Accurate trajectory prediction is crucial for autonomous driving, yet uncertainty in agent behavior and perception noise makes it inherently challenging. While multi-modal trajectory prediction models generate multiple plausible future paths with associated probabilities, effectively quantifying uncertainty remains an open problem. In this work, we propose a novel multi-modal trajectory prediction approach based on evidential deep learning that estimates both positional and mode probability uncertainty in real time. Our approach leverages a Normal Inverse Gamma distribution for positional uncertainty and a Dirichlet distribution for mode uncertainty. Unlike sampling-based methods, it infers both types of uncertainty in a single forward pass, significantly improving efficiency. Additionally, we experimented with uncertainty-driven importance sampling to improve training efficiency by prioritizing underrepresented high-uncertainty samples over redundant ones. We perform extensive evaluations of our method on the Argoverse 1 and Argoverse 2 datasets, demonstrating that it provides reliable uncertainty estimates while maintaining high trajectory prediction accuracy.

Abstract A considerable amount of research is concerned with the generation of realistic sensor data. LiDAR point clouds are generated by complex simulations or learned generative models. The generated data is usually exploited to enable or improve downstream perception algorithms. Two major questions arise from these procedures: First, how to evaluate the realism of the generated data? Second, does more realistic data also lead to better perception performance? This paper addresses both questions and presents a novel metric to quantify the realism of LiDAR point clouds. Relevant features are learned from real-world and synthetic point clouds by training on a proxy classification task. In a series of experiments, we demonstrate the application of our metric to determine the realism of generated LiDAR data and compare the realism estimation of our metric to the performance of a segmentation model. We confirm that our metric provides an indication for the downstream segmentation performance.

Semantic scene completion is the task of jointly estimating 3D geometry and semantics of objects and surfaces within a given extent. This is a particularly challenging task on real-world data that is sparse and occluded. We propose a scene segmentation network based on local Deep Implicit Functions as a novel learning-based method for scene completion. Unlike previous work on scene completion, our method produces a continuous scene representation that is not based on voxelization. We encode raw point clouds into a latent space locally and at multiple spatial resolutions. A global scene completion function is subsequently assembled from the localized function patches. We show that this continuous representation is suitable to encode geometric and semantic properties of extensive outdoor scenes without the need for spatial discretization (thus avoiding the trade-off between level of scene detail and the scene extent that can be covered). We train and evaluate our method on semantically annotated LiDAR scans from the Semantic KITTI dataset. Our experiments verify that our method generates a powerful representation that can be decoded into a dense 3D description of a given scene. The performance of our method surpasses the state of the art on the Semantic KITTI Scene Completion Benchmark in terms of geometric completion intersection-over-union (IoU).

This work proposes a spatially-conditioned neural network to perform semantic segmentation and geometric scene completion in 3D on real-world LiDAR data. Spatially-conditioned scene segmentation (SCSSnet) is a representation suitable to encode properties of large 3D scenes at high resolution. A novel sampling strategy encodes free space information from LiDAR scans explicitly and is both simple and effective. We avoid the need for synthetically generated or volumetric ground truth data and are able to train and evaluate our method on semantically annotated LiDAR scans from the Semantic KITTI dataset. Ultimately, our method is able to predict scene geometry as well as a diverse set of semantic classes over a large spatial extent at arbitrary output resolution instead of a fixed discretization of space. Our experiments confirm that the learned scene representation is versatile and powerful and can be used for multiple downstream tasks. We perform point-wise semantic segmentation, point-of-view depth completion and ground plane segmentation. The semantic segmentation performance of our method surpasses the state of the art by a significant margin of 7% mIoU.

Autonomous vehicles need to have a semantic understanding of the three-dimensional world around them in order to reason about their environment. State of the art methods use deep neural networks to predict semantic classes for each point in a LiDAR scan. A powerful and efficient way to process LiDAR measurements is to use two-dimensional, image-like projections. In this work, we perform a comprehensive experimental study of image-based semantic segmentation architectures for LiDAR point clouds. We demonstrate various techniques to boost the performance and to improve runtime as well as memory constraints. First, we examine the effect of network size and suggest that much faster inference times can be achieved at a very low cost to accuracy. Next, we introduce an improved point cloud projection technique that does not suffer from systematic occlusions. We use a cyclic padding mechanism that provides context at the horizontal field-of-view boundaries. In a third part, we perform experiments with a soft Dice loss function that directly optimizes for the intersection-over-union metric. Finally, we propose a new kind of convolution layer with a reduced amount of weight-sharing along one of the two spatial dimensions, addressing the large difference in appearance along the vertical axis of a LiDAR scan. We propose a final set of the above methods with which the model achieves an increase of 3.2% in mIoU segmentation performance over the baseline while requiring only 42% of the original inference time.

Autonomous vehicles require an accurate understanding of the underlying motion of their surroundings. Traditionally this understanding is acquired using optical flow algorithms on camera images, RADAR sensors which measure velocity directly or by object tracking through various sensors. We propose a novel method to estimate point-wise 3D motion vectors from LiDAR point clouds using fully convolutional networks trained and evaluated on the KITTI dataset. Besides, we show how this motion information can be used to efficiently estimate odometry. We demonstrate that our approach achieves significant speed ups over the current state of the art.

A considerable amount of annotated training data is necessary to achieve state-of-the-art performance in perception tasks using point clouds. Unlike RGB-images, LiDAR point clouds captured with different sensors or varied mounting positions exhibit a significant shift in their input data distribution. This can impede transfer of trained feature extractors between datasets as it degrades performance vastly. We analyze the transferability of point cloud features between two different LiDAR sensor set-ups (32 and 64 vertical scanning planes with different geometry). We propose a supervised training methodology to learn transferable features in a pre-training step on LiDAR datasets that are heterogeneous in their data and label domains. In extensive experiments on object detection and semantic segmentation in a multi-task setup we analyze the performance of our network architecture under the impact of a change in the input data domain. We show that our pre-training approach effectively increases performance for both target tasks at once without having an actual multi-task dataset available for pre-training.

This paper presents a novel CNN-based approach for synthesizing high-resolution LiDAR point cloud data. Our approach generates semantically and perceptually realistic results with guidance from specialized loss-functions. First, we utilize a modified per-point loss that addresses missing LiDAR point measurements. Second, we align the quality of our generated output with real-world sensor data by applying a perceptual loss. In large-scale experiments on real-world datasets, we evaluate both the geometric accuracy and semantic segmentation performance using our generated data vs. ground truth. In a mean opinion score testing we further assess the perceptual quality of our generated point clouds. Our results demonstrate a significant quantitative and qualitative improvement in both geometry and semantics over traditional non CNN-based upsampling methods.