Deep neural network (DNN) is extensively explored for LiDAR-based 3D object detection, a crucial perception task in the field of autonomous driving. However, the presence of redundant parameters and complex computatio...
详细信息
Deep neural network (DNN) is extensively explored for LiDAR-based 3D object detection, a crucial perception task in the field of autonomous driving. However, the presence of redundant parameters and complex computations pose challenges for the practical deployment of DNNs. Despite knowledge distillation (KD) is an effective approach for accelerating models, extremely small number of efforts explore its potential on LiDARbased 3D detectors. Besides, existing studies neglect to elaborately investigate 3D voxel-wise features for compression. To this end, we propose instance-aware knowledge distillation (InstKD) for 3D detector compression. The proposed method conducts KD by fully excavating two types of knowledge related to 3D voxelwise features. Firstly, the 3D voxel-wise feature of teacher is transferred to teach the student. In order to prioritize the knowledge with strong guiding capacity, we introduce expanded bounding box (E-Bbox) to distinguish and balance the foreground and background regions. Besides, we generate contribution map (CM) by calculating the gap between the classification response of teacher and student models to further dynamically balance individual instance for distillation. Secondly, we also align the relation-based knowledge of 3D voxel-wise features between the distillation pairs. To avoid incalculable relation on a massive number of 3D voxel-wise features, we distill the relation among instances selected by E-Bboxes, where the intra-relation of homogeneous instances and inter-relation of heterogeneous instances are transferred in a dual-pathway manner. In the experiments, we compress different models on benchmarks with varying scales. The results demonstrate that our method achieves the lightweight 3D detector with slight performance drop. For example, on KITTI dataset, our 2× compressed SECOND (75.5% parameters and 74.5% FLOPs reduction) achieves 66.83% mAP, surpassing its teacher model. The key code is available at https://***/zhnxj
Wound healing can be delayed if the biomechanical stability of the wound closure is inadequate. Therefore, it is necessary to investigate different suturing techniques for their biomechanical stability. In this study,...
详细信息
Despite the success of Multi-Agent Reinforcement Learning (MARL) algorithms in cooperative tasks, previous works, unfortunately, face challenges in heterogeneous scenarios since they simply disable parameter sharing f...
Despite some efforts and attempts have been made to improve the direction-of-arrival(DOA)estimation performance of the standard Capon beamformer(SCB)in array processing,rigorous statistical performance analyses of the...
详细信息
Despite some efforts and attempts have been made to improve the direction-of-arrival(DOA)estimation performance of the standard Capon beamformer(SCB)in array processing,rigorous statistical performance analyses of these modified Capon estimators are still *** paper studies an improved Capon estimator(ICE)for estimating the DOAs of multiple uncorrelated narrowband signals,where the higherorder inverse(sample)array covariance matrix is used in the Capon-like cost *** establishing the relationship between this nonparametric estimator and the parametric and classic subspace-based MUSIC(multiple signal classification),it is clarified that as long as the power order of the inverse covariance matrix is increased to reduce the influence of signal subspace components in the ICE,the estimation performance of the ICE becomes equivalent to that of the MUSIC regardless of the signal-to-noise ratio(SNR).Furthermore the statistical performance of the ICE is analyzed,and the large-sample mean-squared-error(MSE)expression of the estimated DOA is *** the effectiveness and the theoretical analysis of the ICE are substantiated through numerical examples,where the Cramer-Rao lower bound(CRB)is used to evaluate the validity of the derived asymptotic MSE expression.
The increasing interest in cryogenic circuits is driven by their transformative potential across high-performance computing, medical devices, space exploration, and quantum technologies. Operating transistors at cryog...
详细信息
Artificial intelligence has proven its benefits in many domains. Yet, traditional deep learning models are still too energy and compute-intensive for resource-constrained edge environments. Spiking neural networks (SN...
详细信息
Hyperspectral image(HSI) denoising addresses noise impact during image acquisition. Transformers have gained notable prominence in the field of denoising, but their quadratic self-attention complexity poses computatio...
详细信息
In der sich rasant entwickelnden Landschaft der industriellen Automatisierung läutet das Aufkommen von Industrie 5.0 (I5.0) einen Paradigmenwechsel hin zu einem stärker kollaborativen und menschzentrierten A...
In der sich rasant entwickelnden Landschaft der industriellen Automatisierung läutet das Aufkommen von Industrie 5.0 (I5.0) einen Paradigmenwechsel hin zu einem stärker kollaborativen und menschzentrierten Ansatz ein. In diesem Beitrag wird die Rolle der Mensch-Maschine-Kollaboration und menschzentrierter Werkzeuge bei der Förderung einer symbiotischen Beziehung zwischen fortschrittlichen Technologien und menschlichen Benutzern untersucht, um so das volle Potenzial von I5.0 zu erschließen. Als nächste Stufe in der Entwicklung des Produktionssektors zielt I5.0 darauf ab, ein Gleichgewicht zwischen Automatisierung und menschlichen Fähigkeiten herzustellen und die sich ergänzenden Stärken beider zu nutzen. Es werden Technologien vorgestellt, welche menschzentrierte Lösungen zur Steigerung von Produktivität, Flexibilität und Nachhaltigkeit in der Fabrik der Zukunft fokussieren.
The sensor configuration of an autonomous vehicle (AV) is determined in the early development phase when specific perception algorithms are not yet available. Therefore, approaches based on synthetic raw data are nece...
详细信息
ISBN:
(数字)9781665410205
ISBN:
(纸本)9781665410212
The sensor configuration of an autonomous vehicle (AV) is determined in the early development phase when specific perception algorithms are not yet available. Therefore, approaches based on synthetic raw data are necessary to evaluate different configurations. One sensor type used in AV is LiDAR, but developers should carefully consider the amount and placement of the sensors due to their high costs. In this contribution, we propose the Omni-Lidar Evaluation Score (OLES), a novel metric to evaluate different LiDAR configurations based on their simulated raw data. Our OLES metric combines information-theoretic quantities with coverage-based metrics, considering both the spatial coverage and the uniformity of a LiDAR point cloud distribution. We show the need for a new metric and provide details on implementing OLES using the open-source simulator CARLA. We demonstrate the effectiveness of our new metric in a simulation study and highlight its usefulness in the early phases of vehicle development. This research provides a means to evaluate the quality of LiDAR configurations and provides a basis for further optimizing sensor setups for AVs.
Today, deep learning detectors for autonomous driving are delivering impressive results on public datasets and in real-world applications. However, these detectors require large amounts of data, especially labeled dat...
详细信息
ISBN:
(数字)9798350348811
ISBN:
(纸本)9798350348828
Today, deep learning detectors for autonomous driving are delivering impressive results on public datasets and in real-world applications. However, these detectors require large amounts of data, especially labeled data, to achieve the performance needed to ensure safe driving. The process of collecting and tagging data is expensive and cumbersome. Therefore, the recent focus of the industry has been on how to achieve similar performance while limiting the amount of labeled data required to train such models. Within the cross-modal active learning paradigm, we propose and analyze new strategies to exploit the inconsistencies between camera and LiDAR detectors to improve sampling efficiency and label only the samples that promise improvements for model training. For this, we leverage the 2D projection of the bounding boxes to equalize the output quality of camera and LiDAR detections. Finally, we achieve up to 0.6% AP improvement for camera and 2% improvement for LiDAR over random sampling on the KITTI dataset using a sampling strategy based on the number of detected objects.
暂无评论