we introduce the notion of a generalized synaptic score (GSS). In this work we do not concentrate on a single pruning criterion, but provide a framework for combining arbitrary GSSs to create more powerful pruning strategies. These COmbined Pruning Scores (COPS) are obtained by solving a constrained optimization problem. Optimizing for more than one score prevents the sparse network to overly specialize on an individual task, thus COntrols Pruning before training Starts. The combinatorial optimization problem given by COPS is relaxed on a linear program (LP). This LP is solved analytically and determines a solution for COPS. Furthermore, an algorithm to compute it for two scores numerically is proposed and evaluated. Solving COPS in such a way has lower complexity than the best general LP solver. In our experiments we compared pruning with COPS against state-of-the-art methods for different network architectures and image classification tasks and obtained improved results.

Domain adaptation techniques enable the re-use and transfer of existing labeled datasets from a source to a target domain in which little or no labeled data exists. Recently, image-level domain adaptation approaches have demonstrated impressive results in adapting from synthetic to real-world environments by translating source images to the style of a target domain. However, the domain gap between source and target may not only be caused by a different style but also by a change in viewpoint. This case necessitates a semantically consistent translation of source images and labels to the style and viewpoint of the target domain. In this work, we propose the Novel Viewpoint Adaptation (NoVA) model, which enables unsupervised adaptation to a novel viewpoint in a target domain for which no labeled data is available. NoVA utilizes an explicit representation of the 3D scene geometry to translate source view images and labels to the target view. Experiments on adaptation to synthetic and real-world datasets show the benefit of NoVA compared to state-of-the-art domain adaptation approaches on the task of semantic segmentation.

In this paper we present a modification of the popular Black-Box Variational Inference (BBVI) approach which significantly improves the computational efficiency of the inference. We achieve this performance boost by replacing the standard gradient in the stochastic gradient ascent framework of BBVI with the natural gradient. Our experimental results (e.g. training of neutral networks) show that the proposed method outperforms the original BBVI algorithm on both synthetic and real data.

The concept of “Gaussianization” implies a transformation aimed at changing the distribution of the input random variable to Gaussian. It has been used until now as a means to achieve independence among components in multivariate distributions that in turn was used as a tool for various purposes ranging from density estimation to normalization. In this contribution we propose Gaussianization for pattern recognition applications in support of Gaussianity assumptions made by various classifiers. Previous approaches completely ignore separability considerations, the Gaussianization being conducted over the entire data, irrespective of class affiliation and are not useful for recognition purposes. We instead propose a transform such that the output random variable is distributed according to a Gaussian mixture, where each class accounts for one mixture component. We successfully test our method on both synthetic and real data.

In a pattern classification setup, image segmentation is achieved by assigning each pixel to one of two classes: object or background. The special case of vessel segmentation is characterized by a strong disproportion between the number of representatives of each class (i.e. class skew) and also by a strong overlap between classes. These difficulties can be solved using problem-specific knowledge. The proposed hysteresis classification makes use of such knowledge in an efficient way. We describe a novel, supervised, hysteresis-based classification method that we apply to the segmentation of retina photographies. This procedure is fast and achieves results that comparable or even superior to other hysteresis methods and, for the problem of retina vessel segmentation, to known dedicated methods on similar data sets.

The success of skin transplantations depends on the adequate revascularization of the transplanted dermal matrix. To induce vessel growth or angiogenesis, pharmacological substances may be applied to the dermal matrix. The effectiveness of different such substances has been evaluated in laboratory experiments. For this purpose, the surface and length of newly grown vessels have to be measured in micro-angiograms (x-ray images of the blood vessels recorded after the injection of a radiopaque substance) of tissue transplanted on the back of laboratory animals. To this end we describe in this contribution a vessel analysis environment central to which is a vessel segmentation tool for surface quantification in fasciocutaneous skin transplant micro-angiograms.

The success of skin transplantations depends on a proper revascularization of the transplanted tissue. Angiogenesis (vessel growth) in the transplanted dermal matrices can be stimulated by administering different drugs. To evaluate the effectiveness of different drug treatments in an experimental setting using laboratory animals, vessels in transplant samples are visualized by so-called micro-angiograms, i.e., X-ray images after contrast agent injection. We describe a framework for the acquisition of such micro-angiograms as well as for the subsequent semi-automatic analysis of angiogenesis. Central to our analysis is the segmentation of small fasciocutaneous vessels in the transplant sample.