{"id": "2605.12220", "title": "TriBand-BEV: Real-Time LiDAR-Only 3D Pedestrian Detection via Height-Aware BEV and High-Resolution Feature Fusion", "abstract": "Safe autonomous agents and mobile robots need fast real time 3D perception, especially for vulnerable road users (VRUs) such as pedestrians. We introduce a new bird's eye view (BEV) encoding, which maps the full 3D LiDAR point cloud into a light-weight 2D BEV tensor with three height bands. We explicitly reformulate 3D detection as a 2D detection problem and then reconstruct 3D boxes from the BEV outputs. A single network detects cars, pedestrians, and cyclists in one pass. The backbone uses area attention at deep stages, a hierarchical bidirectional neck over P1 to P4 fuses context and detail, and the head predicts oriented boxes with distribution focal learning for side offsets and a rotated IoU loss. Training applies a small vertical re bin and a mild reflectance jitter in channel space to resist memorization. We use an interquartile range (IQR) filter to remove noisy and outlier LiDAR points during 3D reconstruction. On KITTI dataset, TriBand-BEV attains 58.7/52.6/47.2 pedestrian BEV AP(%) for easy, moderate, and hard at 49 FPS on a single consumer GPU, surpassing Complex-YOLO, with gains of +12.6%, +7.5%, and +3.1%. Qualitative scenes show stable detection under occlusion. The pipeline is compact and ready for real time robotic deployment. Our source code is publicly available on GitHub.", "date": "2026-05-12", "categories": []} {"id": "2411.07742", "title": "Efficient 3D Perception on Multi-Sweep Point Cloud with Gumbel Spatial Pruning", "abstract": "This paper studies point cloud perception within outdoor environments. Existing methods face limitations in recognizing objects located at a distance or occluded, due to the sparse nature of outdoor point clouds. In this work, we observe a significant mitigation of this problem by accumulating multiple temporally consecutive point cloud sweeps, resulting in a remarkable improvement in perception accuracy. However, the computation cost also increases, hindering previous approaches from utilizing a large number of point cloud sweeps. To tackle this challenge, we find that a considerable portion of points in the accumulated point cloud is redundant, and discarding these points has minimal impact on perception accuracy. We introduce a simple yet effective Gumbel Spatial Pruning (GSP) layer that dynamically prunes points based on a learned end-to-end sampling. The GSP layer is decoupled from other network components and thus can be seamlessly integrated into existing point cloud network architectures. Without incurring additional computational overhead, we increase the number of point cloud sweeps from 10, a common practice, to as many as 40. Consequently, there is a significant enhancement in perception performance. For instance, in nuScenes 3D object detection and BEV map segmentation tasks, our pruning strategy improves several 3D perception baseline methods.", "date": "2024-11-12", "categories": []} {"id": "2606.12981", "title": "Camera and LiDAR BEV Fusion for Cooperative 3D Object Detection on TUMTraf V2X", "abstract": "We describe a Camera and LiDAR fusion detector developed for the TUMTraf V2X cooperative 3D object detection track of the DriveX 2026 challenge. The detector fuses three roadside cameras with a fused infrastructure-plus-vehicle point cloud in a shared bird's-eye-view space and predicts boxes through a CenterPoint-style head with a generalized IoU regression loss and an IoU quality re-ranking head. Trained on the provided train and validation splits, the model reaches a 3D mAP of 0.85 on the public Codabench test split. While iterating on the system, we observed that 44 of the 50 test frames are also present in the released train (40) and validation (4) splits with their labels. We therefore conducted two additional studies to quantify how this overlap affects the final score: (1) a finetuning run that oversamples the 44 overlapping frames, reaching 0.89 mAP, and (2) a post-processing run that replaces predictions on those frames with the released ground truth, reaching 0.99 mAP (uploaded to our Codabench account for testing but not published on the leaderboard). All three configurations and their per-class results are reported.", "date": "2026-06-11", "categories": []} {"id": "2606.11573", "title": "Understanding Cross-Sensor Feature Variations for Generalizable 3D Perception", "abstract": "Radar-camera BEV perception often suffers from degraded performance when evaluated across datasets, as changes in driving scenes, sensor configurations, and environmental conditions can alter both the input observations and the internal fused representations. This work studies this issue from the perspective of source-domain variation modeling, aiming to improve the robustness of BEV-based 3D detectors without relying on target-domain samples. We introduce a framework that characterizes visual scene variations in the frequency domain and uses them to synthesize diverse source-domain views. By comparing the resulting fused BEV representations, the framework further captures how image-level variations influence multi-modal BEV features. These variation patterns are then used to regularize the detector, encouraging the learned fusion space to remain stable under latent scene changes. The proposed method is applied only during training and leaves the inference pipeline unchanged. Experiments on cross-dataset radar-camera 3D detection between View-of-Delft and TJ4DRadSet demonstrate consistent improvements over multiple BEV fusion backbones, and the gains remain effective when a small amount of target-domain data is available.", "date": "2026-06-10", "categories": []} {"id": "2606.11739", "title": "Multi-View In-Cabin Monitoring System for Public Transport Vehicles", "abstract": "We introduce a multi-view in-cabin monitoring dataset for public transportation with synchronized RGB and depth images from four inward-facing cameras and a rotating LiDAR covering the vehicle interior of a digitalized and partly automated German city bus. The dataset contains 9.136 synchronized samples with annotations and is accompanied by a calibration and pseudo-labeling pipeline that generates 3D human pose estimates and oriented 3D bounding boxes for occupants. We further provide a nuScenes-format conversion and benchmark representative multi-view 3D detection models (e.g., Lift-Splat-Shoot and BEVFusion), supporting comparative evaluation and small-scale training of multi-view in-cabin perception models. The dataset and tools are available at https://github.com/EvgenyGorelik/multiview_incabin_dataset.", "date": "2026-06-10", "categories": []} {"id": "2606.08680", "title": "Distortion-Aware PETR for BEV Object Detection with Mixed Pinhole-Fisheye Cameras", "abstract": "Fisheye cameras are widely deployed in autonomous driving perception suites for their low cost and full-coverage field of view (FOV), yet their potential remains underleveraged in 3D object detection. Severe radial distortion challenges most BEV detectors by violating the fundamental assumption of uniform sampling. To bridge this gap, we propose Distortion-Aware PETR (DAPETR), a projection-free detector tailored for mixed pinhole-fisheye camera setups. DAPETR incorporates two key learned-adaptive modules: a unified distortion-aware positional embedding that harmonizes positional encodings for image representations with fisheye geometry, and a bidirectional feature-geometry co-modulation module that mutually adapts image features and 3D positional embeddings. In our experiments on a converted KITTI-360 benchmark, we systematically compare our learned adaptive approach against PETR in polar coordinates (PolarPETR). We find that while both methods improve over the baseline, our learned modules achieve superior performance. Crucially, we uncover a negative interaction when combining both strategies, revealing that learned adaptation and explicit geometric reparameterization can conflict. Our final DAPETR model significantly advances the research and benchmark for fisheye BEV detection, providing critical insights into effective distortion-aware 3D perception design other than image rectification.", "date": "2026-06-07", "categories": []} {"id": "2606.08844", "title": "Geometry-Aware Fisheye-LiDAR Fusion for Robust 3D Object Detection in Low-Overlap Setups", "abstract": "As autonomous systems expand from capital-intensive robotaxis to cost-sensitive logistics, sensor configurations are increasingly optimized for coverage-per-cost. A prevalent sparse-view setup utilizes dual-fisheye cameras with a roof-mounted LiDAR, introducing severe geometric challenges: extreme radial distortion, minimal overlap, and misalignment between spherical projections and rectilinear grids. BEV fusion algorithms typically force image and point cloud modalities into unified Cartesian grids early in the pipeline, causing significant feature distortion and information loss for wide-view fisheye cameras. To address this, we propose a Geometry-Aware Hybrid Fusion (GA-HF) framework that explicitly accounts for fisheye geometry and BEV feature distortion, where fisheye features are lifted into a polar BEV grid via a Distortion-Aware Lift-Splat-Shoot (LSS) module to preserve native angular density, while LiDAR features are processed in native Cartesian space for metric fidelity of bounding box regression. To bridge these heterogeneous streams, we introduce a Dual-Attention Warping Correction module that applies spatial and channel attention to the warped camera features before fusion, explicitly suppressing artifacts in low-quality peripheral regions while enhancing high-quality semantic cues. GA-HF is evaluated on three benchmarks: KITTI-360, Dur360BEV, and Fisheye3DOD datasets. To the best of our knowledge, it is the first approach to explore LiDAR-fisheye camera fusion. On KITTI-360, GA-HF improves NDS by 4.2% over Cartesian baselines; on Dur360BEV, it surpasses both LiDAR-only and BEVFusion, while significantly reducing orientation error despite the geometric distortions; on Fisheye3DOD, it attains the highest detection score among all fusion methods.", "date": "2026-06-07", "categories": []} {"id": "2606.09143", "title": "CAMF-Det: Closure-Aware Multimodal Fusion for LiDAR-Camera 3D Object Detection on UAV Platforms", "abstract": "Multimodal 3D object detection based on LiDAR and cameras has demonstrated excellent performance in ground-vehicle scenarios, but has not been explored for Unmanned Aerial Vehicle (UAV) platforms. In UAV top-down scenes, frequent groundobject occlusion dominated by tree canopies causes spatially varying and modality-dependent information degradation. Existing multimodal fusion frameworks neither explicitly model such ground-object occlusion nor embed occlusion awareness into the detection pipeline, limiting their performance in occluded UAV scenes. To address these challenges, we propose CAMF-Det, a closure-aware multimodal fusion framework for LiDAR-camera 3D object detection on UAV platforms, which derives dual-modal occlusion intensity through physics-inspired modeling and embeds them as priors throughout the detection pipeline. First, a dual-modal closure modeling module explicitly constructs occlusion intensity ground truth for both modalities offline via a Beer-Lambert-inspired formulation and building-mask correction. Second, using these ground-truth maps as supervision, a dual-modal prediction network converts the offline modeling results into online occlusion intensity predictions under single-frame inference. Third, both ground-truth and predicted occlusion intensity are injected into data augmentation, feature encoding, multimodal fusion, and detection head, enabling adaptive detection under spatially varying and modality-dependent information degradation. Experiments on two self-built UAV-based multimodal datasets, SI3D-DI and SI3D-DII, demonstrate that CAMF-Det achieves the best performance across all difficulty levels, with hard-level mAP$_{\\mathrm{BEV}}$ improvements of 9.43% and 4.88% over the best competing methods, respectively. These results confirm the effectiveness of explicit occlusion prior modeling and exploitation for robust multimodal 3D detection in UAV scenes.", "date": "2026-06-08", "categories": []} {"id": "2606.09350", "title": "Taming Perception Jitter: Uncertainty-Aware LiDAR Object Detection for Reliable Motion Classification", "abstract": "Reliable motion classification is critical for autonomous driving, as false dynamic predictions of static objects can cascade into unnecessary planner interventions. Unstable bounding box predictions can lead to spurious velocity estimates in tracking and falsely predicted trajectories. We present a deployment-friendly mitigation strategy that augments a 3D object detector with aleatoric uncertainty estimates and applies a two-sample z-test over short observation windows to separate true motion from jitter. Integrated into Autoware with minimal changes, the approach reuses existing data association for minimal compute overhead. Empirical results show parity with velocity thresholding on nuScenes, but substantially fewer false dynamic predictions and unnecessary stops in real-world test drives, explained by the presence of an intermediate jitter band in the recorded data that speed-only rules misclassify. This demonstrates that uncertainty-aware detection and lightweight statistical testing can deliver practical performance gains for autonomous driving in noisier real-world settings.", "date": "2026-06-08", "categories": []} {"id": "2606.09634", "title": "ATN3D: Density-Aware LiDAR-Radar Early 3D Object Detection Under Extreme Sparsity", "abstract": "3D object detection is the backbone of perception for automated vehicles (AV) and broader intelligent transportation systems applications. Long-range detection is challenging because sensing evidence is sparse; yet this ``long-range'' scenario is routine in traffic. Although >30m is often labeled long-range in computer vision, on roadways it affords only approx. 1-2s for perception and decision-making. Under such extreme sparsity, two core challenges arise. First, early multimodal fusion tends to discard sparsity information and inject noise from empty or falsely occupied cells, degrading long-range recall. Second, context-agnostic uniform channel supervision favors dense and near-range samples, leaving far and small objects under-optimized, delaying the earliest detection of distant objects. We propose ``Ask The Neighbor'' (ATN3D), a LiDAR-Radar framework tailored for sparse-range conditions. ATN3D introduces (i) Density-aware early fusion with cross-modal gating that conditions fusion on per-voxel density/sparsity and Radar evidence, (ii) Occupancy-gated neighborhood aggregation with circular kernels to aggregate only from credible cells, (iii) Evidence-conditioned channel self-attention to adapt channel weights with weather/range, and (iv) a Range-aware loss that re-balances classification and localization by distance, aligning training with distance-stratified evaluation. On the VoD benchmark across clear and foggy conditions, ATN3D surpasses strong baselines: +3.55% mAP in clear weather and +8.41% mAP under simulated heavy fog; for >30m objects, gains are +3.33% (clear) and +2.09% (heavy fog). These results indicate earlier and more reliable long-range detections under sparse sensing in on-road traffic.", "date": "2026-06-08", "categories": []} {"id": "2606.03568", "title": "Learned Non-Maximum Suppression for 3D Object Detection", "abstract": "Post-processing is a critical stage in LiDAR-based 3D object detection, where dense and overlapping proposals must be filtered for compact and reliable perception. This work introduces two learned filtering modules that replace heuristic non-maximum suppression (NMS) by leveraging relations among detections. D2D-Rescore employs transformer-based detection-to-detection (D2D) attention, while GossipNet3D adapts the 2D GossipNet concept to 3D through localized message passing in bird's-eye view. A metric-aware matching strategy aligned with the nuScenes evaluation protocol ensures consistent training and validation behavior, improving overall detection performance. Both approaches improve mean average precision (mAP), nuScenes detection score (NDS), and true positive quality compared to CircleNMS, particularly for small and infrequent classes, while adding minimal computational overhead. These results demonstrate that learned, detection-level filtering can enhance 3D detector reliability without modifying the base network, offering a principled alternative to heuristic suppression. Code is available at https://github.com/rst-tu-dortmund/learned-3d-nms .", "date": "2026-06-02", "categories": []} {"id": "2602.11554", "title": "HyperDet: 3D Object Detection with Hyper 4D Radar Point Clouds", "abstract": "How far can 3D object detection go using 4D radar alone? Despite offering weather-robust and velocity-aware sensing for autonomous perception, modern 4D radar still yields sparse, noisy, and unstable point clouds, limiting radar-only 3D detection. We present HyperDet, a detector-agnostic framework that constructs task-aware hyper 4D radar point clouds before detection. HyperDet first refines short-window surround-view radar observations through spatio-temporal accumulation, cross-sensor validation, and Doppler-guided motion compensation, improving return reliability and temporal coherence. It then performs foreground generative enhancement using LiDAR-guided pseudo-radar supervision available only during training, enriching object geometry while preserving measured radar background and radar-native attributes. During detector training, radar-aware object-level augmentation further preserves Doppler consistency under geometric relocation. At inference time, HyperDet requires radar input alone and can be directly paired with standard 3D detectors. Experiments on two public surround-view 4D radar datasets demonstrate consistent improvements over raw radar inputs across standard 3D detectors, validating input-level radar enhancement as an effective approach to radar-only 3D detection.", "date": "2026-02-12", "categories": []} {"id": "2605.20301", "title": "Co-Fusion4D: Spatio-temporal Collaborative Fusion for Robust 3D Object Detection", "abstract": "In autonomous driving, 3D object detection is essential for accurate perception and reliable decision-making. However, object motion and ego-motion often induce cross-frame spatiotemporal inconsistencies in BEV-based detectors, leading to temporal BEV feature misalignment and degraded spatiotemporal consistency.\n To address these challenges, we propose Co-Fusion4D, a unified framework that explicitly preserves cross-frame spatiotemporal consistency and suppresses temporal feature drift. Co-Fusion4D adopts a current-frame-centric strategy, treating the current frame as the primary source of information while selectively incorporating historical frames after spatiotemporal filtering and alignment. This dominant-complementary mechanism effectively mitigates cumulative alignment errors, suppresses noisy feature propagation, and exploits reliable temporal cues for a more consistent BEV representation.\n In addition, Co-Fusion4D integrates a Dual Attention Fusion (DAF) module to further enhance spatiotemporal feature interaction. DAF jointly leverages intra-frame spatial attention and inter-frame temporal attention to adaptively align and fuse multi-frame features, emphasizing motion-consistent regions while suppressing spurious correlations. By departing from conventional uniform fusion paradigms, this design substantially improves the temporal stability and discriminative capability of BEV representations.\n Extensive experiments on the nuScenes benchmark demonstrate that Co-Fusion4D achieves state-of-the-art performance, with 74.9% mAP and 75.6% NDS, without relying on test-time augmentation or external data.", "date": "2026-05-19", "categories": []} {"id": "2606.00688", "title": "Shape-Prior-Based Point Cloud Completion for Single-Stage Fully Sparse 3D Object Detection", "abstract": "Single-stage fully sparse 3D object detectors rely on point clouds data to detect objects in autonomous driving scenarios. However, the sparsity and incompleteness of point clouds significantly limit the performance of 3D object detection. To address this issue, this paper proposes a point clouds completion method specifically designed for single-stage fully sparse detectors. The entire shape-prior-based completion process consists of two consecutive steps. In the first step, we design a novel Instance Selection module, which is capable of identifying point clouds corresponding to foreground objects even when the baseline model does not generate proposals, while effectively ignoring the point clouds of background regions. In the second step, we introduce a novel Alignment-Based Point Completion module, which aligns the point clouds of foreground objects with prototypes in terms of both their centers and orientations. Subsequently, points are selected from the prototype to fill in the missing parts of the foreground object. We evaluated our method on two single-stage fully sparse detectors using the KITTI dataset. The experimental results demonstrate that the proposed method significantly improves the detection performance, confirming its effectiveness and generalizability.", "date": "2026-05-30", "categories": []} {"id": "2606.01757", "title": "PillarDETR: YOLO-Backbone and RT-DETR Head for Real-Time 3D Object Detection", "abstract": "Real-time 3D object detection is a critical component for the safe operation of autonomous driving systems and robotics. While LiDAR point clouds provide accurate spatial information, processing them efficiently remains a significant challenge. Traditional methods rely on complex 3D convolutions or anchor-based paradigms that struggle to balance detection accuracy with inference speed. In this paper, we propose PillarDETR, a novel end-to-end 3D object detection architecture that combines the efficiency of pillar-based LiDAR encoding with the representational power of modern 2D vision models. Specifically, PillarDETR replaces standard convolutional backbones with a Cross Stage Partial (CSP) network derived from YOLOv8, enabling richer feature extraction from pseudoimages. Furthermore, we discard conventional anchor-based or center-based detection heads in favor of a Real-Time Detection Transformer (RT-DETR) decoder. This hybrid design allows the network to capture global context and directly predict 3D bounding boxes without relying on non-maximum suppression (NMS). Extensive experiments on the KITTI and nuScenes benchmarks demonstrate that PillarDETR achieves a compelling trade-off between mean Average Precision (mAP) and inference latency. Our ablation studies confirm that integrating the YOLOv8 backbone and RT-DETR head yields substantial improvements over the PointPillars baseline, establishing PillarDETR as a highly effective solution for real-time 3D perception.", "date": "2026-06-01", "categories": []} {"id": "2605.30983", "title": "Can BEV Perception Gracefully Degrade under Sensor Failures?", "abstract": "Despite the remarkable success of multi-modal bird's-eye view (BEV) perception in autonomous driving, current systems exhibit a critical vulnerability: existing fusion mechanisms are highly brittle to sensor corruptions, often causing catastrophic performance degradation. This vulnerability largely stems from the fact that standard fusion frameworks typically integrate multi-modal representations in a static manner, leading to a precipitous performance collapse under missing or corrupted modalities. In contrast, we show that graceful degradation is achievable through active modality reliability assessment. To this end, we present Grace-BEV, a lightweight and plug-and-play framework that enforces active reliability awareness during multi-modal fusion. Instead of relying on computationally expensive cross-modal interactions, Grace-BEV leverages the aligned BEV space to explicitly assess modality trustworthiness via a TrustGate Router and dynamically recalibrate feature integration using the FailSafe Fusion Block. Furthermore, we devise a Three-Phase Training strategy with Modality Dropout to prevent modality dominance and encourage balanced cross-modal learning under unreliable inputs. Extensive experiments on nuScenes-R and nuScenes-C show that Grace-BEV maintains robust performance across diverse corruption settings. Notably, under catastrophic LiDAR failures where standard baselines collapse to 0.0% mean Average Precision (mAP), Grace-BEV restores performance to as high as 34.7% mAP. Moreover, it improves clean accuracy by up to 1.4%, achieving a strong trade-off between robustness and efficiency.", "date": "2026-05-29", "categories": []} {"id": "2604.12918", "title": "Radar-Camera BEV Multi-Task Learning with Cross-Task Attention Bridge for Joint 3D Detection and Segmentation", "abstract": "Bird's-eye-view (BEV) representations are the dominant paradigm for 3D perception in autonomous driving, providing a unified spatial canvas where detection and segmentation features are geometrically registered to the same physical coordinate system. However, existing radar-camera fusion methods treat these tasks in isolation, missing the opportunity for cross-task feature sharing: object-level geometric cues from detection can sharpen segmentation, while dense road-layout context from segmentation can anchor detection. We propose \\textbf{CTAB} (Cross-Task Attention Bridge), a bidirectional module that exchanges features between detection and segmentation branches via multi-scale deformable attention in shared BEV space. CTAB is integrated into a multi-task framework with an Instance Normalization-based segmentation decoder and learnable BEV upsampling to provide a more detailed BEV representation. On nuScenes, CTAB improves segmentation on 7 classes over the joint multi-task baseline at essentially neutral detection. On a 4-class subset (drivable area, pedestrian crossing, walkway, vehicle), our joint multi-task model achieves 51.0 mIoU-4 while simultaneously providing competitive 3D detection.", "date": "2026-04-14", "categories": []} {"id": "2605.26642", "title": "Adaptation-Free Heterogeneous Collaborative Perception with Unseen Agent Configurations", "abstract": "Collaborative perception improves 3D object detection by enabling agents to share complementary observations, but most existing methods assume fixed or known collaborator encoder configurations, limiting deployment in practice. In this work, we consider an open-world setting in which auxiliary agents with unseen configurations may appear after deployment, such as different LiDAR beam counts or encoder architectures. To address this challenge, we propose ALF, a collaborative perception framework that enables zero-adaptation collaboration with unseen agent configurations by lifting lightweight box-level messages into ego-compatible auxiliary features. ALF converts auxiliary box-level messages into pseudo-BEV maps and synthesizes ego-compatible latent features by combining object-centric cues with scene context from the ego feature. On V2X-Real, under a zero-shot evaluation across 64 case studies, ALF outperforms the strongest prior baseline by 35.91% in relative mAP@0.7 while requiring only 120 bytes per agent per frame (approximately 9.6 Kbps bandwidth at 10 Hz).", "date": "2026-05-26", "categories": []} {"id": "2410.12673", "title": "MambaBEV: An EV-based 3D detection model with Mamba2", "abstract": "Accurate 3D object detection in autonomous driving relies on Bird's Eye View (BEV) perception and effective temporal fusion. However, existing fusion strategies based on convolutional layers or deformable self-attention struggle to model global context in BEV space, leading to reduced accuracy for large objects.To address this limitation, we propose MambaBEV, a novel BEV-based 3D object detection model that leverages Mamba2, an advanced state-space model (SSM) optimized for long-sequence processing. Our key contribution is TemporalMamba, a temporal fusion module that enhances global context modeling through a BEV feature discrete rearrangement mechanism tailored for sequential processing. In addition, we introduce a Mamba-based DETR head to improve multi-object representation. Evaluations on the nuScenes dataset demonstrate that MambaBEV-base achieves 51.7% NDS and an 42.7% mAP. Furthermore, evaluation within an end-to-end autonomous driving paradigm validates its effectiveness in motion forecasting and planning.These results highlight the potential of state-space models for improving global context understanding and large-object detection in autonomous driving perception systems.", "date": "2024-10-16", "categories": []} {"id": "2605.25262", "title": "Semantics-Guided Multimodal Masked Autoencoder Pretraining for 3D BEV Object Detection", "abstract": "Accurate 3D bird's-eye view (BEV) object detection is essential for autonomous driving, and depends strongly on effective multimodal representations from complementary sensors such as cameras and LiDAR. Multimodal masked autoencoders have shown strong potential for learning such representations for downstream 3D BEV object detection. However, existing methods typically apply uniform random masking to camera and LiDAR inputs, treating all regions equally, and learn representations only through masked reconstruction. We propose a semantics-guided multimodal masked autoencoder framework that introduces semantic information during pretraining through two separate components: (i) semantics-guided LiDAR voxel masking, which preserves semantically important LiDAR regions more strongly, and (ii) an auxiliary point-wise LiDAR semantic decoder branch that injects semantic guidance in addition to reconstruction. On BEVFusion 3D object detection, our semantics-guided pretraining strategy improves performance on the nuScenes mini validation set compared to the standard UniM2AE baseline: semantics-guided LiDAR voxel masking yields +1.49% mean Average Precision (mAP) and +1.66% nuScenes Detection Score (NDS), while decoder-side point semantic supervision yields +1.39% mAP and +3.22% NDS over the baseline.", "date": "2026-05-24", "categories": []} {"id": "2605.20390", "title": "STELLAR: Scaling 3D Perception Large Models for Autonomous Driving", "abstract": "Model scaling has demonstrated remarkable success through large-scale training on diverse datasets. It remains an open question whether the same paradigm would apply to autonomous driving perception systems due to unique challenges, such as fusing heterogeneous sensor data and the need for sophisticated 3D spatial understanding. To bridge this gap, we present a comprehensive study on systematically analyzing the impact of scale on these systems. We develop our STELLAR model based on Sparse Window Transformer, by extending the input modalities to include LiDAR, radar, camera, and map prior. We train the model on a large-scale dataset of 50 million driving examples with up to 500 million parameters. Our large-scale experiments reveal empirical scaling trends that connect model performance to model size, data, and compute. The resulting model establishes a new state-of-the-art on the Waymo Open Dataset challenge, outperforming prior arts by a large margin. Our work demonstrates that large-scale training is a highly promising path for advancing the capabilities of perception models for autonomous driving.", "date": "2026-05-19", "categories": []} {"id": "2605.21112", "title": "RCGDet3D: Rethinking 4D Radar-Camera Fusion-based 3D Object Detection with Enhanced Radar Feature Encoding", "abstract": "4D automotive radar is indispensable for autonomous driving due to its low cost and robustness, yet its point cloud sparsity challenges 3D object detection. Existing 4D radar-camera fusion methods focus on complex fusion strategies, trading inference speed for marginal gains. This trade-off hinders real-time deployment due to heavy computation on dense feature maps. In contrast, feature extraction from sparse radar points is less time-consuming but remains under-explored. This work uncovers that simply enhancing radar feature extraction can achieve comparable or even higher performance than elaborate fusion modules, while maintaining real-time performance. Based on this finding, we propose RCGDet3D, which centers on radar feature encoding and simplifies multi-modal fusion. Its encoder inherits from the efficient Gaussian Splatting-based Point Gaussian Encoder (PGE) in RadarGaussianDet3D with two key improvements. First, the Ray-centric PGE (R-PGE) predicts Gaussian attributes in ray-aligned coordinate systems before unifying them to Bird's-Eye View (BEV) space, significantly improving geometric consistency and reducing learning difficulty by decoupling the coordinate transformation from representation learning. Second, a Semantic Injection (SI) module incorporates visual cues from images, producing more geometrically accurate and semantically enriched radar features. Experiments on View-of-Delft (VoD) and TJ4DRadSet show that RCGDet3D outperforms state-of-the-art methods in both accuracy and speed, setting a new benchmark for real-time deployment.", "date": "2026-05-20", "categories": []} {"id": "2512.08237", "title": "Fast-BEV++: Fast by Algorithm, Deployable by Design", "abstract": "The advancement of vision-only Bird's-Eye-View (BEV) perception, a core paradigm for cost-effective autonomous driving, is hindered by the long-standing fundamental trade-off between perception accuracy and on-device deployment efficiency. In this work, we introduce Fast-BEV++, a BEV perception framework that resolves this tension through two fundamental design principles: Fast by Algorithm and Deployable by Design. By decomposing the core view transformation module into a hardware-oriented standard Index-Gather-Reshape pipeline, Fast-BEV++ eliminates dependencies on custom kernels while achieving no less than 3 times speedup over the Fast-BEV baseline across mainstream edge platforms. Empirically, Fast-BEV++ establishes a new state-of-the-art result of 0.488 NDS on the nuScenes 3D object detection benchmark, simultaneously delivering real-time inference at more than 134 FPS via our acceleration design. In particular, our integrated, learnable depth module yields consistent performance gains, maintaining the highest accuracy among comparable methods. Overall, this inherently decomposed architecture enables seamless real-time deployment across diverse production-grade automotive platforms, alleviating hardware limitations without compromising perception accuracy or inference efficiency.", "date": "2025-12-09", "categories": []} {"id": "2603.19538", "title": "MoCA3D: Monocular 3D Bounding Box Prediction in the Image Plane", "abstract": "Monocular 3D object understanding has largely been cast as a 2D RoI-to-3D box lifting problem. However, emerging downstream applications require image-plane geometry (e.g., projected 3D box corners) which cannot be easily obtained without known intrinsics, a problem for object detection in the wild. We introduce MoCA3D, a Monocular, Class-Agnostic 3D model that predicts projected 3D bounding box corners and per-corner depths without requiring camera intrinsics at inference time. MoCA3D formulates pixel-space localization and depth assignment as dense prediction via corner heatmaps and depth maps. To evaluate image-plane geometric fidelity, we propose Pixel-Aligned Geometry (PAG), which directly measures image-plane corner and depth consistency. Extensive experiments demonstrate that MoCA3D achieves state-of-the-art performance, improving image-plane corner PAG by 22.8% while remaining comparable on 3D IoU, using up to 57 times fewer trainable parameters. Finally, we apply MoCA3D to downstream tasks which were previously impractical under unknown intrinsics, highlighting its utility beyond standard baseline models.", "date": "2026-03-20", "categories": []} {"id": "2605.15546", "title": "3DTMDet: A Dual-Path Synergy Network of Transformer and SSM for 3D Object Detection in Point Clouds", "abstract": "A fundamental challenge in point cloud object detection lies in the conflict between the extreme sparsity of distant points and the need for remote context understanding. The existing methods typically use 1D serialization to expand the receptive field, which inevitably discards already scarce local geometric details and reduces detection of distant and small objects. To address this issue, we propose 3DTMDet, a novel detection network that synergistically combines state space models (Mamba) with Transformers. The core idea is to utilize SSM's linear complexity and advantages in long sequence modeling to effectively capture global interactions between sparse and distant points, while using Transformer modules with local attention to encode fine-grained geometric structures in local point sets, preserving accurate shape information. We propose the 3D Hybrid Mamba Transformer (3DHMT) block, which uses an SSM-Attention-SSM pipeline to balance global context understanding and local detail preservation, effectively alleviating the tension between receptive field enlargement and geometric preservation in remote detection. In addition, we introduced a voxel generation block inspired by LiDAR physics, which diffuses features along the sensor observation direction to reconstruct the complete object structure of occlusion and distant areas. Extensive experiments conducted on the KITTI and ONCE datasets have shown that 3DTMDet outperforms state-of-the-art detectors. The code is available at https://github.com/QiuBingwen/3DTMDet.", "date": "2026-05-15", "categories": []} {"id": "2506.04499", "title": "FALO: Fast and Accurate LiDAR 3D Object Detection on Resource-Constrained Devices", "abstract": "Existing LiDAR 3D object detection methods predominantely rely on sparse convolutions and/or transformers, which can be challenging to run on resource-constrained edge devices, due to irregular memory access patterns and high computational costs. In this paper, we propose FALO, a hardware-friendly approach to LiDAR 3D detection, which offers both state-of-the-art (SOTA) detection accuracy and fast inference speed. More specifically, given the 3D point cloud and after voxelization, FALO first arranges sparse 3D voxels into a 1D sequence based on their coordinates and proximity. The sequence is then processed by our proposed ConvDotMix blocks, consisting of large-kernel convolutions, Hadamard products, and linear layers. ConvDotMix provides sufficient mixing capability in both spatial and embedding dimensions, and introduces higher-order nonlinear interaction among spatial features. Furthermore, when going through the ConvDotMix layers, we introduce implicit grouping, which balances the tensor dimensions for more efficient inference and takes into account the growing receptive field. All these operations are friendly to run on resource-constrained platforms and proposed FALO can readily deploy on compact, embedded devices. Our extensive evaluation on LiDAR 3D detection benchmarks such as nuScenes and Waymo shows that FALO achieves competitive performance. Meanwhile, FALO is 1.6~9.8x faster than the latest SOTA on mobile Graphics Processing Unit (GPU) and mobile Neural Processing Unit (NPU).", "date": "2025-06-04", "categories": []} {"id": "2605.14110", "title": "SToRe3D: Sparse Token Relevance in ViTs for Efficient Multi-View 3D Object Detection", "abstract": "Vision Transformers (ViTs) enable strong multi-view 3D detection but are limited by high inference latency from dense token and query processing across multiple views and large 3D regions. Existing sparsity methods, designed mainly for 2D vision, prune or merge image tokens but do not extend to full-model sparsity or address 3D object queries. We introduce SToRe3D, a relevance-aligned sparsity framework that jointly selects 2D image tokens and 3D object queries while storing filtered features for reactivation. Mutual 2D-3D relevance heads allocate compute to driving-critical content and preserve other embeddings. Evaluated on nuScenes and our new nuScenes-Relevance benchmark, SToRe3D achieves up to 3x faster inference with marginal accuracy loss, establishing real-time large-scale ViT-based 3D detection while maintaining accuracy on planning-critical agents.", "date": "2026-05-13", "categories": []} {"id": "2605.14601", "title": "Towards Accurate Single Panoramic 3D Detection: A Semantic Gaussian Centric Approach", "abstract": "Three-dimensional object detection in panoramic imagery is crucial for comprehensive scene understanding, yet accurately mapping 2D features to 3D remains a significant challenge. Prevailing methods often project 2D features onto discrete 3D grids, which break geometric continuity and limit representation efficiency. To overcome this limitation, this paper proposes PanoGSDet, a monocular panoramic 3D detection framework built upon continuous semantic 3D Gaussian representations. The proposed framework comprises a panoramic depth estimation component and a semantic Gaussian component. The panoramic depth estimation component extracts the equirectangular semantic and depth features from the monocular panorama input. The semantic Gaussian component includes a semantic Gaussian lifting module that projects spherical features into 3D semantic Gaussians, a semantic Gaussian optimization module that refines these semantic Gaussians, and a Gaussian guided prediction head that generates 3D bounding boxes from optimized Gaussian representations. Extensive experiments on the Structured3D dataset demonstrate that our method significantly outperforms existing methods.", "date": "2026-05-14", "categories": []} {"id": "2605.14781", "title": "MonoPRIO: Adaptive Prior Conditioning for Unified Monocular 3D Object Detection", "abstract": "Monocular 3D object detection remains challenging because metric size and depth are underdetermined by single-view evidence, particularly under occlusion, truncation, and projection-induced scale-depth ambiguity. Although recent methods improve depth and geometric reasoning, metric size remains unstable in unified multi-class settings, where class variability and partial visibility broaden plausible size modes. We propose MonoPRIO, a unified monocular 3D detector that targets this bottleneck through adaptive prior conditioning in the size pathway. MonoPRIO constructs class-aware size prototypes offline, routes each decoder query to a soft mixture prior, applies uncertainty-aware log-space conditioning, and uses Cluster-Aligned Prior (CAP) regularisation on matched positives during training. On the official KITTI test server, MonoPRIO achieves the strongest fully reported unified multi-class result among methods reporting complete Car, Pedestrian, and Cyclist metrics. In the car-only setting, it also achieves the strongest 3D bounding-box AP across Easy/Moderate/Hard categories among compared methods without extra data, while using substantially less compute than MonoCLUE. Ablations and diagnostics show complementary gains from routed injection and CAP, with the largest benefits in ambiguity-prone, partially occluded, and low-data regimes. These findings indicate that adaptive priors are most effective when image evidence underdetermines metric size, while atypical geometry or extreme visibility loss can still cause mismatch between routed priors and true instance geometry. Code, trained models, result logs, and reproducibility material are available at https://github.com/bigggs/MonoPRIO.", "date": "2026-05-14", "categories": []} {"id": "2411.13311", "title": "A Resource Efficient Fusion Network for Object Detection in Bird's-Eye View using Camera and Raw Radar Data", "abstract": "Cameras can be used to perceive the environment around the vehicle, while affordable radar sensors are popular in autonomous driving systems as they can withstand adverse weather conditions unlike cameras. However, radar point clouds are sparser with low azimuth and elevation resolution that lack semantic and structural information of the scenes, resulting in generally lower radar detection performance. In this work, we directly use the raw range-Doppler (RD) spectrum of radar data, thus avoiding radar signal processing. We independently process camera images within the proposed comprehensive image processing pipeline. Specifically, first, we transform the camera images to Bird's-Eye View (BEV) Polar domain and extract the corresponding features with our camera encoder-decoder architecture. The resultant feature maps are fused with Range-Azimuth (RA) features, recovered from the RD spectrum input from the radar decoder to perform object detection. We evaluate our fusion strategy with other existing methods not only in terms of accuracy but also on computational complexity metrics on RADIal dataset.", "date": "2024-11-20", "categories": []} {"id": "2605.11799", "title": "SB-BEVFusion: Enhancing the Robustness against Sensor Malfunction and Corruptions", "abstract": "Multimodal sensor fusion has demonstrated remarkable performance improvements over unimodal approaches in 3D object detection for autonomous vehicles. Typically, existing methods transform multimodal data from independent sensors, such as camera and LiDAR, into a unified bird's-eye view (BEV) representation for fusion. Although effective in ideal conditions, this strategy suffers from substantial performance deterioration when camera or LiDAR data are missing, corrupted, or noisy. To address this vulnerability, we develop a framework-agnostic fusion module for camera and LiDAR data that allows for handling cases when one of the two modalities is missing or corrupted. To demonstrate the effectiveness of our module, we instantiate it in BEVFusion [1], a well-established framework to combine camera and LiDAR data for 3D object detection. By means of quantitative experiments on the MultiCorrupt dataset, we demonstrate that our module achieves favorable performance improvements under scenarios of missing and corrupted modalities, substantially outperforming existing unified representation approaches across a wide range of sensor deterioration scenarios and reaching state-of-the-art performance in scenarios of corrupted modality due to extreme weather conditions and sensor failure.", "date": "2026-05-12", "categories": []} {"id": "2603.11566", "title": "R4Det: 4D Radar-Camera Fusion for High-Performance 3D Object Detection", "abstract": "4D radar-camera sensing configuration has gained increasing importance in autonomous driving. However, existing 3D object detection methods that fuse 4D Radar and camera data confront several challenges. First, their absolute depth estimation module is not robust and accurate enough, leading to inaccurate 3D localization. Second, the performance of their temporal fusion module will degrade dramatically or even fail when the ego vehicle's pose is missing or inaccurate. Third, for some small objects, the sparse radar point clouds may completely fail to reflect from their surfaces. In such cases, detection must rely solely on visual unimodal priors. To address these limitations, we propose R4Det, which enhances depth estimation quality via the Panoramic Depth Fusion module, enabling mutual reinforcement between absolute and relative depth. For temporal fusion, we design a Deformable Gated Temporal Fusion module that does not rely on the ego vehicle's pose. In addition, we built an Instance-Guided Dynamic Refinement module that extracts semantic prototypes from 2D instance guidance. Experiments show that R4Det achieves state-of-the-art 3D object detection results on the TJ4DRadSet and VoD datasets. The source code and models will be released at https://github.com/VDIGPKU/R4Det.", "date": "2026-03-12", "categories": []} {"id": "2605.09899", "title": "Hyperbolic Distillation: Geometry-Guided Cross-Modal Transfer for Robust 3D Object Detection", "abstract": "Cross-modal knowledge distillation has emerged as an effective strategy for integrating point cloud and image features in 3D perception tasks. However, the modality heterogeneity, spatial misalignment, and the representation crisis of multiple modalities often limit the efficient of these cross-modal distillation methods. To address these limitations in existing approaches, we propose a hyperbolic constrained cross-modal distillation method for multimodal 3D object detection (HGC-Det). The proposed HGC-Det framework includes an image branch and a point cloud branch to extract semantic features from two different modalities. The point cloud branch comprises three core components: a 2D semantic-guided voxel optimization component (SGVO), a hyperbolic geometry constrained cross-modal feature transfer component (HFT), and a feature aggregation-based geometry optimization component (FAGO). Specifically, the SGVO component adaptively refines the spatial representation of the 3D branch by leveraging semantic cues from the image branch, thereby mitigating the issue of inadequate representation fusion. The HFT component exploits the intrinsic geometric properties of hyperbolic space to alleviate semantic loss during the fusion of high-dimensional image features and low-dimensional point cloud features. Finally, the FAGO compensates for potential spatial feature degradation introduced by the 2D semantic-guided voxel optimization component. Extensive experiments on indoor datasets (SUN RGB-D, ARKitScenes) and outdoor datasets (KITTI, nuScenes) demonstrate that our method achieves a better trade-off between detection accuracy and computational cost.", "date": "2026-05-11", "categories": []} {"id": "2605.10026", "title": "MUSDA: Multi-source Multi-modality Unsupervised Domain Adaptive 3D Object Detection for Autonomous Driving", "abstract": "With the advancement of autonomous driving, numerous annotated multi-modality datasets have become available. This presents an opportunity to develop domain-adaptive 3D object detectors for new environments without relying on labor-intensive manual annotations. However, traditional domain adaptation methods typically focus on a single source domain or a single modality, limiting their effectiveness in multi-source, multi-modality scenarios. In this paper, we propose a novel framework for multi-source, multi-modality unsupervised domain adaptation in 3D object detection for autonomous driving. Given multiple labeled source domains and one unlabeled target domain, our framework first introduces hierarchical spatially-conditioned (HSC) domain classifiers, which jointly align features from both camera and LiDAR modalities at two distinct levels for each source-target domain pair. To effectively leverage information from multiple source domains, we construct a prototype graph between each pair of domains. Based on this, we develop a prototype graph weighted (PGW) multi-source fusion strategy to aggregate predictions from multiple source detection heads. Experimental results on three widely used 3D object detection datasets - Waymo, nuScenes, and Lyft - demonstrate that our proposed framework effectively integrates information across both modalities and source domains, consistently outperforming state-of-the-art methods.", "date": "2026-05-11", "categories": []} {"id": "2512.02991", "title": "GraphFusion3D: Dynamic Graph Attention Convolution with Adaptive Cross-Modal Transformer for 3D Object Detection", "abstract": "Despite significant progress in 3D object detection, point clouds remain challenging due to sparse data, incomplete structures, and limited semantic information. Capturing contextual relationships between distant objects presents additional difficulties. To address these challenges, we propose GraphFusion3D, a unified framework combining multi-modal fusion with advanced feature learning. Our approach introduces the Adaptive Cross-Modal Transformer (ACMT), which adaptively integrates image features into point representations to enrich both geometric and semantic information. For proposal refinement, we introduce the Graph Reasoning Module (GRM), a novel mechanism that models neighborhood relationships to simultaneously capture local geometric structures and global semantic context. The module employs multi-scale graph attention to dynamically weight both spatial proximity and feature similarity between proposals. We further employ a cascade decoder that progressively refines detections through multi-stage predictions. Extensive experiments on SUN RGB-D (70.6% AP$_{25}$ and 51.2% AP$_{50}$) and ScanNetV2 (75.1% AP$_{25}$ and 60.8% AP$_{50}$) demonstrate a substantial performance improvement over existing approaches.", "date": "2025-12-02", "categories": []} {"id": "2605.05328", "title": "Query2Uncertainty: Robust Uncertainty Quantification and Calibration for 3D Object Detection under Distribution Shift", "abstract": "Reliable uncertainty estimation for 3D object detection is critical for deploying safe autonomous systems, yet modern detectors remain poorly calibrated, especially under distribution shifts. Although post-hoc calibration methods address this issue and provide improved calibration for in-distribution tests, they fail to adapt in distribution-shifted scenarios. In this work, we address this issue and introduce a density-aware calibration method that couples post-hoc calibrators with the feature density of latent object queries from DETR-style 3D object detectors. These queries form a compact, location and class-aware feature, ideal for density estimation, allowing our approach to adjust model confidences in distribution-shift scenarios. By fitting a density estimator on these query features, our approach jointly recalibrates both classification and bounding box regression uncertainties. On both a multi-view camera and LiDAR-based detector, our approach consistently outperforms standard post-hoc methods in both in-distribution and distribution-shifted scenarios. Code available https://tillbeemelmanns.github.io/query2uncertainty/ .", "date": "2026-05-06", "categories": []} {"id": "2605.03437", "title": "Learning Discriminative Signed Distance Functions from Multi-scale Level-of-detail Features for 3D Anomaly Detection", "abstract": "Detecting anomalies from 3D point clouds has received increasing attention in the field of computer vision, with some group-based or point-based methods achieving impressive results in recent years. However, learning accurate point-wise representations for 3D anomaly detection faces great challenges due to the large scale and sparsity of point clouds. In this study, a surface-based method is proposed for 3D anomaly detection, which learns a discriminative signed distance function using multi-scale level-of-detail features. We first present a Noisy Points Generation (NPG) module to generate different types of noise, thereby facilitating the learning of discriminative features by exposing abnormal points. Then, we introduce a Multi-scale Level-of-detail Feature (MLF) module to capture multi-scale information from a point cloud, which provides both fine-grained local and coarse-grained global feature information. Finally, we design an Implicit Surface Discrimination (ISD) module that leverages the extracted multi-scale features to learn an implicit surface representation of point clouds, which effectively trains a signed distance function to distinguish between abnormal and normal points. Experimental results demonstrate that the proposed method achieves an average object-level AUROC of 92.1\\% and 85.9\\% on the Anomaly-ShapeNet and Real3D-AD datasets, outperforming the current best approach by 2.1\\% and 3.6\\%, respectively. Codes are available at https://anonymous.4open.science/r/DLF-3AD-DA61.", "date": "2026-05-05", "categories": []} {"id": "2605.01924", "title": "SimPB++: Simultaneously Detecting 2D and 3D Objects from Multiple Cameras", "abstract": "Simultaneous perception of 2D objects in perspective view and 3D objects in Bird's Eye View (BEV) is challenging for multi-camera autonomous driving. Existing two-stage pipelines use 2D results only as a one-time cue for 3D detection. We propose SimPB++, which simultaneously detects 2D objects in perspective and 3D objects in BEV from multiple cameras. It unifies both tasks into an end-to-end model with a hybrid decoder architecture, coupling multi-view 2D and 3D decoders interactively. Two novel modules enable deep interaction: Dynamic Query Allocation adaptively assigns 2D queries to 3D candidates, and Adaptive Query Aggregation refines 3D representations using multi-view 2D features, forming a cyclic 3D-2D-3D refinement. For multi-view 2D detection, we use Query-group Attention for intra-group communication. We also design a Crop-and-Scale strategy for long-range perception and a Propagating Denoising strategy with an auxiliary RoI detector. SimPB++ supports mixed supervision with 2D-only and fully annotated data, reducing reliance on expensive 3D labels. Experiments show state-of-the-art performance on nuScenes for both tasks and strong long-range detection (up to 150m) on Argoverse2.", "date": "2026-05-03", "categories": []} {"id": "2605.00595", "title": "Robust Fusion of Object-Level V2X for Learned 3D Object Detection", "abstract": "Perception for automated driving is largely based on onboard environmental sensors, such as cameras and radar, which are cost-effective but limited by line-of-sight and field-of-view constraints. These inherent limitations may cause onboard perception to fail under occlusions or poor visibility conditions. In parallel, cooperative awareness via vehicle-to-everything (V2X) communication is becoming increasingly available, enabling vehicles and infrastructure to share their own state as object-level information that complements onboard perception. In this work, we study how such V2X information can be integrated into 3D object detection and how robust the resulting system is to realistic V2X imperfections. Using the nuScenes dataset, we emulate object-level cooperative awareness messages from ground truth, injecting controlled noise and object dropout to mimic real-world conditions such as latency, localization errors, and low V2X penetration rates. We convert these messages into a dedicated bird's-eye view (BEV) input and fuse them into a BEVFusion-style detector. Our results demonstrate that while object-level cooperative information can substantially improve detection performance, achieving an NDS of 0.80 under favorable conditions, models trained on idealized data become fragile and over-reliant on V2X. Conversely, our proposed noise-aware training strategy, coupled with explicit confidence encoding, enhances robustness, maintaining performance gains even under severe noise and reduced V2X penetration.", "date": "2026-05-01", "categories": []} {"id": "2604.25405", "title": "Leveraging Previous-Traversal Point Cloud Map Priors for Camera-Based 3D Object Detection and Tracking", "abstract": "Camera-based 3D object detection and tracking are central to autonomous driving, yet precise 3D object localization remains fundamentally constrained by depth ambiguity when no expensive, depth-rich online LiDAR is available at inference. In many deployments, however, vehicles repeatedly traverse the same environments, making static point cloud maps from prior traversals a practical source of geometric priors. We propose DualViewMapDet, a camera-only inference framework that retrieves such map priors online and leverages them to mitigate the absence of a LiDAR sensor during deployment. The key idea is a dual-space camera-map fusion strategy that avoids one-sided view conversion. Specifically, we (i) project the map into perspective view (PV) and encode multi-channel geometric cues to enrich image features and support BEV lifting, and (ii) encode the map directly in bird's-eye view (BEV) with a sparse voxel backbone and fuse it with lifted camera features in a shared metric space. Extensive evaluations on nuScenes and Argoverse 2 demonstrate consistent improvements over strong camera-only baselines, with particularly strong gains in object localization. Ablations further validate the contributions of PV/BEV fusion and prior-map coverage. We make the code and pre-trained models available at https://dualviewmapdet.cs.uni-freiburg.de .", "date": "2026-04-28", "categories": []} {"id": "2604.25574", "title": "Control Your Queries: Heterogeneous Query Interaction for Camera-Radar Fusion", "abstract": "In autonomous driving, camera-radar fusion offers complementary sensing and low deployment cost. Existing methods perform fusion through input mixing, feature map mixing, or query-based feature sampling. We propose a new fusion paradigm, termed heterogeneous query interaction, and present ConFusion, a camera-radar 3D object detector. ConFusion combines image queries, radar queries, and learnable world queries distributed in 3D space to improve query initialization and object coverage. To encourage cross-type interaction among heterogeneous queries, we introduce heterogeneous query mixing (QMix), which performs dedicated cross-type attention after feature sampling to consolidate complementary object evidence. We further propose interactive query swap sampling (QSwap), which improves feature sampling by allowing related queries to exchange informative feature tokens under attention and geometric constraints. Experiments on the nuScenes dataset show that ConFusion achieves state-of-the-art performance, reaching 59.1 mAP and 65.6 NDS on the validation set, and 61.6 mAP and 67.9 NDS on the test set.", "date": "2026-04-28", "categories": []} {"id": "2604.24044", "title": "CLLAP: Contrastive Learning-based LiDAR-Augmented Pretraining for Enhanced Radar-Camera Fusion", "abstract": "Accurate 3D object detection is critical for autonomous driving, necessitating reliable, cost-effective sensors capable of operating in adverse weather conditions. Camera and millimeter-wave radar fusion has emerged as a promising solution; however, these methods often rely on finely annotated radar data, which is scarce and labor-intensive to produce. To address this challenge, we present CLLAP, a Contrastive Learning-based LiDAR-Augmented Pretraining framework that enhances the performance of existing radar-camera fusion methods for 3D object detection. CLLAP leverages abundant LiDAR data to generate pseudo-radar data using the proposed L2R (LiDAR-to-Radar) Sampling method. Then, it incorporates this data into a novel dual-stage, dual-modality contrastive learning strategy, enabling effective self-supervised learning from paired pseudo-radar and image data. This approach facilitates effective pretraining of existing radar-camera fusion models in a plug-and-play manner, enhancing their feature extraction capabilities and improving detection accuracy and robustness. Experimental results using NuScenes and Lyft Level 5 datasets demonstrate significant performance improvements across three baseline models, highlighting CLLAP's effectiveness in advancing radar-camera fusion for autonomous driving applications.", "date": "2026-04-27", "categories": []} {"id": "2604.24718", "title": "WildLIFT: Lifting monocular drone video to 3D for species-agnostic wildlife monitoring", "abstract": "Monocular RGB cameras mounted on drones are widely used for wildlife monitoring, yet most analytical pipelines remain confined to two-dimensional image space, leaving geometric information in video underexploited. We present WildLIFT, a computational framework that integrates three-dimensional scene geometry from monocular drone video with open-vocabulary 2D instance segmentation to enable species-agnostic 3D detection and tracking. Oriented 3D bounding box labels with semantic face information enable quantitative assessment of viewpoint coverage and inter-animal occlusion, producing structured metadata for downstream ecological analyses. We validate the framework on 2,581 manually curated frames comprising over 6,700 3D detections across four large mammal species. WildLIFT maintains high identity consistency in multi-animal scenes and substantially reduces manual 3D annotation effort through keyframe-based refinement. By transforming standard drone footage into structured 3D and viewpoint-aware representations, WildLIFT extends the analytical utility of aerial wildlife datasets for behavioural research and population monitoring.", "date": "2026-04-27", "categories": []} {"id": "2604.08626", "title": "WildDet3D: Scaling Promptable 3D Detection in the Wild", "abstract": "Understanding objects in 3D from a single image is a cornerstone of spatial intelligence. A key step toward this goal is monocular 3D object detection--recovering the extent, location, and orientation of objects from an input RGB image. To be practical in the open world, such a detector must generalize beyond closed-set categories, support diverse prompt modalities, and leverage geometric cues when available. Progress is hampered by two bottlenecks: existing methods are designed for a single prompt type and lack a mechanism to incorporate additional geometric cues, and current 3D datasets cover only narrow categories in controlled environments, limiting open-world transfer. In this work we address both gaps. First, we introduce WildDet3D, a unified geometry-aware architecture that natively accepts text, point, and box prompts and can incorporate auxiliary depth signals at inference time. Second, we present WildDet3D-Data, the largest open 3D detection dataset to date, constructed by generating candidate 3D boxes from existing 2D annotations and retaining only human-verified ones, yielding over 1M images across 13.5K categories in diverse real-world scenes. WildDet3D establishes a new state-of-the-art across multiple benchmarks and settings. In the open-world setting, it achieves 22.6/24.8 AP3D on our newly introduced WildDet3D-Bench with text and box prompts. On Omni3D, it reaches 34.2/36.4 AP3D with text and box prompts, respectively. In zero-shot evaluation, it achieves 40.3/48.9 ODS on Argoverse 2 and ScanNet. Notably, incorporating depth cues at inference time yields substantial additional gains (+20.7 AP on average across settings).", "date": "2026-04-09", "categories": []} {"id": "2604.16696", "title": "LOD-Net: Locality-Aware 3D Object Detection Using Multi-Scale Transformer Network", "abstract": "3D object detection in point cloud data remains a challenging task due to the sparsity and lack of global structure inherent in the input. In this work, we propose a novel Multi-Scale Attention (MSA) mechanism integrated into the 3DETR architecture to better capture both local geometry and global context. Our method introduces an upsampling operation that generates high-resolution feature maps, enabling the network to better detect smaller and semantically related objects. Experiments conducted on the ScanNetv2 dataset demonstrate that our 3DETR + MSA model improves detection performance, achieving a gain of almost 1% in mAP@25 and 4.78% in mAP@50 over the baseline. While applying MSA to the 3DETR-m variant shows limited improvement, our analysis reveals the importance of adapting the upsampling strategy for lightweight models. These results highlight the effectiveness of combining hierarchical feature extraction with attention mechanisms in enhancing 3D scene understanding.", "date": "2026-04-17", "categories": []} {"id": "2604.17024", "title": "CAM3DNet: Comprehensively mining the multi-scale features for 3D Object Detection with Multi-View Cameras", "abstract": "Query-based 3D object detection methods using multi-view images often struggle to efficiently leverage dynamic multi-scale information, e.g., the relationship between the object features and the geometric of the queries are not sufficiently learned, directly exploring the multi-scale spatiotemporal features will pay too many costs. To address these challenges, we propose CAM3DNet, a novel sparse query-based framework which combines three new modules, composite query (CQ), adaptive self-attention (ASA), and multi-scale hybrid sampling (MSHS). First, the core idea in the CQ module is a multi-scale projection strategy to transform 2D queries into 3D space. Second, the ASA module learns the interactions between the spatiotemporal multi-scale queries. Third, the MSHS module uses the deformable attention mechanism to sample multi-scale object information by considering multi-scales queries, pyramid feature maps, and 2D-camera prior knowledge. The entire model employs a backbone network and a feature pyramid network (FPN) as the encoder, then introduces a YOLOX and a DepthNet as a ROI\\_Head to produce CQ, and repeatedly utilizes ASA and MSHS as the decoder to gain detection features. Extensive experiments on the nuScenes, Waymo, and Argoverse benchmark datasets demonstrate the effectiveness of our CAM3DNet, and most existing camera-based 3D object detection methods are outperformed. Besides, we make comprehensive ablation studies to check the individual effect of CQ, ASA, and MSHS, as well as their cost of space and computation complexity.", "date": "2026-04-18", "categories": []} {"id": "2604.18476", "title": "SemLT3D: Semantic-Guided Expert Distillation for Camera-only Long-Tailed 3D Object Detection", "abstract": "Camera-only 3D object detection has emerged as a cost-effective and scalable alternative to LiDAR for autonomous driving, yet existing methods primarily prioritize overall performance while overlooking the severe long-tail imbalance inherent in real-world datasets. In practice, many rare but safety-critical categories such as children, strollers, or emergency vehicles are heavily underrepresented, leading to biased learning and degraded performance. This challenge is further exacerbated by pronounced inter-class ambiguity (e.g., visually similar subclasses) and substantial intra-class diversity (e.g., objects varying widely in appearance, scale, pose, or context), which together hinder reliable long-tail recognition. In this work, we introduce SemLT3D, a Semantic-Guided Expert Distillation framework designed to enrich the representation space for underrepresented classes through semantic priors. SemLT3D consists of: (1) a language-guided mixture-of-experts module that routes 3D queries to specialized experts according to their semantic affinity, enabling the model to better disentangle confusing classes and specialize on tail distributions; and (2) a semantic projection distillation pipeline that aligns 3D queries with CLIP-informed 2D semantics, producing more coherent and discriminative features across diverse visual manifestations. Although motivated by long-tail imbalance, the semantically structured learning in SemLT3D also improves robustness under broader appearance variations and challenging corner cases, offering a principled step toward more reliable camera-only 3D perception.", "date": "2026-04-20", "categories": []} {"id": "2604.15795", "title": "Fed3D: Federated 3D Object Detection", "abstract": "3D object detection models trained in one server plays an important role in autonomous driving, robotics manipulation, and augmented reality scenarios. However, most existing methods face severe privacy concern when deployed on a multi-robot perception network to explore large-scale 3D scene. Meanwhile, it is highly challenging to employ conventional federated learning methods on 3D object detection scenes, due to the 3D data heterogeneity and limited communication bandwidth. In this paper, we take the first attempt to propose a novel Federated 3D object detection framework (i.e., Fed3D), to enable distributed learning for 3D object detection with privacy preservation. Specifically, considering the irregular input 3D object in local robot and various category distribution between robots could cause local heterogeneity and global heterogeneity, respectively. We then propose a local-global class-aware loss for the 3D data heterogeneity issue, which could balance gradient back-propagation rate of different 3D categories from local and global aspects. To reduce communication cost on each round, we develop a federated 3D prompt module, which could only learn and communicate the prompts with few learnable parameters. To the end, several extensive experiments on federated 3D object detection show that our Fed3D model significantly outperforms state-of-the-art algorithms with lower communication cost when providing the limited local training data.", "date": "2026-04-17", "categories": []} {"id": "2604.14563", "title": "Revisiting Token Compression for Accelerating ViT-based Sparse Multi-View 3D Object Detectors", "abstract": "Vision Transformer (ViT)-based sparse multi-view 3D object detectors have achieved remarkable accuracy but still suffer from high inference latency due to heavy token processing. To accelerate these models, token compression has been widely explored. However, our revisit of existing strategies, such as token pruning, merging, and patch size enlargement, reveals that they often discard informative background cues, disrupt contextual consistency, and lose fine-grained semantics, negatively affecting 3D detection. To overcome these limitations, we propose SEPatch3D, a novel framework that dynamically adjusts patch sizes while preserving critical semantic information within coarse patches. Specifically, we design Spatiotemporal-aware Patch Size Selection (SPSS) that assigns small patches to scenes containing nearby objects to preserve fine details and large patches to background-dominated scenes to reduce computation cost. To further mitigate potential detail loss, Informative Patch Selection (IPS) selects the informative patches for feature refinement, and Cross-Granularity Feature Enhancement (CGFE) injects fine-grained details into selected coarse patches, enriching semantic features. Experiments on the nuScenes and Argoverse 2 validation sets show that SEPatch3D achieves up to \\textbf{57\\%} faster inference than the StreamPETR baseline and \\textbf{20\\%} higher efficiency than the state-of-the-art ToC3D-faster, while preserving comparable detection accuracy. Code is available at https://github.com/Mingqj/SEPatch3D.", "date": "2026-04-16", "categories": []} {"id": "2604.13586", "title": "Efficient Multi-View 3D Object Detection by Dynamic Token Selection and Fine-Tuning", "abstract": "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}.", "date": "2026-04-15", "categories": []} {"id": "2602.21484", "title": "Unified Unsupervised and Sparsely-Supervised 3D Object Detection by Semantic Pseudo-Labeling and Prototype Learning", "abstract": "3D object detection is essential for autonomous driving and robotic perception, yet its reliance on large-scale manually annotated data limits scalability and adaptability. To reduce annotation dependency, unsupervised and sparsely-supervised paradigms have emerged. However, they face intertwined challenges: low-quality pseudo-labels, unstable feature mining, and a lack of a unified training framework. This paper proposes SPL, a unified training framework for both unsupervised and sparsely-supervised 3D object detection via \\underline{S}emantic \\underline{P}seudo-labeling and prototype \\underline{L}earning. SPL first generates high-quality pseudo-labels by integrating image semantics, point cloud geometry, and temporal cues, producing both 3D bounding boxes for dense objects and 3D point labels for sparse ones. These pseudo-labels are not used directly but as probabilistic priors within a novel, multi-stage prototype learning strategy. This strategy stabilizes feature representation learning through memory-based initialization and momentum-based prototype updating, effectively mining features from both labeled and unlabeled data. Extensive experiments on KITTI and nuScenes datasets demonstrate that SPL significantly outperforms state-of-the-art methods in both settings. Our work provides a robust and generalizable solution for learning 3D object detectors with minimal or no manual annotations. Our code is available at https://github.com/TossherO/SPL.", "date": "2026-02-25", "categories": []} {"id": "2604.10305", "title": "Class-Adaptive Cooperative Perception for Multi-Class LiDAR-based 3D Object Detection in V2X Systems", "abstract": "Cooperative perception allows connected vehicles and roadside infrastructure to share sensor observations, creating a fused scene representation beyond the capability of any single platform. However, most cooperative 3D object detectors use a uniform fusion strategy for all object classes, which limits their ability to handle the different geometric structures and point-sampling patterns of small and large objects. This problem is further reinforced by narrow evaluation protocols that often emphasize a single dominant class or only a few cooperation settings, leaving robust multi-class detection across diverse vehicle-to-everything interactions insufficiently explored. To address this gap, we propose a class-adaptive cooperative perception architecture for multi-class 3D object detection from LiDAR data. The model integrates four components: multi-scale window attention with learned scale routing for spatially adaptive feature extraction, a class-specific fusion module that separates small and large objects into attentive fusion pathways, bird's-eye-view enhancement through parallel dilated convolution and channel recalibration for richer contextual representation, and class-balanced objective weighting to reduce bias toward frequent categories. Experiments on the V2X-Real benchmark cover vehicle-centric, infrastructure-centric, vehicle-to-vehicle, infrastructure-to-infrastructure, and vehicle-to-infrastructure settings under identical backbone and training configurations. The proposed method consistently improves mean detection performance over strong intermediate-fusion baselines, with the largest gains on trucks, clear improvements on pedestrians, and competitive results on cars. These results show that aligning feature extraction and fusion with class-dependent geometry and point density leads to more balanced cooperative perception in realistic vehicle-to-everything deployments.", "date": "2026-04-11", "categories": []} {"id": "2604.11400", "title": "EagleVision: A Multi-Task Benchmark for Cross-Domain Perception in High-Speed Autonomous Racing", "abstract": "High-speed autonomous racing presents extreme perception challenges, including large relative velocities and substantial domain shifts from conventional urban-driving datasets. Existing benchmarks do not adequately capture these high-dynamic conditions. We introduce EagleVision, a unified LiDAR-based multi-task benchmark for 3D detection and trajectory prediction in high-speed racing, providing newly annotated 3D bounding boxes for the Indy Autonomous Challenge dataset (14,893 frames) and the A2RL Real competition dataset (1,163 frames), together with 12,000 simulator-generated annotated frames, all standardized under a common evaluation protocol. Using a dataset-centric transfer framework, we quantify cross-domain generalization across urban, simulator, and real racing domains. Urban pretraining improves detection over scratch training (NDS 0.72 vs. 0.69), while intermediate pretraining on real racing data achieves the best transfer to A2RL (NDS 0.726), outperforming simulator-only adaptation. For trajectory prediction, Indy-trained models surpass in-domain A2RL training on A2RL test sequences (FDE 0.947 vs. 1.250), highlighting the role of motion-distribution coverage in cross-domain forecasting. EagleVision enables systematic study of perception generalization under extreme high-speed dynamics. The dataset and benchmark are publicly available at https://avlab.io/EagleVision", "date": "2026-04-13", "categories": []} {"id": "2604.09206", "title": "Long-SCOPE: Fully Sparse Long-Range Cooperative 3D Perception", "abstract": "Cooperative 3D perception via Vehicle-to-Everything communication is a promising paradigm for enhancing autonomous driving, offering extended sensing horizons and occlusion resolution. However, the practical deployment of existing methods is hindered at long distances by two critical bottlenecks: the quadratic computational scaling of dense BEV representations and the fragility of feature association mechanisms under significant observation and alignment errors. To overcome these limitations, we introduce Long-SCOPE, a fully sparse framework designed for robust long-distance cooperative 3D perception. Our method features two novel components: a Geometry-guided Query Generation module to accurately detect small, distant objects, and a learnable Context-Aware Association module that robustly matches cooperative queries despite severe positional noise. Experiments on the V2X-Seq and Griffin datasets validate that Long-SCOPE achieves state-of-the-art performance, particularly in challenging 100-150 m long-range settings, while maintaining highly competitive computation and communication costs.", "date": "2026-04-10", "categories": []} {"id": "2604.09210", "title": "Adding Another Dimension to Image-based Animal Detection", "abstract": "Monocular imaging of animals inherently reduces 3D structures to 2D projections. Detection algorithms lead to 2D bounding boxes that lack information about animal's orientation relative to the camera. To build 3D detection methods for RGB animal images, there is a lack of labeled datasets; such labeling processes require 3D input streams along with RGB data. We present a pipeline that utilises Skinned Multi Animal Linear models to estimate 3D bounding boxes and to project them as robust labels into 2D image space using a dedicated camera pose refinement algorithm. To assess which sides of the animal are captured, cuboid face visibility metrics are computed. These 3D bounding boxes and metrics form a crucial step toward developing and benchmarking future monocular 3D animal detection algorithms. We evaluate our method on the Animal3D dataset, demonstrating accurate performance across species and settings.", "date": "2026-04-10", "categories": []} {"id": "2505.17732", "title": "RQR3D: Reparametrizing the regression targets for BEV-based 3D object detection", "abstract": "Accurate, fast, and reliable 3D perception is essential for autonomous driving. Recently, bird's-eye view (BEV)-based perception approaches have emerged as superior alternatives to perspective-based solutions, offering enhanced spatial understanding and more natural outputs for planning. Existing BEV-based 3D object detection methods, typically using an angle-based representation, directly estimate the size and orientation of rotated bounding boxes. We observe that BEV-based 3D object detection is analogous to aerial oriented object detection, where angle-based methods are known to suffer from discontinuities in their loss functions. Drawing inspiration from this domain, we propose \\textbf{R}estricted \\textbf{Q}uadrilateral \\textbf{R}epresentation to define \\textbf{3D} regression targets. RQR3D regresses the smallest horizontal bounding box encapsulating the oriented box, along with the offsets between the corners of these two boxes, thereby transforming the oriented object detection problem into a keypoint regression task. We employ RQR3D within an anchor-free single-stage object detection method achieving state-of-the-art performance. We show that the proposed architecture is compatible with different object detection approaches. Furthermore, we introduce a simplified radar fusion backbone that applies standard 2D convolutions to radar features. This backbone leverages the inherent 2D structure of the data for efficient and geometrically consistent processing without over-parameterization, thereby eliminating the need for voxel grouping and sparse convolutions. Extensive evaluations on the nuScenes dataset show that RQR3D achieves SotA camera-radar 3D object detection performance despite its lightweight design, reaching 67.5 NDS and 59.7 mAP with reduced translation and orientation errors, which are crucial for safe autonomous driving.", "date": "2025-05-23", "categories": []} {"id": "2604.07997", "title": "Few-Shot Incremental 3D Object Detection in Dynamic Indoor Environments", "abstract": "Incremental 3D object perception is a critical step toward embodied intelligence in dynamic indoor environments. However, existing incremental 3D detection methods rely on extensive annotations of novel classes for satisfactory performance. To address this limitation, we propose FI3Det, a Few-shot Incremental 3D Detection framework that enables efficient 3D perception with only a few novel samples by leveraging vision-language models (VLMs) to learn knowledge of unseen categories. FI3Det introduces a VLM-guided unknown object learning module in the base stage to enhance perception of unseen categories. Specifically, it employs VLMs to mine unknown objects and extract comprehensive representations, including 2D semantic features and class-agnostic 3D bounding boxes. To mitigate noise in these representations, a weighting mechanism is further designed to re-weight the contributions of point- and box-level features based on their spatial locations and feature consistency within each box. Moreover, FI3Det proposes a gated multimodal prototype imprinting module, where category prototypes are constructed from aligned 2D semantic and 3D geometric features to compute classification scores, which are then fused via a multimodal gating mechanism for novel object detection. As the first framework for few-shot incremental 3D object detection, we establish both batch and sequential evaluation settings on two datasets, ScanNet V2 and SUN RGB-D, where FI3Det achieves strong and consistent improvements over baseline methods. Code is available at https://github.com/zyrant/FI3Det.", "date": "2026-04-09", "categories": []} {"id": "2604.05212", "title": "Boxer: Robust Lifting of Open-World 2D Bounding Boxes to 3D", "abstract": "Detecting and localizing objects in space is a fundamental computer vision problem. While much progress has been made to solve 2D object detection, 3D object localization is much less explored and far from solved, especially for open-world categories. To address this research challenge, we propose Boxer, an algorithm to estimate static 3D bounding boxes (3DBBs) from 2D open-vocabulary object detections, posed images and optional depth either represented as a sparse point cloud or dense depth. At its core is BoxerNet, a transformer-based network which lifts 2D bounding box (2DBB) proposals into 3D, followed by multi-view fusion and geometric filtering to produce globally consistent de-duplicated 3DBBs in metric world space. Boxer leverages the power of existing 2DBB detection algorithms (e.g. DETIC, OWLv2, SAM3) to localize objects in 2D. This allows the main BoxerNet model to focus on lifting to 3D rather than detecting, ultimately reducing the demand for costly annotated 3DBB training data. Extending the CuTR formulation, we incorporate an aleatoric uncertainty for robust regression, a median depth patch encoding to support sparse depth inputs, and large-scale training with over 1.2 million unique 3DBBs. BoxerNet outperforms state-of-the-art baselines in open-world 3DBB lifting, including CuTR in egocentric settings without dense depth (0.532 vs. 0.010 mAP) and on CA-1M with dense depth available (0.412 vs. 0.250 mAP).", "date": "2026-04-06", "categories": []} {"id": "2604.05316", "title": "Indoor Asset Detection in Large Scale 360{\\deg} Drone-Captured Imagery via 3D Gaussian Splatting", "abstract": "We present an approach for object-level detection and segmentation of target indoor assets in 3D Gaussian Splatting (3DGS) scenes, reconstructed from 360{\\deg} drone-captured imagery. We introduce a 3D object codebook that jointly leverages mask semantics and spatial information of their corresponding Gaussian primitives to guide multi-view mask association and indoor asset detection. By integrating 2D object detection and segmentation models with semantically and spatially constrained merging procedures, our method aggregates masks from multiple views into coherent 3D object instances. Experiments on two large indoor scenes demonstrate reliable multi-view mask consistency, improving F1 score by 65% over state-of-the-art baselines, and accurate object-level 3D indoor asset detection, achieving an 11% mAP gain over baseline methods.", "date": "2026-04-07", "categories": []} {"id": "2604.05405", "title": "Weather-Conditioned Branch Routing for Robust LiDAR-Radar 3D Object Detection", "abstract": "Robust 3D object detection in adverse weather is highly challenging due to the varying reliability of different sensors. While existing LiDAR-4D radar fusion methods improve robustness, they predominantly rely on fixed or weakly adaptive pipelines, failing to dy-namically adjust modality preferences as environmental conditions change. To bridge this gap, we reformulate multi-modal perception as a weather-conditioned branch routing problem. Instead of computing a single fused output, our framework explicitly maintains three parallel 3D feature streams: a pure LiDAR branch, a pure 4D radar branch, and a condition-gated fusion branch. Guided by a condition token extracted from visual and semantic prompts, a lightweight router dynamically predicts sample-specific weights to softly aggregate these representations. Furthermore, to prevent branch collapse, we introduce a weather-supervised learning strategy with auxiliary classification and diversity regularization to enforce distinct, condition-dependent routing behaviors. Extensive experiments on the K-Radar benchmark demonstrate that our method achieves state-of-the-art performance. Furthermore, it provides explicit and highly interpretable insights into modality preferences, transparently revealing how adaptive routing robustly shifts reliance between LiDAR and 4D radar across diverse adverse-weather scenarios. The source code with be released.", "date": "2026-04-07", "categories": []} {"id": "2604.01646", "title": "MonoSAOD: Monocular 3D Object Detection with Sparsely Annotated Label", "abstract": "Monocular 3D object detection has achieved impressive performance on densely annotated datasets. However, it struggles when only a fraction of objects are labeled due to the high cost of 3D annotation. This sparsely annotated setting is common in real-world scenarios where annotating every object is impractical. To address this, we propose a novel framework for sparsely annotated monocular 3D object detection with two key modules. First, we propose Road-Aware Patch Augmentation (RAPA), which leverages sparse annotations by augmenting segmented object patches onto road regions while preserving 3D geometric consistency. Second, we propose Prototype-Based Filtering (PBF), which generates high-quality pseudo-labels by filtering predictions through prototype similarity and depth uncertainty. It maintains global 2D RoI feature prototypes and selects pseudo-labels that are both feature-consistent with learned prototypes and have reliable depth estimates. Our training strategy combines geometry-preserving augmentation with prototype-guided pseudo-labeling to achieve robust detection under sparse supervision. Extensive experiments demonstrate the effectiveness of the proposed method. The source code is available at https://github.com/VisualAIKHU/MonoSAOD .", "date": "2026-04-02", "categories": []} {"id": "2604.03685", "title": "DSERT-RoLL: Robust Multi-Modal Perception for Diverse Driving Conditions with Stereo Event-RGB-Thermal Cameras, 4D Radar, and Dual-LiDAR", "abstract": "In this paper, we present DSERT-RoLL, a driving dataset that incorporates stereo event, RGB, and thermal cameras together with 4D radar and dual LiDAR, collected across diverse weather and illumination conditions. The dataset provides precise 2D and 3D bounding boxes with track IDs and ego vehicle odometry, enabling fair comparisons within and across sensor combinations. It is designed to alleviate data scarcity for novel sensors such as event cameras and 4D radar and to support systematic studies of their behavior. We establish unified 3D and 2D benchmarks that enable direct comparison of characteristics and strengths across sensor families and within each family. We report baselines for representative single modality and multimodal methods and provide protocols that encourage research on different fusion strategies and sensor combinations. In addition, we propose a fusion framework that integrates sensor specific cues into a unified feature space and improves 3D detection robustness under varied weather and lighting.", "date": "2026-04-04", "categories": []} {"id": "2604.04797", "title": "Multi-Modal Sensor Fusion using Hybrid Attention for Autonomous Driving", "abstract": "Accurate 3D object detection for autonomous driving requires complementary sensors. Cameras provide dense semantics but unreliable depth, while millimeter-wave radar offers precise range and velocity measurements with sparse geometry. We propose MMF-BEV, a radar-camera BEV fusion framework that leverages deformable attention for cross-modal feature alignment on the View-of-Delft (VoD) 4D radar dataset [1]. MMF-BEV builds a BEVDepth [2] camera branch and a RadarBEVNet [3] radar branch, each enhanced with Deformable Self-Attention, and fuses them via a Deformable Cross-Attention module. We evaluate three configurations: camera-only, radar-only, and hybrid fusion. A sensor contribution analysis quantifies per-distance modality weighting, providing interpretable evidence of sensor complementarity. A two-stage training strategy - pre-training the camera branch with depth supervision, then jointly training radar and fusion modules stabilizes learning. Experiments on VoD show that MMF-BEV consistently outperforms unimodal baselines and achieves competitive results against prior fusion methods across all object classes in both the full annotated area and near-range Region of Interest.", "date": "2026-04-06", "categories": []} {"id": "2509.23880", "title": "Learning Adaptive Pseudo-Label Selection for Semi-Supervised 3D Object Detection", "abstract": "Semi-supervised 3D object detection (SS3DOD) aims to reduce costly 3D annotations utilizing unlabeled data. Recent studies adopt pseudo-label-based teacher-student frameworks and demonstrate impressive performance. The main challenge of these frameworks is in selecting high-quality pseudo-labels from the teacher's predictions. Most previous methods, however, select pseudo-labels by comparing confidence scores over thresholds manually set. The latest works tackle the challenge either by dynamic thresholding or refining the quality of pseudo-labels. Such methods still overlook contextual information e.g. object distances, classes, and learning states, and inadequately assess the pseudo-label quality using partial information available from the networks. In this work, we propose a novel SS3DOD framework featuring a learnable pseudo-labeling module designed to automatically and adaptively select high-quality pseudo-labels. Our approach introduces two networks at the teacher output level. These networks reliably assess the quality of pseudo-labels by the score fusion and determine context-adaptive thresholds, which are supervised by the alignment of pseudo-labels over GT bounding boxes. Additionally, we introduce a soft supervision strategy that can learn robustly under pseudo-label noises. This helps the student network prioritize cleaner labels over noisy ones in semi-supervised learning. Extensive experiments on the KITTI and Waymo datasets demonstrate the effectiveness of our method. The proposed method selects high-precision pseudo-labels while maintaining a wider coverage of contexts and a higher recall rate, significantly improving relevant SS3DOD methods.", "date": "2025-09-28", "categories": []} {"id": "2604.02903", "title": "RayMamba: Ray-Aligned Serialization for Long-Range 3D Object Detection", "abstract": "Long-range 3D object detection remains challenging because LiDAR observations become highly sparse and fragmented in the far field, making reliable context modeling difficult for existing detectors. To address this issue, recent state space model (SSM)-based methods have improved long-range modeling efficiency. However, their effectiveness is still limited by generic serialization strategies that fail to preserve meaningful contextual neighborhoods in sparse scenes. To address this issue, we propose RayMamba, a geometry-aware plug-and-play enhancement for voxel-based 3D detectors. RayMamba organizes sparse voxels into sector-wise ordered sequences through a ray-aligned serialization strategy, which preserves directional continuity and occlusion-related context for subsequent Mamba-based modeling. It is compatible with both LiDAR-only and multimodal detectors, while introducing only modest overhead. Extensive experiments on nuScenes and Argoverse 2 demonstrate consistent improvements across strong baselines. In particular, RayMamba achieves up to 2.49 mAP and 1.59 NDS gain in the challenging 40--50 m range on nuScenes, and further improves VoxelNeXt on Argoverse 2 from 30.3 to 31.2 mAP.", "date": "2026-04-03", "categories": []} {"id": "2409.11018", "title": "Unleashing the Potential of Mamba: Boosting a LiDAR 3D Sparse Detector by Using Cross-Model Knowledge Distillation", "abstract": "The LiDAR 3D object detector that strikes a balance between accuracy and speed is crucial for achieving real-time perception in autonomous driving. However, many existing LiDAR detection models depend on complex feature transformations, leading to poor real-time performance and high resource consumption, which limits their practical effectiveness. In this work, we propose a faster LiDAR 3D object detector, a framework that adaptively aligns sparse voxels to enable efficient heterogeneous knowledge distillation, called FASD. We aim to distill the Transformer sequence modeling capability into Mamba models, significantly boosting accuracy through knowledge transfer. Specifically, we first design the architecture for cross-model knowledge distillation to impart the global contextual understanding capabilities of the Transformer to Mamba. Transformer-based teacher model employ a scale-adaptive attention mechanism to enhance multiscale fusion. In contrast, Mamba-based student model leverages feature alignment through spatial-based adapters, supervised with latent space feature and span-head distillation losses, leading to improved performance and efficiency. We evaluated the FASD on the Waymo and nuScenes datasets, achieving a 4x reduction in resource consumption and a 1-2% performance improvement over the baseline, while also delivering significant gains in accuracy and efficiency in real deployment.", "date": "2024-09-17", "categories": []} {"id": "2505.04594", "title": "Unleashing the Power of Chain-of-Prediction for Monocular 3D Object Detection", "abstract": "Monocular 3D detection (Mono3D) aims to infer 3D bounding boxes from a single RGB image. Without auxiliary sensors such as LiDAR, this task is inherently ill-posed since the 3D-to-2D projection introduces depth ambiguity. Previous works often predict 3D attributes (e.g., depth, size, and orientation) in parallel, overlooking that these attributes are inherently correlated through the 3D-to-2D projection. However, simply enforcing such correlations through sequential prediction can propagate errors across attributes, especially when objects are occluded or truncated, where inaccurate size or orientation predictions can further amplify depth errors. Therefore, neither parallel nor sequential prediction is optimal. In this paper, we propose MonoCoP, an adaptive framework that learns when and how to leverage inter-attribute correlations with two complementary designs. A Chain-of-Prediction (CoP) explores inter-attribute correlations through feature-level learning, propagation, and aggregation, while an Uncertainty-Guided Selector (GS) dynamically switches between CoP and parallel paradigms for each object based on the predicted uncertainty. By combining their strengths, MonoCoP achieves state-of-the-art (SOTA) performance on KITTI, nuScenes, and Waymo, significantly improving depth accuracy, particularly for distant objects.", "date": "2025-05-07", "categories": []} {"id": "2603.27059", "title": "Towards Intrinsic-Aware Monocular 3D Object Detection", "abstract": "Monocular 3D object detection (Mono3D) aims to infer object locations and dimensions in 3D space from a single RGB image. Despite recent progress, existing methods remain highly sensitive to camera intrinsics and struggle to generalize across diverse settings, since intrinsics govern how 3D scenes are projected onto the image plane. We propose MonoIA, a unified intrinsic-aware framework that models and adapts to intrinsic variation through a language-grounded representation. The key insight is that intrinsic variation is not a numeric difference but a perceptual transformation that alters apparent scale, perspective, and spatial geometry. To capture this effect, MonoIA employs large language models and vision-language models to generate intrinsic embeddings that encode the visual and geometric implications of camera parameters. These embeddings are hierarchically integrated into the detection network via an Intrinsic Adaptation Module, allowing the model to modulate its feature representations according to camera-specific configurations and maintain consistent 3D detection across intrinsics. This shifts intrinsic modeling from numeric conditioning to semantic representation, enabling robust and unified perception across cameras. Extensive experiments show that MonoIA achieves new state-of-the-art results on standard benchmarks including KITTI, Waymo, and nuScenes (e.g., +1.18% on the KITTI leaderboard), and further improves performance under multi-dataset training (e.g., +4.46% on KITTI Val).", "date": "2026-03-28", "categories": []} {"id": "2603.27818", "title": "Benchmarking Multi-View BEV Object Detection with Mixed Pinhole and Fisheye Cameras", "abstract": "Modern autonomous driving systems increasingly rely on mixed camera configurations with pinhole and fisheye cameras for full view perception. However, Bird's-Eye View (BEV) 3D object detection models are predominantly designed for pinhole cameras, leading to performance degradation under fisheye distortion. To bridge this gap, we introduce a multi-view BEV detection benchmark with mixed cameras by converting KITTI-360 into nuScenes format. Our study encompasses three adaptations: rectification for zero-shot evaluation and fine-tuning of nuScenes-trained models, distortion-aware view transformation modules (VTMs) via the MEI camera model, and polar coordinate representations to better align with radial distortion. We systematically evaluate three representative BEV architectures, BEVFormer, BEVDet and PETR, across these strategies. We demonstrate that projection-free architectures are inherently more robust and effective against fisheye distortion than other VTMs. This work establishes the first real-data 3D detection benchmark with fisheye and pinhole images and provides systematic adaptation and practical guidelines for designing robust and cost-effective 3D perception systems. The code is available at https://github.com/CesarLiu/FishBEVOD.git.", "date": "2026-03-29", "categories": []} {"id": "2603.27995", "title": "UniDA3D: A Unified Domain-Adaptive Framework for Multi-View 3D Object Detection", "abstract": "Camera-only 3D object detection is critical for autonomous driving, offering a cost-effective alternative to LiDAR based methods. In particular, multi-view 3D object detection has emerged as a promising direction due to its balanced trade-off between performance and cost. However, existing methods often suffer significant performance degradation under complex environmental conditions such as nighttime, fog, and rain, primarily due to their reliance on training data collected mostly in ideal conditions. To address this challenge, we propose UniDA3D, a unified domain-adaptive multi-view 3D object detector designed for robust perception under diverse adverse conditions. UniDA3D formulates nighttime, rainy, and foggy scenes as a unified multi target domain adaptation problem and leverages a novel query guided domain discrepancy mitigation (QDDM) module to align object features between source and target domains at both batch and global levels via query-centric adversarial and contrastive learning. Furthermore, we introduce a domain-adaptive teacher student training pipeline with an exponential-moving-average teacher and dynamically updated high-quality pseudo labels to enhance consistency learning and suppress background noise in unlabeled target domains. In contrast to prior approaches that require separate training for each condition, UniDA3D performs a single unified training process across multiple domains, enabling robust all-weather 3D perception. On a synthesized multi-view 3D benchmark constructed by generating nighttime, rainy, and foggy counterparts from nuScenes (nuScenes-Night, nuScenes-Rain, and nuScenes-Haze), UniDA3D consistently outperforms state of-the-art camera-only multi-view 3D detectors under extreme conditions, achieving substantial gains in mAP and NDS while maintaining real-time inference efficiency.", "date": "2026-03-30", "categories": []} {"id": "2603.05042", "title": "CoIn3D: Revisiting Configuration-Invariant Multi-Camera 3D Object Detection", "abstract": "Multi-camera 3D object detection (MC3D) has attracted increasing attention with the growing deployment of multi-sensor physical agents, such as robots and autonomous vehicles. However, MC3D models still struggle to generalize to unseen platforms with new multi-camera configurations. Current solutions simply employ a meta-camera for unified representation but lack comprehensive consideration. In this paper, we revisit this issue and identify that the devil lies in spatial prior discrepancies across source and target configurations, including different intrinsics, extrinsics, and array layouts. To address this, we propose CoIn3D, a generalizable MC3D framework that enables strong transferability from source configurations to unseen target ones. CoIn3D explicitly incorporates all identified spatial priors into both feature embedding and image observation through spatial-aware feature modulation (SFM) and camera-aware data augmentation (CDA), respectively. SFM enriches feature space by integrating four spatial representations, such as focal length, ground depth, ground gradient, and Pl\\\"ucker coordinate. CDA improves observation diversity under various configurations via a training-free dynamic novel-view image synthesis scheme. Extensive experiments demonstrate that CoIn3D achieves strong cross-configuration performance on landmark datasets such as NuScenes, Waymo, and Lyft, under three dominant MC3D paradigms represented by BEVDepth, BEVFormer, and PETR.", "date": "2026-03-05", "categories": []} {"id": "2603.25275", "title": "V2U4Real: A Real-world Large-scale Dataset for Vehicle-to-UAV Cooperative Perception", "abstract": "Modern autonomous vehicle perception systems are often constrained by occlusions, blind spots, and limited sensing range. While existing cooperative perception paradigms, such as Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I), have demonstrated their effectiveness in mitigating these challenges, they remain limited to ground-level collaboration and cannot fully address large-scale occlusions or long-range perception in complex environments. To advance research in cross-view cooperative perception, we present V2U4Real, the first large-scale real-world multi-modal dataset for Vehicle-to-UAV (V2U) cooperative object perception. V2U4Real is collected by a ground vehicle and a UAV equipped with multi-view LiDARs and RGB cameras. The dataset covers urban streets, university campuses, and rural roads under diverse traffic scenarios, comprising over 56K LiDAR frames, 56K multi-view camera images, and 700K annotated 3D bounding boxes across four classes. To support a wide range of research tasks, we establish benchmarks for single-agent 3D object detection, cooperative 3D object detection, and object tracking. Comprehensive evaluations of several state-of-the-art models demonstrate the effectiveness of V2U cooperation in enhancing perception robustness and long-range awareness. The V2U4Real dataset and codebase is available at https://github.com/VjiaLi/V2U4Real.", "date": "2026-03-26", "categories": []} {"id": "2603.24294", "title": "VERIA: Verification-Centric Multimodal Instance Augmentation for Long-Tailed 3D Object Detection", "abstract": "Long-tail distributions in driving datasets pose a fundamental challenge for 3D perception, as rare classes exhibit substantial intra-class diversity yet available samples cover this variation space only sparsely. Existing instance augmentation methods based on copy-paste or asset libraries improve rare-class exposure but are often limited in fine-grained diversity and scene-context placement. We propose VERIA, an image-first multimodal augmentation framework that synthesizes synchronized RGB--LiDAR instances using off-the-shelf foundation models and curates them with sequential semantic and geometric verification. This verification-centric design tends to select instances that better match real LiDAR statistics while spanning a wider range of intra-class variation. Stage-wise yield decomposition provides a log-based diagnostic of pipeline reliability. On nuScenes and Lyft, VERIA improves rare-class 3D object detection in both LiDAR-only and multimodal settings. Our code is available at https://sgvr.kaist.ac.kr/VERIA/.", "date": "2026-03-25", "categories": []} {"id": "2603.23276", "title": "CCF: Complementary Collaborative Fusion for Domain Generalized Multi-Modal 3D Object Detection", "abstract": "Multi-modal fusion has emerged as a promising paradigm for accurate 3D object detection. However, performance degrades substantially when deployed in target domains different from training. In this work, focusing on dual-branch proposal-level detectors, we identify two factors that limit robust cross-domain generalization: 1) in challenging domains such as rain or nighttime, one modality may undergo severe degradation; 2) the LiDAR branch often dominates the detection process, leading to systematic underutilization of visual cues and vulnerability when point clouds are compromised. To address these challenges, we propose three components. First, Query-Decoupled Loss provides independent supervision for 2D-only, 3D-only, and fused queries, rebalancing gradient flow across modalities. Second, LiDAR-Guided Depth Prior augments 2D queries with instance-aware geometric priors through probabilistic fusion of image-predicted and LiDAR-derived depth distributions, improving their spatial initialization. Third, Complementary Cross-Modal Masking applies complementary spatial masks to the image and point cloud, encouraging queries from both modalities to compete within the fused decoder and thereby promoting adaptive fusion. Extensive experiments demonstrate substantial gains over state-of-the-art baselines while preserving source-domain performance. Code and models are publicly available at https://github.com/IMPL-Lab/CCF.", "date": "2026-03-24", "categories": []} {"id": "2603.21944", "title": "Group3D: MLLM-Driven Semantic Grouping for Open-Vocabulary 3D Object Detection", "abstract": "Open-vocabulary 3D object detection aims to localize and recognize objects beyond a fixed training taxonomy. In multi-view RGB settings, recent approaches often decouple geometry-based instance construction from semantic labeling, generating class-agnostic fragments and assigning open-vocabulary categories post hoc. While flexible, such decoupling leaves instance construction governed primarily by geometric consistency, without semantic constraints during merging. When geometric evidence is view-dependent and incomplete, this geometry-only merging can lead to irreversible association errors, including over-merging of distinct objects or fragmentation of a single instance. We propose Group3D, a multi-view open-vocabulary 3D detection framework that integrates semantic constraints directly into the instance construction process. Group3D maintains a scene-adaptive vocabulary derived from a multimodal large language model (MLLM) and organizes it into semantic compatibility groups that encode plausible cross-view category equivalence. These groups act as merge-time constraints: 3D fragments are associated only when they satisfy both semantic compatibility and geometric consistency. This semantically gated merging mitigates geometry-driven over-merging while absorbing multi-view category variability. Group3D supports both pose-known and pose-free settings, relying only on RGB observations. Experiments on ScanNet and ARKitScenes demonstrate that Group3D achieves state-of-the-art performance in multi-view open-vocabulary 3D detection, while exhibiting strong generalization in zero-shot scenarios. The project page is available at https://ubin108.github.io/Group3D/.", "date": "2026-03-23", "categories": []} {"id": "2511.08015", "title": "Invisible Triggers, Visible Threats! Road-Style Adversarial Creation Attack for Visual 3D Detection in Autonomous Driving", "abstract": "Modern autonomous driving (AD) systems leverage 3D object detection to perceive foreground objects in 3D environments for subsequent prediction and planning. Visual 3D detection based on RGB cameras provides a cost-effective solution compared to the LiDAR paradigm. While achieving promising detection accuracy, current deep neural network-based models remain highly susceptible to adversarial examples. The underlying safety concerns motivate us to investigate realistic adversarial attacks in AD scenarios. Previous work has demonstrated the feasibility of placing adversarial posters on the road surface to induce hallucinations in the detector. However, the unnatural appearance of the posters makes them easily noticeable by humans, and their fixed content can be readily targeted and defended. To address these limitations, we propose the AdvRoad to generate diverse road-style adversarial posters. The adversaries have naturalistic appearances resembling the road surface while compromising the detector to perceive non-existent objects at the attack locations. We employ a two-stage approach, termed Road-Style Adversary Generation and Scenario-Associated Adaptation, to maximize the attack effectiveness on the input scene while ensuring the natural appearance of the poster, allowing the attack to be carried out stealthily without drawing human attention. Extensive experiments show that AdvRoad generalizes well to different detectors, scenes, and spoofing locations. Moreover, physical attacks further demonstrate the practical threats in real-world environments.", "date": "2025-11-11", "categories": []} {"id": "2603.17470", "title": "VirPro: Visual-referred Probabilistic Prompt Learning for Weakly-Supervised Monocular 3D Detection", "abstract": "Monocular 3D object detection typically relies on pseudo-labeling techniques to reduce dependency on real-world annotations. Recent advances demonstrate that deterministic linguistic cues can serve as effective auxiliary weak supervision signals, providing complementary semantic context. However, hand-crafted textual descriptions struggle to capture the inherent visual diversity of individuals across scenes, limiting the model's ability to learn scene-aware representations. To address this challenge, we propose Visual-referred Probabilistic Prompt Learning (VirPro), an adaptive multi-modal pretraining paradigm that can be seamlessly integrated into diverse weakly supervised monocular 3D detection frameworks. Specifically, we generate a diverse set of learnable, instance-conditioned prompts across scenes and store them in an Adaptive Prompt Bank (APB). Subsequently, we introduce Multi-Gaussian Prompt Modeling (MGPM), which incorporates scene-based visual features into the corresponding textual embeddings, allowing the text prompts to express visual uncertainties. Then, from the fused vision-language embeddings, we decode a prompt-targeted Gaussian, from which we derive a unified object-level prompt embedding for each instance. RoI-level contrastive matching is employed to enforce modality alignment, bringing embeddings of co-occurring objects within the same scene closer in the latent space, thus enhancing semantic coherence. Extensive experiments on the KITTI benchmark demonstrate that integrating our pretraining paradigm consistently yields substantial performance gains, achieving up to a 4.8% average precision improvement than the baseline. Code is available at https://github.com/AustinLCP/VirPro.", "date": "2026-03-18", "categories": []} {"id": "2507.16861", "title": "Look Before You Fuse: 2D-Guided Cross-Modal Alignment for Robust 3D Detection", "abstract": "Integrating LiDAR and camera inputs into a unified Bird's-Eye-View (BEV) representation is crucial for enhancing 3D perception capabilities of autonomous vehicles. However, existing methods suffer from spatial misalignment between LiDAR and camera features, which causes inaccurate depth supervision in camera branch and erroneous fusion during cross-modal feature aggregation. The root cause of this misalignment lies in projection errors, stemming from calibration inaccuracies and rolling shutter effect. The key insight of this work is that locations of these projection errors are not random but highly predictable, as they are concentrated at object-background boundaries which 2D detectors can reliably identify. Based on this, our main motivation is to utilize 2D object priors to pre-align cross-modal features before fusion. To address local misalignment, we propose Prior Guided Depth Calibration (PGDC), which leverages 2D priors to alleviate misalignment and preserve correct cross-modal feature pairs. To resolve global misalignment, we introduce Discontinuity Aware Geometric Fusion (DAGF) to suppress residual noise from PGDC and explicitly enhance sharp depth transitions at object-background boundaries, yielding a structurally aware representation. To effectively utilize these aligned representations, we incorporate Structural Guidance Depth Modulator (SGDM), using a gated attention mechanism to efficiently fuse aligned depth and image features. Our method achieves SOTA performance on nuScenes validation dataset, with its mAP and NDS reaching 71.5% and 73.6% respectively. Additionally, on the Argoverse 2 validation set, we achieve a competitive mAP of 41.7%.", "date": "2025-07-21", "categories": []} {"id": "2603.18298", "title": "Sparse3DTrack: Monocular 3D Object Tracking Using Sparse Supervision", "abstract": "Monocular 3D object tracking aims to estimate temporally consistent 3D object poses across video frames, enabling autonomous agents to reason about scene dynamics. However, existing state-of-the-art approaches are fully supervised and rely on dense 3D annotations over long video sequences, which are expensive to obtain and difficult to scale. In this work, we address this fundamental limitation by proposing the first sparsely supervised framework for monocular 3D object tracking. Our approach decomposes the task into two sequential sub-problems: 2D query matching and 3D geometry estimation. Both components leverage the spatio-temporal consistency of image sequences to augment a sparse set of labeled samples and learn rich 2D and 3D representations of the scene. Leveraging these learned cues, our model automatically generates high-quality 3D pseudolabels across entire videos, effectively transforming sparse supervision into dense 3D track annotations. This enables existing fully-supervised trackers to effectively operate under extreme label sparsity. Extensive experiments on the KITTI and nuScenes datasets demonstrate that our method significantly improves tracking performance, achieving an improvement of up to 15.50 p.p. while using at most four ground truth annotations per track.", "date": "2026-03-18", "categories": []} {"id": "2512.05663", "title": "LeAD-M3D: Leveraging Asymmetric Distillation for Real-Time Monocular 3D Detection", "abstract": "Real-time monocular 3D object detection remains challenging due to severe depth ambiguity, viewpoint shifts, and the high computational cost of 3D reasoning. Existing approaches either rely on LiDAR or geometric priors to compensate for missing depth or sacrifice efficiency to achieve competitive accuracy. We introduce LeAD-M3D, a monocular 3D detector that achieves state-of-the-art accuracy and real-time inference without extra modalities. Our method is enabled by three key components. Asymmetric Augmentation Denoising Distillation (A2D2) transfers geometric knowledge from a clean-image teacher to a MixUp-noised student via a quality- and importance-weighted depth-feature loss, enabling stronger depth reasoning without LiDAR. 3D-aware Consistent Matching (CM$_{\\text{3D}}$) improves prediction-to-ground truth assignment by integrating 3D MGIoU into the matching score, yielding stable and precise supervision. Finally, Confidence-Gated 3D Inference (CGI$_{\\text{3D}}$) accelerates inference by restricting expensive 3D regression to confident regions. Together, these contributions set a new Pareto frontier for monocular 3D detection: LeAD-M3D achieves state-of-the-art accuracy on KITTI and Waymo, and the best reported car AP on Rope3D, while running up to 3.6$\\,\\times$ faster than prior high-accuracy models (e.g., MonoDiff). LeAD-M3D demonstrates that high fidelity and real-time monocular 3D detection is simultaneously attainable, without LiDAR, stereo, or strong geometric assumptions.", "date": "2025-12-05", "categories": []} {"id": "2603.16261", "title": "AW-MoE: All-Weather Mixture of Experts for Robust Multi-Modal 3D Object Detection", "abstract": "Robust 3D object detection under adverse weather conditions is crucial for autonomous driving. However, most existing methods simply combine all weather samples for training while overlooking data distribution discrepancies across different weather scenarios, leading to performance conflicts. To address this issue, we introduce AW-MoE, the framework that innovatively integrates Mixture of Experts (MoE) into weather-robust multi-modal 3D object detection approaches. AW-MoE incorporates Image-guided Weather-aware Routing (IWR), which leverages the superior discriminability of image features across weather conditions and their invariance to scene variations for precise weather classification. Based on this accurate classification, IWR selects the top-K most relevant Weather-Specific Experts (WSE) that handle data discrepancies, ensuring optimal detection under all weather conditions. Additionally, we propose a Unified Dual-Modal Augmentation (UDMA) for synchronous LiDAR and 4D Radar dual-modal data augmentation while preserving the realism of scenes. Extensive experiments on the real-world dataset demonstrate that AW-MoE achieves ~ 15% improvement in adverse-weather performance over state-of-the-art methods, while incurring negligible inference overhead. Moreover, integrating AW-MoE into established baseline detectors yields performance improvements surpassing current state-of-the-art methods. These results show the effectiveness and strong scalability of our AW-MoE. We will release the code publicly at https://github.com/windlinsherlock/AW-MoE.", "date": "2026-03-17", "categories": []} {"id": "2603.16742", "title": "When the City Teaches the Car: Label-Free 3D Perception from Infrastructure", "abstract": "Building robust 3D perception for self-driving still relies heavily on large-scale data collection and manual annotation, yet this paradigm becomes impractical as deployment expands across diverse cities and regions. Meanwhile, modern cities are increasingly instrumented with roadside units (RSUs), static sensors deployed along roads and at intersections to monitor traffic. This raises a natural question: can the city itself help train the vehicle? We propose infrastructure-taught, label-free 3D perception, a paradigm in which RSUs act as stationary, unsupervised teachers for ego vehicles. Leveraging their fixed viewpoints and repeated observations, RSUs learn local 3D detectors from unlabeled data and broadcast predictions to passing vehicles, which are aggregated as pseudo-label supervision for training a standalone ego detector. The resulting model requires no infrastructure or communication at test time. We instantiate this idea as a fully label-free three-stage pipeline and conduct a concept-and-feasibility study in a CARLA-based multi-agent environment. With CenterPoint, our pipeline achieves 82.3% AP for detecting vehicles, compared to a fully supervised ego upper bound of 94.4%. We further systematically analyze each stage, evaluate its scalability, and demonstrate complementarity with existing ego-centric label-free methods. Together, these results suggest that city infrastructure itself can potentially provide a scalable supervisory signal for autonomous vehicles, positioning infrastructure-taught learning as a promising orthogonal paradigm for reducing annotation cost in 3D perception.", "date": "2026-03-17", "categories": []} {"id": "2509.16119", "title": "RadarGaussianDet3D: Gaussian Representation-based Real-time 3D Object Detection with 4D Automotive Radars", "abstract": "4D automotive radars have gained increasing attention for autonomous driving due to their low cost, robustness, and inherent velocity measurement capability. However, existing 4D radar-based 3D detectors rely heavily on pillar encoders for BEV feature extraction, where each point contributes to only a single BEV grid, resulting in sparse feature maps and degraded representation quality. In addition, they also optimize bounding box attributes independently, leading to sub-optimal detection accuracy. Moreover, their inference speed, while sufficient for high-end GPUs, may fail to meet the real-time requirement on vehicle-mounted embedded devices. To overcome these limitations, an efficient and effective Gaussian-based 3D detector, namely RadarGaussianDet3D is introduced, leveraging Gaussian primitives and distributions as intermediate representations for radar points and bounding boxes. In RadarGaussianDet3D, a novel Point Gaussian Encoder (PGE) is designed to transform each point into a Gaussian primitive after feature aggregation and employs the 3D Gaussian Splatting (3DGS) technique for BEV rasterization, yielding denser feature maps. PGE exhibits exceptionally low latency, owing to the optimized algorithm for point feature aggregation and fast rendering of 3DGS. In addition, a new Box Gaussian Loss (BGL) is proposed, which converts bounding boxes into 3D Gaussian distributions and measures their distance to enable more comprehensive and consistent optimization. Extensive experiments on TJ4DRadSet and View-of-Delft demonstrate that RadarGaussianDet3D achieves high detection accuracy while delivering substantially faster inference, highlighting its potential for real-time deployment in autonomous driving.", "date": "2025-09-19", "categories": []} {"id": "2603.13919", "title": "OpenCOOD-Air: Prompting Heterogeneous Ground-Air Collaborative Perception with Spatial Conversion and Offset Prediction", "abstract": "While Vehicle-to-Vehicle (V2V) collaboration extends sensing ranges through multi-agent data sharing, its reliability remains severely constrained by ground-level occlusions and the limited perspective of chassis-mounted sensors, which often result in critical perception blind spots. We propose OpenCOOD-Air, a novel framework that integrates UAVs as extensible platforms into V2V collaborative perception to overcome these constraints. To mitigate gradient interference from ground-air domain gaps and data sparsity, we adopt a transfer learning strategy to fine-tune UAV weights from pre-trained V2V models. To prevent the spatial information loss inherent in this transition, we formulate ground-air collaborative perception as a heterogeneous integration task with explicit altitude supervision and introduce a Cross-Domain Spatial Converter (CDSC) and a Spatial Offset Prediction Transformer (SOPT). Furthermore, we present the OPV2V-Air benchmark to validate the transition from V2V to Vehicle-to-Vehicle-to-UAV. Compared to state-of-the-art methods, our approach improves 2D and 3D AP@0.7 by 4% and 7%, respectively.", "date": "2026-03-14", "categories": []} {"id": "2603.14822", "title": "RadarXFormer: Robust Object Detection via Cross-Dimension Fusion of 4D Radar Spectra and Images for Autonomous Driving", "abstract": "Reliable perception is essential for autonomous driving systems to operate safely under diverse real-world traffic conditions. However, camera- and LiDAR-based perception systems suffer from performance degradation under adverse weather and lighting conditions, limiting their robustness and large-scale deployment in intelligent transportation systems. Radar-vision fusion provides a promising alternative by combining the environmental robustness and cost efficiency of millimeter-wave (mmWave) radar with the rich semantic information captured by cameras. Nevertheless, conventional 3D radar measurements lack height resolution and remain highly sparse, while emerging 4D mmWave radar introduces elevation information but also brings challenges such as signal noise and large data volume. To address these issues, this paper proposes RadarXFormer, a 3D object detection framework that enables efficient cross-modal fusion between 4D radar spectra and RGB images. Instead of relying on sparse radar point clouds, RadarXFormer directly leverages raw radar spectra and constructs an efficient 3D representation that reduces data volume while preserving complete 3D spatial information. The \"X\" highlights the proposed cross-dimension (3D-2D) fusion mechanism, in which multi-scale 3D spherical radar feature cubes are fused with complementary 2D image feature maps. Experiments on the K-Radar dataset demonstrate improved detection accuracy and robustness under challenging conditions while maintaining real-time inference capability.", "date": "2026-03-16", "categories": []} {"id": "2410.01647", "title": "3DGS-DET: Empower 3D Gaussian Splatting with Boundary Guidance and Box-Focused Sampling for Indoor 3D Object Detection", "abstract": "Neural Radiance Fields (NeRF) have been adapted for indoor 3D Object Detection (3DOD), offering a promising approach to indoor 3DOD via view-synthesis representation. But its implicit nature limits representational capacity. Recently, 3D Gaussian Splatting (3DGS) has emerged as an explicit 3D representation that addresses the limitation. This work introduces 3DGS into indoor 3DOD for the first time, identifying two main challenges: (i) Ambiguous spatial distribution of Gaussian blobs -- 3DGS primarily relies on 2D pixel-level supervision, resulting in unclear 3D spatial distribution of Gaussian blobs and poor differentiation between objects and background, which hinders indoor 3DOD; (ii) Excessive background blobs -- 2D images typically include numerous background pixels, leading to densely reconstructed 3DGS with many noisy Gaussian blobs representing the background, negatively affecting detection. To tackle (i), we leverage the fact that 3DGS reconstruction is derived from 2D images, and propose an elegant solution by incorporating 2D Boundary Guidance to significantly enhance the spatial distribution of Gaussian blobs, resulting in clearer differentiation between objects and their background (please see fig:teaser). To address (ii), we propose a Box-Focused Sampling strategy using 2D boxes to generate object probability distribution in 3D space, allowing effective probabilistic sampling in 3D to retain more object blobs and reduce noisy background blobs. Benefiting from these innovations, 3DGS-DET significantly outperforms the state-of-the-art NeRF-based method, NeRF-Det++, achieving improvements of +6.0 on mAP@0.25 and +7.8 on mAP@0.5 for the ScanNet, and the +14.9 on mAP@0.25 for the ARKITScenes.", "date": "2024-10-02", "categories": []} {"id": "2603.12599", "title": "A Prediction-as-Perception Framework for 3D Object Detection", "abstract": "Humans combine prediction and perception to observe the world. When faced with rapidly moving birds or insects, we can only perceive them clearly by predicting their next position and focusing our gaze there. Inspired by this, this paper proposes the Prediction-As-Perception (PAP) framework, integrating a prediction-perception architecture into 3D object perception tasks to enhance the model's perceptual accuracy. The PAP framework consists of two main modules: prediction and perception, primarily utilizing continuous frame information as input. Firstly, the prediction module forecasts the potential future positions of ego vehicles and surrounding traffic participants based on the perception results of the current frame. These predicted positions are then passed as queries to the perception module of the subsequent frame. The perceived results are iteratively fed back into the prediction module. We evaluated the PAP structure using the end-to-end model UniAD on the nuScenes dataset. The results demonstrate that the PAP structure improves UniAD's target tracking accuracy by 10% and increases the inference speed by 15%. This indicates that such a biomimetic design significantly enhances the efficiency and accuracy of perception models while reducing computational resource consumption.", "date": "2026-03-13", "categories": []} {"id": "2603.09695", "title": "DRIFT: Dual-Representation Inter-Fusion Transformer for Automated Driving Perception with 4D Radar Point Clouds", "abstract": "4D radars, which provide 3D point cloud data along with Doppler velocity, are attractive components of modern automated driving systems due to their low cost and robustness under adverse weather conditions. However, they provide a significantly lower point cloud density than LiDAR sensors. This makes it important to exploit not only local but also global contextual scene information. This paper proposes DRIFT, a model that effectively captures and fuses both local and global contexts through a dual-path architecture. The model incorporates a point path to aggregate fine-grained local features and a pillar path to encode coarse-grained global features. These two parallel paths are intertwined via novel feature-sharing layers at multiple stages, enabling full utilization of both representations. DRIFT is evaluated on the widely used View-of-Delft (VoD) dataset and a proprietary internal dataset. It outperforms the baselines on the tasks of object detection and/or free road estimation. For example, DRIFT achieves a mean average precision (mAP) of 52.6% (compared to, say, 45.4% of CenterPoint) on the VoD dataset.", "date": "2026-03-10", "categories": []} {"id": "2506.07737", "title": "SpikeSMOKE: Spiking Neural Networks for Monocular 3D Object Detection with Cross-Scale Gated Coding", "abstract": "With the wide application of 3D object detection in some fields such as autonomous driving, its energy consumption is constantly increasing, making the research on low-power consumption alternatives a key research area. The spiking neural networks (SNNs), possessing low-power consumption characteristics, offer a novel solution for this research. Consequently, we apply SNNs to monocular 3D object detection and propose the SpikeSMOKE architecture, which represents a new attempt at low-power monocular 3D object detection. It's well known that the discrete signals of SNNs can lead to information loss compared to artificial neural networks (ANNs), which limits their feature representation capabilities. To solve this problem, inspired by the synaptic filtering mechanism of biological neurons, we propose a new Cross-Scale Gating Coding Mechanism (CSGC), which can enhance feature representation by combining cross-scale fusion of attentional methods and gated filtering mechanisms. In addition, to reduce the computation and accelerate training, we present a novel light-weight residual block that can maintain spiking computing paradigm and the highest possible detection performance. Our method is effective on the KITTI, NuScenes-mini and CIFAR10/100 datasets. Compared to the baseline SpikeSMOKE under the 3D Object Detection, the proposed SpikeSMOKE with CSGC can achieve 11.78 (+2.82, Easy), 10.69 (+3.2, Moderate), and 10.48 (+3.17, Hard) on the KITTI autonomous driving dataset by AP|R11 at 0.7 IoU threshold, respectively. It is worth noting that the results of SpikeSMOKE can significantly reduce energy consumption compared with the results of SMOKE. And SpikeSMOKE-L (lightweight) can further reduce the amount of parameters by 3 times and computation by 10 times compared to SMOKE.", "date": "2025-06-09", "categories": []} {"id": "2511.06702", "title": "SPAN: Spatial-Projection Alignment for Monocular 3D Object Detection", "abstract": "Existing monocular 3D detectors typically tame the pronounced nonlinear regression of 3D bounding box through decoupled prediction paradigm, which employs multiple branches to estimate geometric center, depth, dimensions, and rotation angle separately. Although this decoupling strategy simplifies the learning process, it inherently ignores the geometric collaborative constraints between different attributes, resulting in the lack of geometric consistency prior, thereby leading to suboptimal performance. To address this issue, we propose novel Spatial-Projection Alignment (SPAN) with two pivotal components: (i). Spatial Point Alignment enforces an explicit global spatial constraint between the predicted and ground-truth 3D bounding boxes, thereby rectifying spatial drift caused by decoupled attribute regression. (ii). 3D-2D Projection Alignment ensures that the projected 3D box is aligned tightly within its corresponding 2D detection bounding box on the image plane, mitigating projection misalignment overlooked in previous works. To ensure training stability, we further introduce a Hierarchical Task Learning strategy that progressively incorporates spatial-projection alignment as 3D attribute predictions refine, preventing early stage error propagation across attributes. Extensive experiments demonstrate that the proposed method can be easily integrated into any established monocular 3D detector and delivers significant performance improvements.", "date": "2025-11-10", "categories": []} {"id": "2509.01487", "title": "PointSlice: Accurate and Efficient Slice-Based Representation for 3D Object Detection from Point Clouds", "abstract": "3D object detection from point clouds plays a critical role in autonomous driving. Currently, the primary methods for point cloud processing are voxel-based and pillar-based approaches. Voxel-based methods offer high accuracy through fine-grained spatial segmentation but suffer from slower inference speeds. Pillar-based methods enhance inference speed but typically lag behind voxel-based methods in detection accuracy. To address this trade-off, we propose a novel point cloud processing method, PointSlice, which slices point clouds along the horizontal plane and incorporates a dedicated detection network. The main contributions of PointSlice are: (1) A novel slice-based representation that converts 3D point clouds into multiple sets of 2D (x-y) data slices. The model explicitly learns 2D data distributions by treating the 3D point cloud as separate batches of 2D data, which significantly reduces the parameter count and enhances inference speed; (2) The introduction of a Slice Interaction Network (SIN). To preserve vertical geometric relationships across slices, we incorporate SIN into the 2D backbone network, thereby improving the model's 3D perception capability. Extensive experiments demonstrate that PointSlice achieves a superior balance between detection accuracy and efficiency. On the Waymo Open Dataset, PointSlice achieves a 1.13$\\times$ speedup and uses 0.79$\\times$ the parameters of the state-of-the-art voxel-based method (SAFDNet), with a marginal 1.2 mAPH accuracy reduction. On the nuScenes dataset, we achieve a state-of-the-art 66.7 mAP. On the Argoverse 2 dataset, PointSlice is 1.10$\\times$ faster with 0.66$\\times$ the parameters, while showing a negligible accuracy drop of 1.0 mAP. The source code is available at https://github.com/qifeng22/PointSlice2.", "date": "2025-09-01", "categories": []} {"id": "2602.20627", "title": "Object-Scene-Camera Decomposition and Recomposition for Data-Efficient Monocular 3D Object Detection", "abstract": "Monocular 3D object detection (M3OD) is intrinsically ill-posed, hence training a high-performance deep learning based M3OD model requires a humongous amount of labeled data with complicated visual variation from diverse scenes, variety of objects and camera poses.However, we observe that, due to strong human bias, the three independent entities, i.e., object, scene, and camera pose, are always tightly entangled when an image is captured to construct training data. More specifically, specific 3D objects are always captured in particular scenes with fixed camera poses, and hence lacks necessary diversity. Such tight entanglement induces the challenging issues of insufficient utilization and overfitting to uniform training data. To mitigate this, we propose an online object-scene-camera decomposition and recomposition data manipulation scheme to more efficiently exploit the training data. We first fully decompose training images into textured 3D object point models and background scenes in an efficient computation and storage manner. We then continuously recompose new training images in each epoch by inserting the 3D objects into the freespace of the background scenes, and rendering them with perturbed camera poses from textured 3D point representation. In this way, the refreshed training data in all epochs can cover the full spectrum of independent object, scene, and camera pose combinations. This scheme can serve as a plug-and-play component to boost M3OD models, working flexibly with both fully and sparsely supervised settings. In the sparsely-supervised setting, objects closest to the ego-camera for all instances are sparsely annotated. We then can flexibly increase the annotated objects to control annotation cost. For validation, our method is widely applied to five representative M3OD models and evaluated on both the KITTI and the more complicated Waymo datasets.", "date": "2026-02-24", "categories": []} {"id": "2603.07464", "title": "Selective Transfer Learning of Cross-Modality Distillation for Monocular 3D Object Detection", "abstract": "Monocular 3D object detection is a promising yet ill-posed task for autonomous vehicles due to the lack of accurate depth information. Cross-modality knowledge distillation could effectively transfer depth information from LiDAR to image-based network. However, modality gap between image and LiDAR seriously limits its accuracy. In this paper, we systematically investigate the negative transfer problem induced by modality gap in cross-modality distillation for the first time, including not only the architecture inconsistency issue but more importantly the feature overfitting issue. We propose a selective learning approach named MonoSTL to overcome these issues, which encourages positive transfer of depth information from LiDAR while alleviates the negative transfer on image-based network. On the one hand, we utilize similar architectures to ensure spatial alignment of features between image-based and LiDAR-based networks. On the other hand, we develop two novel distillation modules, namely Depth-Aware Selective Feature Distillation (DASFD) and Depth-Aware Selective Relation Distillation (DASRD), which selectively learn positive features and relationships of objects by integrating depth uncertainty into feature and relation distillations, respectively. Our approach can be seamlessly integrated into various CNN-based and DETR-based models, where we take three recent models on KITTI and a recent model on NuScenes for validation. Extensive experiments show that our approach considerably improves the accuracy of the base models and thereby achieves the best accuracy compared with all recently released SOTA models.", "date": "2026-03-08", "categories": []} {"id": "2603.07486", "title": "Multi-Modal Decouple and Recouple Network for Robust 3D Object Detection", "abstract": "Multi-modal 3D object detection with bird's eye view (BEV) has achieved desired advances on benchmarks. Nonetheless, the accuracy may drop significantly in the real world due to data corruption such as sensor configurations for LiDAR and scene conditions for camera. One design bottleneck of previous models resides in the tightly coupling of multi-modal BEV features during fusion, which may degrade the overall system performance if one modality or both is corrupted. To mitigate, we propose a Multi-Modal Decouple and Recouple Network for robust 3D object detection under data corruption. Different modalities commonly share some high-level invariant features. We observe that these invariant features across modalities do not always fail simultaneously, because different types of data corruption affect each modality in distinct ways.These invariant features can be recovered across modalities for robust fusion under data corruption.To this end, we explicitly decouple Camera/LiDAR BEV features into modality-invariant and modality-specific parts. It allows invariant features to compensate each other while mitigates the negative impact of a corrupted modality on the other.We then recouple these features into three experts to handle different types of data corruption, respectively, i.e., LiDAR, camera, and both.For each expert, we use modality-invariant features as robust information, while modality-specific features serve as a complement.Finally, we adaptively fuse the three experts to exact robust features for 3D object detection. For validation, we collect a benchmark with a large quantity of data corruption for LiDAR, camera, and both based on nuScenes. Our model is trained on clean nuScenes and tested on all types of data corruption. Our model consistently achieves the best accuracy on both corrupted and clean data compared to recent models.", "date": "2026-03-08", "categories": []} {"id": "2603.07493", "title": "RayD3D: Distilling Depth Knowledge Along the Ray for Robust Multi-View 3D Object Detection", "abstract": "Multi-view 3D detection with bird's eye view (BEV) is crucial for autonomous driving and robotics, but its robustness in real-world is limited as it struggles to predict accurate depth values. A mainstream solution, cross-modal distillation, transfers depth information from LiDAR to camera models but also unintentionally transfers depth-irrelevant information (e.g. LiDAR density). To mitigate this issue, we propose RayD3D, which transfers crucial depth knowledge along the ray: a line projecting from the camera to true location of an object. It is based on the fundamental imaging principle that predicted location of this object can only vary along this ray, which is finally determined by predicted depth value. Therefore, distilling along the ray enables more effective depth information transfer. More specifically, we design two ray-based distillation modules. Ray-based Contrastive Distillation (RCD) incorporates contrastive learning into distillation by sampling along the ray to learn how LiDAR accurately locates objects. Ray-based Weighted Distillation (RWD) adaptively adjusts distillation weight based on the ray to minimize the interference of depth-irrelevant information in LiDAR. For validation, we widely apply RayD3D into three representative types of BEV-based models, including BEVDet, BEVDepth4D, and BEVFormer. Our method is trained on clean NuScenes, and tested on both clean NuScenes and RoboBEV with a variety types of data corruptions. Our method significantly improves the robustness of all the three base models in all scenarios without increasing inference costs, and achieves the best when compared to recently released multi-view and distillation models.", "date": "2026-03-08", "categories": []} {"id": "2603.07985", "title": "On the Feasibility and Opportunity of Autoregressive 3D Object Detection", "abstract": "LiDAR-based 3D object detectors typically rely on proposal heads with hand-crafted components like anchor assignment and non-maximum suppression (NMS), complicating training and limiting extensibility. We present AutoReg3D, an autoregressive 3D detector that casts detection as sequence generation. Given point-cloud features, AutoReg3D emits objects in a range-causal (near-to-far) order and encodes each object as a short, discrete-token sequence consisting of its center, size, orientation, velocity, and class. This near-to-far ordering mirrors LiDAR geometry--near objects occlude far ones but not vice versa--enabling straightforward teacher forcing during training and autoregressive decoding at test time. AutoReg3D is compatible across diverse point-cloud or backbones and attains competitive nuScenes performance without anchors or NMS. Beyond parity, the sequential formulation unlocks language-model advances for 3D perception, including GRPO-style reinforcement learning for task-aligned objectives. These results position autoregressive decoding as a viable, flexible alternative for LiDAR-based detection and open a path to importing modern sequence-modeling tools into 3D perception.", "date": "2026-03-09", "categories": []} {"id": "2603.08199", "title": "Fusion-Poly: A Polyhedral Framework Based on Spatial-Temporal Fusion for 3D Multi-Object Tracking", "abstract": "LiDAR-camera 3D multi-object tracking (MOT) combines rich visual semantics with accurate depth cues to improve trajectory consistency and tracking reliability. In practice, however, LiDAR and cameras operate at different sampling rates. To maintain temporal alignment, existing data pipelines usually synchronize heterogeneous sensor streams and annotate them at a reduced shared frequency, forcing most prior methods to perform spatial fusion only at synchronized timestamps through projection-based or learnable cross-sensor association. As a result, abundant asynchronous observations remain underexploited, despite their potential to support more frequent association and more robust trajectory estimation over short temporal intervals.\n To address this limitation, we propose Fusion-Poly, a spatial-temporal fusion framework for 3D MOT that integrates asynchronous LiDAR and camera data. Fusion-Poly associates trajectories with multi-modal observations at synchronized timestamps and with single-modal observations at asynchronous timestamps, enabling higher-frequency updates of motion and existence states. The framework contains three key components: a frequency-aware cascade matching module that adapts to synchronized and asynchronous frames according to available detection modalities; a frequency-aware trajectory estimation module that maintains trajectories through high-frequency motion prediction, differential updates, and confidence-calibrated lifecycle management; and a full-state observation alignment module that improves cross-modal consistency at synchronized timestamps by optimizing image-projection errors.\n On the nuScenes test set, Fusion-Poly achieves 76.5% AMOTA, establishing a new state of the art among tracking-by-detection 3D MOT methods. Extensive ablation studies further validate the effectiveness of each component. Code will be released.", "date": "2026-03-09", "categories": []} {"id": "2603.08611", "title": "FOMO-3D: Using Vision Foundation Models for Long-Tailed 3D Object Detection", "abstract": "In order to navigate complex traffic environments, self-driving vehicles must recognize many semantic classes pertaining to vulnerable road users or traffic control devices. However, many safety-critical objects (e.g., construction worker) appear infrequently in nominal traffic conditions, leading to a severe shortage of training examples from driving data alone. Recent vision foundation models, which are trained on a large corpus of data, can serve as a good source of external prior knowledge to improve generalization. We propose FOMO-3D, the first multi-modal 3D detector to leverage vision foundation models for long-tailed 3D detection. Specifically, FOMO-3D exploits rich semantic and depth priors from OWLv2 and Metric3Dv2 within a two-stage detection paradigm that first generates proposals with a LiDAR-based branch and a novel camera-based branch, and refines them with attention especially to image features from OWL. Evaluations on real-world driving data show that using rich priors from vision foundation models with careful multi-modal fusion designs leads to large gains for long-tailed 3D detection. Project website is at https://waabi.ai/fomo3d/.", "date": "2026-03-09", "categories": []} {"id": "2603.05623", "title": "Post Fusion Bird's Eye View Feature Stabilization for Robust Multimodal 3D Detection", "abstract": "Camera-LiDAR fusion is widely used in autonomous driving to enable accurate 3D object detection. However, bird's-eye view (BEV) fusion detectors can degrade significantly under domain shift and sensor failures, limiting reliability in real-world deployment. Existing robustness approaches often require modifying the fusion architecture or retraining specialized models, making them difficult to integrate into already deployed systems.\n We propose a Post Fusion Stabilizer (PFS), a lightweight module that operates on intermediate BEV representations of existing detectors and produces a refined feature map for the original detection head. The design stabilizes feature statistics under domain shift, suppresses spatial regions affected by sensor degradation, and adaptively restores weakened cues through residual correction. Designed as a near-identity transformation, PFS preserves performance while improving robustness under diverse camera and LiDAR corruptions. Evaluations on the nuScenes benchmark demonstrate that PFS achieves state-of-the-art results in several failure modes, notably improving camera dropout robustness by +1.2% and low-light performance by +4.4% mAP while maintaining a lightweight footprint of only 3.3 M parameters.", "date": "2026-03-05", "categories": []} {"id": "2603.04938", "title": "Person Detection and Tracking from an Overhead Crane LiDAR", "abstract": "This paper investigates person detection and tracking in an industrial indoor workspace using a LiDAR mounted on an overhead crane. The overhead viewpoint introduces a strong domain shift from common vehicle-centric LiDAR benchmarks, and limited availability of suitable public training data. Henceforth, we curate a site-specific overhead LiDAR dataset with 3D human bounding-box annotations and adapt selected candidate 3D detectors under a unified training and evaluation protocol. We further integrate lightweight tracking-by-detection using AB3DMOT and SimpleTrack to maintain person identities over time. Detection performance is reported with distance-sliced evaluation to quantify the practical operating envelope of the sensing setup. The best adapted detector configurations achieve average precision (AP) up to 0.84 within a 5.0 m horizontal radius, increasing to 0.97 at 1.0 m, with VoxelNeXt and SECOND emerging as the most reliable backbones across this range. The acquired results contribute in bridging the domain gap between standard driving datasets and overhead sensing for person detection and tracking. We also report latency measurements, highlighting practical real-time feasibility. Finally, we release our dataset and implementations in GitHub to support further research", "date": "2026-03-05", "categories": []} {"id": "2603.05305", "title": "Fusion4CA: Boosting 3D Object Detection via Comprehensive Image Exploitation", "abstract": "Nowadays, an increasing number of works fuse LiDAR and RGB data in the bird's-eye view (BEV) space for 3D object detection in autonomous driving systems. However, existing methods suffer from over-reliance on the LiDAR branch, with insufficient exploration of RGB information. To tackle this issue, we propose Fusion4CA, which is built upon the classic BEVFusion framework and dedicated to fully exploiting visual input with plug-and-play components. Specifically, a contrastive alignment module is designed to calibrate image features with 3D geometry, and a camera auxiliary branch is introduced to mine RGB information sufficiently during training. For further performance enhancement, we leverage an off-the-shelf cognitive adapter to make the most of pretrained image weights, and integrate a standard coordinate attention module into the fusion stage as a supplementary boost. Experiments on the nuScenes dataset demonstrate that our method achieves 69.7% mAP with only 6 training epochs and a mere 3.48% increase in inference parameters, yielding a 1.2% improvement over the baseline which is fully trained for 20 epochs. Extensive experiments in a simulated lunar environment further validate the effectiveness and generalization of our method. Our code will be released through Fusion4CA.", "date": "2026-03-05", "categories": []} {"id": "2603.02481", "title": "ModalPatch: A Plug-and-Play Module for Robust Multi-Modal 3D Object Detection under Modality Drop", "abstract": "Multi-modal 3D object detection is pivotal for autonomous driving, integrating complementary sensors like LiDAR and cameras. However, its real-world reliability is challenged by transient data interruptions and missing, where modalities can momentarily drop due to hardware glitches, adverse weather, or occlusions. This poses a critical risk, especially during a simultaneous modality drop, where the vehicle is momentarily blind. To address this problem, we introduce ModalPatch, the first plug-and-play module designed to enable robust detection under arbitrary modality-drop scenarios. Without requiring architectural changes or retraining, ModalPatch can be seamlessly integrated into diverse detection frameworks. Technically, ModalPatch leverages the temporal nature of sensor data for perceptual continuity, using a history-based module to predict and compensate for transiently unavailable features. To improve the fidelity of the predicted features, we further introduce an uncertainty-guided cross-modality fusion strategy that dynamically estimates the reliability of compensated features, suppressing biased signals while reinforcing informative ones. Extensive experiments show that ModalPatch consistently enhances both robustness and accuracy of state-of-the-art 3D object detectors under diverse modality-drop conditions.", "date": "2026-03-03", "categories": []} {"id": "2406.17297", "title": "Towards Camera Open-set 3D Object Detection for Autonomous Driving Scenarios", "abstract": "Conventional camera-based 3D object detectors in autonomous driving are limited to recognizing a predefined set of objects, which poses a safety risk when encountering novel or unseen objects in real-world scenarios. To address this limitation, we present OS-Det3D, a two-stage training framework designed for camera-based open-set 3D object detection. In the first stage, our proposed 3D object discovery network (ODN3D) uses geometric cues from LiDAR point clouds to generate class-agnostic 3D object proposals, each of which are assigned a 3D objectness score. This approach allows the network to discover objects beyond known categories, allowing for the detection of unfamiliar objects. However, due to the absence of class constraints, ODN3D-generated proposals may include noisy data, particularly in cluttered or dynamic scenes. To mitigate this issue, we introduce a joint selection (JS) module in the second stage. The JS module uses both camera bird's eye view (BEV) feature responses and 3D objectness scores to filter out low-quality proposals, yielding high-quality pseudo ground truth for unknown objects. OS-Det3D significantly enhances the ability of camera 3D detectors to discover and identify unknown objects while also improving the performance on known objects, as demonstrated through extensive experiments on the nuScenes and KITTI datasets.", "date": "2024-06-25", "categories": []} {"id": "2512.23147", "title": "GeoTeacher: Geometry-Guided Semi-Supervised 3D Object Detection", "abstract": "Semi-supervised 3D object detection, aiming to explore unlabeled data for boosting 3D object detectors, has emerged as an active research area in recent years. Some previous methods have shown substantial improvements by either employing heterogeneous teacher models to provide high-quality pseudo labels or enforcing feature-perspective consistency between the teacher and student networks. However, these methods overlook the fact that the model usually tends to exhibit low sensitivity to object geometries with limited labeled data, making it difficult to capture geometric information, which is crucial for enhancing the student model's ability in object perception and localization. In this paper, we propose GeoTeacher to enhance the student model's ability to capture geometric relations of objects with limited training data, especially unlabeled data. We design a keypoint-based geometric relation supervision module that transfers the teacher model's knowledge of object geometry to the student, thereby improving the student's capability in understanding geometric relations. Furthermore, we introduce a voxel-wise data augmentation strategy that increases the diversity of object geometries, thereby further improving the student model's ability to comprehend geometric structures. To preserve the integrity of distant objects during augmentation, we incorporate a distance-decay mechanism into this strategy. Moreover, GeoTeacher can be combined with different SS3D methods to further improve their performance. Extensive experiments on the ONCE and Waymo datasets indicate the effectiveness and generalization of our method and we achieve the new state-of-the-art results. Code will be available at https://github.com/SII-Whaleice/GeoTeacher", "date": "2025-12-29", "categories": []} {"id": "2603.00912", "title": "VGGT-Det: Mining VGGT Internal Priors for Sensor-Geometry-Free Multi-View Indoor 3D Object Detection", "abstract": "Current multi-view indoor 3D object detectors rely on sensor geometry that is costly to obtain (i.e., precisely calibrated multi-view camera poses) to fuse multi-view information into a global scene representation, limiting deployment in real-world scenes. We target a more practical setting: Sensor-Geometry-Free (SG-Free) multi-view indoor 3D object detection, where there are no sensor-provided geometric inputs (multi-view poses or depth). Recent Visual Geometry Grounded Transformer (VGGT) shows that strong 3D cues can be inferred directly from images. Building on this insight, we present VGGT-Det, the first framework tailored for SG-Free multi-view indoor 3D object detection. Rather than merely consuming VGGT predictions, our method integrates VGGT encoder into a transformer-based pipeline. To effectively leverage both the semantic and geometric priors from inside VGGT, we introduce two novel key components: (i) Attention-Guided Query Generation (AG): exploits VGGT attention maps as semantic priors to initialize object queries, improving localization by focusing on object regions while preserving global spatial structure; (ii) Query-Driven Feature Aggregation (QD): a learnable See-Query interacts with object queries to 'see' what they need, and then dynamically aggregates multi-level geometric features across VGGT layers that progressively lift 2D features into 3D. Experiments show that VGGT-Det significantly surpasses the best-performing method in the SG-Free setting by 4.4 and 8.6 mAP@0.25 on ScanNet and ARKitScenes, respectively. Ablation study shows that VGGT's internally learned semantic and geometric priors can be effectively leveraged by our AG and QD.", "date": "2026-03-01", "categories": []} {"id": "2603.01947", "title": "physfusion: A Transformer-based Dual-Stream Radar and Vision Fusion Framework for Open Water Surface Object Detection", "abstract": "Detecting water-surface targets for Unmanned Surface Vehicles (USVs) is challenging due to wave clutter,\n specular reflections, and weak appearance cues in long-range observations. Although 4D millimeter-wave\n radar complements cameras under degraded illumination, maritime radar point clouds are sparse and\n intermittent, with reflectivity attributes exhibiting heavy-tailed variations under scattering and\n multipath, making conventional fusion designs struggle to exploit radar cues effectively.\n We propose PhysFusion, a physics-informed radar-image detection framework for water-surface perception.\n The framework integrates: (1) a Physics-Informed Radar Encoder (PIR Encoder) with an RCS Mapper and\n Quality Gate, transforming per-point radar attributes into compact scattering priors and predicting\n point-wise reliability for robust feature learning under clutter; (2) a Radar-guided Interactive Fusion\n Module (RIFM) performing query-level radar-image fusion between semantically enriched radar features and\n multi-scale visual features, with the radar branch modeled by a dual-stream backbone including a\n point-based local stream and a transformer-based global stream using Scattering-Aware Self-Attention\n (SASA); and (3) a Temporal Query Aggregation module (TQA) aggregating frame-wise fused queries over a\n short temporal window for temporally consistent representations.\n Experiments on WaterScenes and FLOW demonstrate that PhysFusion achieves 59.7% mAP50:95 and 90.3% mAP50\n on WaterScenes (T=5 radar history) using 5.6M parameters and 12.5G FLOPs, and reaches 94.8% mAP50 and\n 46.2% mAP50:95 on FLOW under radar+camera setting. Ablation studies quantify the contributions of PIR\n Encoder, SASA-based global reasoning, and RIFM.", "date": "2026-03-02", "categories": []} {"id": "2503.08422", "title": "JiSAM: Alleviate Labeling Burden and Corner Case Problems in Autonomous Driving via Minimal Real-World Data", "abstract": "Deep-learning-based autonomous driving (AD) perception introduces a promising picture for safe and environment-friendly transportation. However, the over-reliance on real labeled data in LiDAR perception limits the scale of on-road attempts. 3D real world data is notoriously time-and-energy-consuming to annotate and lacks corner cases like rare traffic participants. On the contrary, in simulators like CARLA, generating labeled LiDAR point clouds with corner cases is a piece of cake. However, introducing synthetic point clouds to improve real perception is non-trivial. This stems from two challenges: 1) sample efficiency of simulation datasets 2) simulation-to-real gaps. To overcome both challenges, we propose a plug-and-play method called JiSAM , shorthand for Jittering augmentation, domain-aware backbone and memory-based Sectorized AlignMent. In extensive experiments conducted on the famous AD dataset NuScenes, we demonstrate that, with SOTA 3D object detector, JiSAM is able to utilize the simulation data and only labels on 2.5% available real data to achieve comparable performance to models trained on all real data. Additionally, JiSAM achieves more than 15 mAPs on the objects not labeled in the real training set.", "date": "2025-03-11", "categories": []} {"id": "2505.17442", "title": "Reflectance Prediction-based Knowledge Distillation for Robust 3D Object Detection in Compressed Point Clouds", "abstract": "Regarding intelligent transportation systems, low-bitrate transmission via lossy point cloud compression is vital for facilitating real-time collaborative perception among connected agents, such as vehicles and infrastructures, under restricted bandwidth. In existing compression transmission systems, the sender lossily compresses point coordinates and reflectance to generate a transmission code stream, which faces transmission burdens from reflectance encoding and limited detection robustness due to information loss. To address these issues, this paper proposes a 3D object detection framework with reflectance prediction-based knowledge distillation (RPKD). We compress point coordinates while discarding reflectance during low-bitrate transmission, and feed the decoded non-reflectance compressed point clouds into a student detector. The discarded reflectance is then reconstructed by a geometry-based reflectance prediction (RP) module within the student detector for precise detection. A teacher detector with the same structure as the student detector is designed for performing reflectance knowledge distillation (RKD) and detection knowledge distillation (DKD) from raw to compressed point clouds. Our cross-source distillation training strategy (CDTS) equips the student detector with robustness to low-quality compressed data while preserving the accuracy benefits of raw data through transferred distillation knowledge. Experimental results on the KITTI and DAIR-V2X-V datasets demonstrate that our method can boost detection accuracy for compressed point clouds across multiple code rates. We will release the code publicly at https://github.com/HaoJing-SX/RPKD.", "date": "2025-05-23", "categories": []} {"id": "2508.16069", "title": "Voxel Densification for Serialized 3D Object Detection: Mitigating Sparsity via Pre-serialization Expansion", "abstract": "Recent advances in point cloud object detection have increasingly adopted Transformer-based and State Space Models (SSMs) to capture long-range dependencies. However, these serialized frameworks strictly maintain the consistency of input and output voxel dimensions, inherently lacking the capability for voxel expansion. This limitation hinders performance, as expanding the voxel set is known to significantly enhance detection accuracy, particularly for sparse foreground objects. To bridge this gap, we propose a novel Voxel Densification Module (VDM). Unlike standard convolutional stems, VDM is explicitly designed to promote pre-serialization spatial expansion. It leverages sparse 3D convolutions to propagate foreground semantics to neighboring empty voxels, effectively densifying the feature representation before it is flattened into a sequence. Simultaneously, VDM incorporates residual sparse blocks to aggregate fine-grained local context, ensuring rich geometric feature extraction. To balance the computational overhead of increased voxel density, we introduce a strategic cascaded downsampling mechanism. We integrate VDM into both Transformer-based (DSVT) and SSM-based (LION) detectors. Extensive experiments demonstrate that VDM consistently improves detection accuracy across multiple benchmarks. Specifically, our method achieves 74.8 mAPH (L2) on the Waymo validation set and 70.5 mAP on the nuScenes test set. Furthermore, it attains 42.6 mAP on the Argoverse 2 validation set and 67.6 mAP on the ONCE validation set, consistently outperforming the baseline models. The source code will be made publicly available at https://github.com/qifeng22/VDM.", "date": "2025-08-22", "categories": []} {"id": "2504.13647", "title": "An Efficient LiDAR-Camera Fusion Network for Multi-Class 3D Dynamic Object Detection and Trajectory Prediction", "abstract": "Service mobile robots are often required to avoid dynamic objects while performing their tasks, but they usually have only limited computational resources. To further advance the practical application of service robots in complex dynamic environments, we propose an efficient multi-modal framework for 3D object detection and trajectory prediction, which synergistically integrates LiDAR and camera inputs to achieve real-time perception of pedestrians, vehicles, and riders in 3D space.The framework incorporates two novel models: 1) a Unified modality detector with Mamba and Transformer (UniMT) for object detection, which achieves high-accuracy object detection with fast inference speed, and 2) a Reference Trajectory-based Multi-Class Transformer (RTMCT) for efficient and diverse trajectory prediction of multi-class objects with flexible-length trajectories. Evaluations on the CODa benchmark demonstrate that our method outperforms existing ones in both detection (+3.71\\% in mAP) and trajectory prediction (-0.408m in minADE$_5$ of pedestrians) metrics. Furthermore, on the challenging nuScenes detection benchmark, our detection model achieves competitive performance among LiDAR-camera fusion methods, with a mAP of 72.7\\% and NDS of 75.3\\%. Remarkably, the system demonstrates exceptional generalizability and practical deployment potential. When transferred and implemented on a wheelchair robot with an entry-level NVIDIA RTX 3060 GPU, it achieves real-time inference at 13.9 frames per second (FPS) with satisfactory accuracy. To facilitate reproducibility and practical deployment, we release the related code of the method at \\href{https://github.com/TossherO/3D_Perception}{https://github.com/TossherO/3D\\_Perception} and its ROS inference version at \\href{https://github.com/TossherO/ros_packages}{https://github.com/TossherO/ros\\_packages}.", "date": "2025-04-18", "categories": []} {"id": "2602.20632", "title": "Boosting Instance Awareness via Cross-View Correlation with 4D Radar and Camera for 3D Object Detection", "abstract": "4D millimeter-wave radar has emerged as a promising sensing modality for autonomous driving due to its robustness and affordability. However, its sparse and weak geometric cues make reliable instance activation difficult, limiting the effectiveness of existing radar-camera fusion paradigms. BEV-level fusion offers global scene understanding but suffers from weak instance focus, while perspective-level fusion captures instance details but lacks holistic context. To address these limitations, we propose SIFormer, a scene-instance aware transformer for 3D object detection using 4D radar and camera. SIFormer first suppresses background noise during view transformation through segmentation- and depth-guided localization. It then introduces a cross-view activation mechanism that injects 2D instance cues into BEV space, enabling reliable instance awareness under weak radar geometry. Finally, a transformer-based fusion module aggregates complementary image semantics and radar geometry for robust perception. As a result, with the aim of enhancing instance awareness, SIFormer bridges the gap between the two paradigms, combining their complementary strengths to address inherent sparse nature of radar and improve detection accuracy. Experiments demonstrate that SIFormer achieves state-of-the-art performance on View-of-Delft, TJ4DRadSet and NuScenes datasets. Source code is available at github.com/shawnnnkb/SIFormer.", "date": "2026-02-24", "categories": []} {"id": "2602.20653", "title": "SD4R: Sparse-to-Dense Learning for 3D Object Detection with 4D Radar", "abstract": "4D radar measurements offer an affordable and weather-robust solution for 3D perception. However, the inherent sparsity and noise of radar point clouds present significant challenges for accurate 3D object detection, underscoring the need for effective and robust point clouds densification. Despite recent progress, existing densification methods often fail to address the extreme sparsity of 4D radar point clouds and exhibit limited robustness when processing scenes with a small number of points. In this paper, we propose SD4R, a novel framework that transforms sparse radar point clouds into dense representations. SD4R begins by utilizing a foreground point generator (FPG) to mitigate noise propagation and produce densified point clouds. Subsequently, a logit-query encoder (LQE) enhances conventional pillarization, resulting in robust feature representations. Through these innovations, our SD4R demonstrates strong capability in both noise reduction and foreground point densification. Extensive experiments conducted on the publicly available View-of-Delft dataset demonstrate that SD4R achieves state-of-the-art performance. Source code is available at https://github.com/lancelot0805/SD4R.", "date": "2026-02-24", "categories": []} {"id": "2511.20648", "title": "LocateAnything3D: Vision-Language 3D Detection with Chain-of-Sight", "abstract": "To act in the world, a model must name what it sees and know where it is in 3D. Today's vision-language models (VLMs) excel at open-ended 2D description and grounding, yet multi-object 3D detection remains largely missing from the VLM toolbox. We present LocateAnything3D, a VLM-native recipe that casts 3D detection as a next-token prediction problem. The key is a short, explicit Chain-of-Sight (CoS) sequence that mirrors how human reason from images: find an object in 2D, then infer its distance, size, and pose. The decoder first emits 2D detections as a visual chain-of-thought, then predicts 3D boxes under an easy-to-hard curriculum: across objects, a near-to-far order reduces early ambiguity and matches ego-centric utility; within each object, a center-from-camera, dimensions, and rotation factorization ranks information by stability and learnability. This VLM-native interface preserves open-vocabulary and visual-prompting capability without specialized heads. On the challenging Omni3D benchmark, our model achieves state-of-the-art results, with 38.90 AP_3D, surpassing the previous best by +13.98 absolute improvement even when the baseline is given ground-truth 2D boxes. It also generalizes zero-shot to held-out categories with strong robustness. By turning 3D detection into a disciplined next-token problem, LocateAnything3D offers a practical foundation for models to perceive in 3D.", "date": "2025-11-25", "categories": []} {"id": "2602.19536", "title": "Fore-Mamba3D: Mamba-based Foreground-Enhanced Encoding for 3D Object Detection", "abstract": "Linear modeling methods like Mamba have been merged as the effective backbone for the 3D object detection task. However, previous Mamba-based methods utilize the bidirectional encoding for the whole non-empty voxel sequence, which contains abundant useless background information in the scenes. Though directly encoding foreground voxels appears to be a plausible solution, it tends to degrade detection performance. We attribute this to the response attenuation and restricted context representation in the linear modeling for fore-only sequences. To address this problem, we propose a novel backbone, termed Fore-Mamba3D, to focus on the foreground enhancement by modifying Mamba-based encoder. The foreground voxels are first sampled according to the predicted scores. Considering the response attenuation existing in the interaction of foreground voxels across different instances, we design a regional-to-global slide window (RGSW) to propagate the information from regional split to the entire sequence. Furthermore, a semantic-assisted and state spatial fusion module (SASFMamba) is proposed to enrich contextual representation by enhancing semantic and geometric awareness within the Mamba model. Our method emphasizes foreground-only encoding and alleviates the distance-based and causal dependencies in the linear autoregression model. The superior performance across various benchmarks demonstrates the effectiveness of Fore-Mamba3D in the 3D object detection task.", "date": "2026-02-23", "categories": []} {"id": "2602.19994", "title": "RADE-Net: Robust Attention Network for Radar-Only Object Detection in Adverse Weather", "abstract": "Automotive perception systems are obligated to meet high requirements. While optical sensors such as Camera and Lidar struggle in adverse weather conditions, Radar provides a more robust perception performance, effectively penetrating fog, rain, and snow. Since full Radar tensors have large data sizes and very few datasets provide them, most Radar-based approaches work with sparse point clouds or 2D projections, which can result in information loss. Additionally, deep learning methods show potential to extract richer and more dense features from low level Radar data and therefore significantly increase the perception performance. Therefore, we propose a 3D projection method for fast-Fourier-transformed 4D Range-Azimuth-Doppler-Elevation (RADE) tensors. Our method preserves rich Doppler and Elevation features while reducing the required data size for a single frame by 91.9% compared to a full tensor, thus achieving higher training and inference speed as well as lower model complexity. We introduce RADE-Net, a lightweight model tailored to 3D projections of the RADE tensor. The backbone enables exploitation of low-level and high-level cues of Radar tensors with spatial and channel-attention. The decoupled detection heads predict object center-points directly in the Range-Azimuth domain and regress rotated 3D bounding boxes from rich feature maps in the cartesian scene. We evaluate the model on scenes with multiple different road users and under various weather conditions on the large-scale K-Radar dataset and achieve a 16.7% improvement compared to their baseline, as well as 6.5% improvement over current Radar-only models. Additionally, we outperform several Lidar approaches in scenarios with adverse weather conditions. The code is available under https://github.com/chr-is-tof/RADE-Net.", "date": "2026-02-23", "categories": []} {"id": "2602.18394", "title": "Self-Aware Object Detection via Degradation Manifolds", "abstract": "Object detectors achieve strong performance under nominal imaging conditions but can fail silently when exposed to blur, noise, compression, adverse weather, or resolution changes. In safety-critical settings, it is therefore insufficient to produce predictions without assessing whether the input remains within the detector's nominal operating regime. We refer to this capability as self-aware object detection.\n We introduce a degradation-aware self-awareness framework based on degradation manifolds, which explicitly structure a detector's feature space according to image degradation rather than semantic content. Our method augments a standard detection backbone with a lightweight embedding head trained via multi-layer contrastive learning. Images sharing the same degradation composition are pulled together, while differing degradation configurations are pushed apart, yielding a geometrically organized representation that captures degradation type and severity without requiring degradation labels or explicit density modeling.\n To anchor the learned geometry, we estimate a pristine prototype from clean training embeddings, defining a nominal operating point in representation space. Self-awareness emerges as geometric deviation from this reference, providing an intrinsic, image-level signal of degradation-induced shift that is independent of detection confidence.\n Extensive experiments on synthetic corruption benchmarks, cross-dataset zero-shot transfer, and natural weather-induced distribution shifts demonstrate strong pristine-degraded separability, consistent behavior across multiple detector architectures, and robust generalization under semantic shift. These results suggest that degradation-aware representation geometry provides a practical and detector-agnostic foundation.", "date": "2026-02-20", "categories": []} {"id": "2508.01423", "title": "3DRot: Rediscovering the Missing Primitive for RGB-Based 3D Augmentation", "abstract": "RGB-based 3D tasks, e.g., 3D detection, depth estimation, 3D keypoint estimation, still suffer from scarce, expensive annotations and a thin augmentation toolbox, since many image transforms, including rotations and warps, disrupt geometric consistency. While horizontal flipping and color jitter are standard, rigorous 3D rotation augmentation has surprisingly remained absent from RGB-based pipelines, largely due to the misconception that it requires scene depth or scene reconstruction. In this paper, we introduce 3DRot, a plug-and-play augmentation that rotates and mirrors images about the camera's optical center while synchronously updating RGB images, camera intrinsics, object poses, and 3D annotations to preserve projective geometry, achieving geometry-consistent rotations and reflections without relying on any scene depth. We first validate 3DRot on a classical RGB-based 3D task, monocular 3D detection. On SUN RGB-D, inserting 3DRot into a frozen DINO-X + Cube R-CNN pipeline raises $IoU_{3D}$ from 43.21 to 44.51, cuts rotation error (ROT) from 22.91$^\\circ$ to 20.93$^\\circ$, and boosts $mAP_{0.5}$ from 35.70 to 38.11; smaller but consistent gains appear on a cross-domain IN10 split. Beyond monocular detection, adding 3DRot on top of the standard BTS augmentation schedule further improves NYU Depth v2 from 0.1783 to 0.1685 in abs-rel (and 0.7472 to 0.7548 in $\\delta<1.25$), and reduces cross-dataset error on SUN RGB-D. On KITTI, applying the same camera-centric rotations in MVX-Net (LiDAR+RGB) raises moderate 3D AP from about 63.85 to 65.16 while remaining compatible with standard 3D augmentations.", "date": "2025-08-02", "categories": []} {"id": "2411.02747", "title": "Efficient Feature Aggregation and Scale-Aware Regression for Monocular 3D Object Detection", "abstract": "Monocular 3D object detection has attracted great attention due to simplicity and low cost. Existing methods typically follow conventional 2D detection paradigms, first locating object centers and then predicting 3D attributes via neighboring features. However, these methods predominantly rely on progressive cross-scale feature aggregation and focus solely on local information, which may result in a lack of global awareness and the omission of small-scale objects. In addition, due to large variation in object scales across different scenes and depths, inaccurate receptive fields often lead to background noise and degraded feature representation. To address these issues, we introduces MonoASRH, a novel monocular 3D detection framework composed of Efficient Hybrid Feature Aggregation Module (EH-FAM) and Adaptive Scale-Aware 3D Regression Head (ASRH). Specifically, EH-FAM employs multi-head attention with a global receptive field to extract semantic features for small-scale objects and leverages lightweight convolutional modules to efficiently aggregate visual features across different scales. The ASRH encodes 2D bounding box dimensions and then fuses scale features with the semantic features aggregated by EH-FAM through a scale-semantic feature fusion module. The scale-semantic feature fusion module guides ASRH in learning dynamic receptive field offsets, incorporating scale priors into 3D position prediction for better scale-awareness. Extensive experiments on the KITTI and Waymo datasets demonstrate that MonoASRH achieves state-of-the-art performance.", "date": "2024-11-05", "categories": []} {"id": "2602.08126", "title": "MambaFusion: Adaptive State-Space Fusion for Multimodal 3D Object Detection", "abstract": "Reliable 3D object detection is fundamental to autonomous driving, and multimodal fusion algorithms using cameras and LiDAR remain a persistent challenge. Cameras provide dense visual cues but ill posed depth; LiDAR provides a precise 3D structure but sparse coverage. Existing BEV-based fusion frameworks have made good progress, but they have difficulties including inefficient context modeling, spatially invariant fusion, and reasoning under uncertainty. We introduce MambaFusion, a unified multi-modal detection framework that achieves efficient, adaptive, and physically grounded 3D perception. MambaFusion interleaves selective state-space models (SSMs) with windowed transformers to propagate the global context in linear time while preserving local geometric fidelity. A multi-modal token alignment (MTA) module and reliability-aware fusion gates dynamically re-weight camera-LiDAR features based on spatial confidence and calibration consistency. Finally, a structure-conditioned diffusion head integrates graph-based reasoning with uncertainty-aware denoising, enforcing physical plausibility, and calibrated confidence. MambaFusion establishes new state-of-the-art performance on nuScenes benchmarks while operating with linear-time complexity. The framework demonstrates that coupling SSM-based efficiency with reliability-driven fusion yields robust, temporally stable, and interpretable 3D perception for real-world autonomous driving systems.", "date": "2026-02-08", "categories": []} {"id": "2403.02818", "title": "Are Dense Labels Always Necessary for 3D Object Detection from Point Cloud?", "abstract": "Current state-of-the-art (SOTA) 3D object detection methods often require a large amount of 3D bounding box annotations for training. However, collecting such large-scale densely-supervised datasets is notoriously costly. To reduce the cumbersome data annotation process, we propose a novel sparsely-annotated framework, in which we just annotate one 3D object per scene. Such a sparse annotation strategy could significantly reduce the heavy annotation burden, while inexact and incomplete sparse supervision may severely deteriorate the detection performance. To address this issue, we develop the SS3D++ method that alternatively improves 3D detector training and confident fully-annotated scene generation in a unified learning scheme. Using sparse annotations as seeds, we progressively generate confident fully-annotated scenes based on designing a missing-annotated instance mining module and reliable background mining module. Our proposed method produces competitive results when compared with SOTA weakly-supervised methods using the same or even more annotation costs. Besides, compared with SOTA fully-supervised methods, we achieve on-par or even better performance on the KITTI dataset with about 5x less annotation cost, and 90% of their performance on the Waymo dataset with about 15x less annotation cost. The additional unlabeled training scenes could further boost the performance.", "date": "2024-03-05", "categories": []} {"id": "2602.10806", "title": "DMP-3DAD: Cross-Category 3D Anomaly Detection via Realistic Depth Map Projection with Few Normal Samples", "abstract": "Cross-category anomaly detection for 3D point clouds aims to determine whether an unseen object belongs to a target category using only a few normal examples. Most existing methods rely on category-specific training, which limits their flexibility in few-shot scenarios. In this paper, we propose DMP-3DAD, a training-free framework for cross-category 3D anomaly detection based on multi-view realistic depth map projection. Specifically, by converting point clouds into a fixed set of realistic depth images, our method leverages a frozen CLIP visual encoder to extract multi-view representations and performs anomaly detection via weighted feature similarity, which does not require any fine-tuning or category-dependent adaptation. Extensive experiments on the ShapeNetPart dataset demonstrate that DMP-3DAD achieves state-of-the-art performance under few-shot setting. The results show that the proposed approach provides a simple yet effective solution for practical cross-category 3D anomaly detection.", "date": "2026-02-11", "categories": []} {"id": "2512.18187", "title": "ALIGN: Advanced Query Initialization with LiDAR-Image Guidance for Occlusion-Robust 3D Object Detection", "abstract": "Recent query-based 3D object detection methods using camera and LiDAR inputs have shown strong performance, but existing query initialization strategies,such as random sampling or BEV heatmap-based sampling, often result in inefficient query usage and reduced accuracy, particularly for occluded or crowded objects. To address this limitation, we propose ALIGN (Advanced query initialization with LiDAR and Image GuidaNce), a novel approach for occlusion-robust, object-aware query initialization. Our model consists of three key components: (i) Occlusion-aware Center Estimation (OCE), which integrates LiDAR geometry and image semantics to estimate object centers accurately (ii) Adaptive Neighbor Sampling (ANS), which generates object candidates from LiDAR clustering and supplements each object by sampling spatially and semantically aligned points around it and (iii) Dynamic Query Balancing (DQB), which adaptively balances queries between foreground and background regions. Our extensive experiments on the nuScenes benchmark demonstrate that ALIGN consistently improves performance across multiple state-of-the-art detectors, achieving gains of up to +0.9 mAP and +1.2 NDS, particularly in challenging scenes with occlusions or dense crowds. Our code will be publicly available upon publication.", "date": "2025-12-20", "categories": []} {"id": "2602.05538", "title": "A Comparative Study of 3D Person Detection: Sensor Modalities and Robustness in Diverse Indoor and Outdoor Environments", "abstract": "Accurate 3D person detection is critical for safety in applications such as robotics, industrial monitoring, and surveillance. This work presents a systematic evaluation of 3D person detection using camera-only, LiDAR-only, and camera-LiDAR fusion. While most existing research focuses on autonomous driving, we explore detection performance and robustness in diverse indoor and outdoor scenes using the JRDB dataset. We compare three representative models - BEVDepth (camera), PointPillars (LiDAR), and DAL (camera-LiDAR fusion) - and analyze their behavior under varying occlusion and distance levels. Our results show that the fusion-based approach consistently outperforms single-modality models, particularly in challenging scenarios. We further investigate robustness against sensor corruptions and misalignments, revealing that while DAL offers improved resilience, it remains sensitive to sensor misalignment and certain LiDAR-based corruptions. In contrast, the camera-based BEVDepth model showed the lowest performance and was most affected by occlusion, distance, and noise. Our findings highlight the importance of utilizing sensor fusion for enhanced 3D person detection, while also underscoring the need for ongoing research to address the vulnerabilities inherent in these systems.", "date": "2026-02-05", "categories": []} {"id": "2602.06406", "title": "Point Virtual Transformer", "abstract": "LiDAR-based 3D object detectors often struggle to detect far-field objects due to the sparsity of point clouds at long ranges, which limits the availability of reliable geometric cues. To address this, prior approaches augment LiDAR data with depth-completed virtual points derived from RGB images; however, directly incorporating all virtual points leads to increased computational cost and introduces challenges in effectively fusing real and virtual information. We present Point Virtual Transformer (PointViT), a transformer-based 3D object detection framework that jointly reasons over raw LiDAR points and selectively sampled virtual points. The framework examines multiple fusion strategies, ranging from early point-level fusion to BEV-based gated fusion, and analyses their trade-offs in terms of accuracy and efficiency. The fused point cloud is voxelized and encoded using sparse convolutions to form a BEV representation, from which a compact set of high-confidence object queries is initialised and refined through a transformer-based context aggregation module. Experiments on the KITTI benchmark report 91.16% 3D AP, 95.94% BEV AP, and 99.36% AP on the KITTI 2D detection benchmark for the Car class.", "date": "2026-02-04", "categories": []} {"id": "2602.05557", "title": "PIRATR: Parametric Object Inference for Robotic Applications with Transformers in 3D Point Clouds", "abstract": "We present PIRATR, an end-to-end 3D object detection framework for robotic use cases in point clouds. Extending PI3DETR, our method streamlines parametric 3D object detection by jointly estimating multi-class 6-DoF poses and class-specific parametric attributes directly from occlusion-affected point cloud data. This formulation enables not only geometric localization but also the estimation of task-relevant properties for parametric objects, such as a gripper's opening, where the 3D model is adjusted according to simple, predefined rules. The architecture employs modular, class-specific heads, making it straightforward to extend to novel object types without re-designing the pipeline. We validate PIRATR on an automated forklift platform, focusing on three structurally and functionally diverse categories: crane grippers, loading platforms, and pallets. Trained entirely in a synthetic environment, PIRATR generalizes effectively to real outdoor LiDAR scans, achieving a detection mAP of 0.919 without additional fine-tuning. PIRATR establishes a new paradigm of pose-aware, parameterized perception. This bridges the gap between low-level geometric reasoning and actionable world models, paving the way for scalable, simulation-trained perception systems that can be deployed in dynamic robotic environments. Code available at https://github.com/swingaxe/piratr.", "date": "2026-02-05", "categories": []} {"id": "2602.00450", "title": "Model Optimization for Multi-Camera 3D Detection and Tracking", "abstract": "Outside-in multi-camera perception is increasingly important in indoor environments, where networks of static cameras must support multi-target tracking under occlusion and heterogeneous viewpoints. We evaluate Sparse4D, a query-based spatiotemporal 3D detection and tracking framework that fuses multi-view features in a shared world frame and propagates sparse object queries via instance memory. We study reduced input frame rates, post-training quantization (INT8 and FP8), transfer to the WILDTRACK benchmark, and Transformer Engine mixed-precision fine-tuning. To better capture identity stability, we report Average Track Duration (AvgTrackDur), which measures identity persistence in seconds. Sparse4D remains stable under moderate FPS reductions, but below 2 FPS, identity association collapses even when detections are stable. Selective quantization of the backbone and neck offers the best speed-accuracy trade-off, while attention-related modules are consistently sensitive to low precision. On WILDTRACK, low-FPS pretraining yields large zero-shot gains over the base checkpoint, while small-scale fine-tuning provides limited additional benefit. Transformer Engine mixed precision reduces latency and improves camera scalability, but can destabilize identity propagation, motivating stability-aware validation.", "date": "2026-01-31", "categories": []} {"id": "2601.12638", "title": "Mixed Precision PointPillars for Efficient 3D Object Detection with TensorRT", "abstract": "LIDAR 3D object detection is one of the important tasks for autonomous vehicles. Ensuring that this task operates in real-time is crucial. Toward this, model quantization can be used to accelerate the runtime. However, directly applying model quantization often leads to performance degradation due to LIDAR's wide numerical distributions and extreme outliers. To address the wide numerical distribution, we proposed a mixed precision framework designed for PointPillars. Our framework first searches for sensitive layers with post-training quantization (PTQ) by quantizing one layer at a time to 8-bit integer (INT8) and evaluating each model for average precision (AP). The top-k most sensitive layers are assigned as floating point (FP). Combinations of these layers are greedily searched to produce candidate mixed precision models, which are finalized with either PTQ or quantization-aware training (QAT). Furthermore, to handle outliers, we observe that using a very small number of calibration data reduces the likelihood of encountering outliers, thereby improving PTQ performance. Our methods provides mixed precision models without training in the PTQ pipeline, while our QAT pipeline achieves the performance competitive to FP models. With TensorRT deployment, our mixed precision models offer less latency by up to 2.538 times compared to FP32 models.", "date": "2026-01-19", "categories": []} {"id": "2602.00149", "title": "SDCM: Simulated Densifying and Compensatory Modeling Fusion for Radar-Vision 3-D Object Detection in Internet of Vehicles", "abstract": "3-D object detection based on 4-D radar-vision is an important part in Internet of Vehicles (IoV). However, there are two challenges which need to be faced. First, the 4-D radar point clouds are sparse, leading to poor 3-D representation. Second, vision datas exhibit representation degradation under low-light, long distance detection and dense occlusion scenes, which provides unreliable texture information during fusion stage. To address these issues, a framework named SDCM is proposed, which contains Simulated Densifying and Compensatory Modeling Fusion for radar-vision 3-D object detection in IoV. Firstly, considering point generation based on Gaussian simulation of key points obtained from 3-D Kernel Density Estimation (3-D KDE), and outline generation based on curvature simulation, Simulated Densifying (SimDen) module is designed to generate dense radar point clouds. Secondly, considering that radar data could provide more real time information than vision data, due to the all-weather property of 4-D radar. Radar Compensatory Mapping (RCM) module is designed to reduce the affects of vision datas' representation degradation. Thirdly, considering that feature tensor difference values contain the effective information of every modality, which could be extracted and modeled for heterogeneity reduction and modalities interaction, Mamba Modeling Interactive Fusion (MMIF) module is designed for reducing heterogeneous and achieving interactive Fusion. Experiment results on the VoD, TJ4DRadSet and Astyx HiRes 2019 dataset show that SDCM achieves best performance with lower parameter quantity and faster inference speed. Our code will be available.", "date": "2026-01-29", "categories": []} {"id": "2601.22616", "title": "UniGeo: A Unified 3D Indoor Object Detection Framework Integrating Geometry-Aware Learning and Dynamic Channel Gating", "abstract": "The growing adoption of robotics and augmented reality in real-world applications has driven considerable research interest in 3D object detection based on point clouds. While previous methods address unified training across multiple datasets, they fail to model geometric relationships in sparse point cloud scenes and ignore the feature distribution in significant areas, which ultimately restricts their performance. To deal with this issue, a unified 3D indoor detection framework, called UniGeo, is proposed. To model geometric relations in scenes, we first propose a geometry-aware learning module that establishes a learnable mapping from spatial relationships to feature weights, which enabes explicit geometric feature enhancement. Then, to further enhance point cloud feature representation, we propose a dynamic channel gating mechanism that leverages learnable channel-wise weighting. This mechanism adaptively optimizes features generated by the sparse 3D U-Net network, significantly enhancing key geometric information. Extensive experiments on six different indoor scene datasets clearly validate the superior performance of our method.", "date": "2026-01-30", "categories": []} {"id": "2512.08557", "title": "SSCATeR: Sparse Scatter-Based Convolution Algorithm with Temporal Data Recycling for Real-Time 3D Object Detection in LiDAR Point Clouds", "abstract": "This work leverages the continuous sweeping motion of LiDAR scanning to concentrate object detection efforts on specific regions that receive a change in point data from one frame to another. We achieve this by using a sliding time window with short strides and consider the temporal dimension by storing convolution results between passes. This allows us to ignore unchanged regions, significantly reducing the number of convolution operations per forward pass without sacrificing accuracy. This data reuse scheme introduces extreme sparsity to detection data. To exploit this sparsity, we extend our previous work on scatter-based convolutions to allow for data reuse, and as such propose Sparse Scatter-Based Convolution Algorithm with Temporal Data Recycling (SSCATeR). This operation treats incoming LiDAR data as a continuous stream and acts only on the changing parts of the point cloud. By doing so, we achieve the same results with as much as a 6.61-fold reduction in processing time. Our test results show that the feature maps output by our method are identical to those produced by traditional sparse convolution techniques, whilst greatly increasing the computational efficiency of the network.", "date": "2025-12-09", "categories": []} {"id": "2601.20720", "title": "Li-ViP3D++: Query-Gated Deformable Camera-LiDAR Fusion for End-to-End Perception and Trajectory Prediction", "abstract": "End-to-end perception and trajectory prediction from raw sensor data is one of the key capabilities for autonomous driving. Modular pipelines restrict information flow and can amplify upstream errors. Recent query-based, fully differentiable perception-and-prediction (PnP) models mitigate these issues, yet the complementarity of cameras and LiDAR in the query-space has not been sufficiently explored. Models often rely on fusion schemes that introduce heuristic alignment and discrete selection steps which prevent full utilization of available information and can introduce unwanted bias. We propose Li-ViP3D++, a query-based multimodal PnP framework that introduces Query-Gated Deformable Fusion (QGDF) to integrate multi-view RGB and LiDAR in query space. QGDF (i) aggregates image evidence via masked attention across cameras and feature levels, (ii) extracts LiDAR context through fully differentiable BEV sampling with learned per-query offsets, and (iii) applies query-conditioned gating to adaptively weight visual and geometric cues per agent. The resulting architecture jointly optimizes detection, tracking, and multi-hypothesis trajectory forecasting in a single end-to-end model. On nuScenes, Li-ViP3D++ improves end-to-end behavior and detection quality, achieving higher EPA (0.335) and mAP (0.502) while substantially reducing false positives (FP ratio 0.147), and it is faster than the prior Li-ViP3D variant (139.82 ms vs. 145.91 ms). These results indicate that query-space, fully differentiable camera-LiDAR fusion can increase robustness of end-to-end PnP without sacrificing deployability.", "date": "2026-01-28", "categories": []} {"id": "2601.19314", "title": "Instance-Guided Radar Depth Estimation for 3D Object Detection", "abstract": "Accurate depth estimation is fundamental to 3D perception in autonomous driving, supporting tasks such as detection, tracking, and motion planning. However, monocular camera-based 3D detection suffers from depth ambiguity and reduced robustness under challenging conditions. Radar provides complementary advantages such as resilience to poor lighting and adverse weather, but its sparsity and low resolution limit its direct use in detection frameworks. This motivates the need for effective Radar-camera fusion with improved preprocessing and depth estimation strategies. We propose an end-to-end framework that enhances monocular 3D object detection through two key components. First, we introduce InstaRadar, an instance segmentation-guided expansion method that leverages pre-trained segmentation masks to enhance Radar density and semantic alignment, producing a more structured representation. InstaRadar achieves state-of-the-art results in Radar-guided depth estimation, showing its effectiveness in generating high-quality depth features. Second, we integrate the pre-trained RCDPT into the BEVDepth framework as a replacement for its depth module. With InstaRadar-enhanced inputs, the RCDPT integration consistently improves 3D detection performance. Overall, these components yield steady gains over the baseline BEVDepth model, demonstrating the effectiveness of InstaRadar and the advantage of explicit depth supervision in 3D object detection. Although the framework lags behind Radar-camera fusion models that directly extract BEV features, since Radar serves only as guidance rather than an independent feature stream, this limitation highlights potential for improvement. Future work will extend InstaRadar to point cloud-like representations and integrate a dedicated Radar branch with temporal cues for enhanced BEV fusion.", "date": "2026-01-27", "categories": []} {"id": "2512.12884", "title": "Cross-Level Sensor Fusion with Object Lists via Transformer for 3D Object Detection", "abstract": "In automotive sensor fusion systems, smart sensors and Vehicle-to-Everything (V2X) modules are commonly utilized. Sensor data from these systems are typically available only as processed object lists rather than raw sensor data from traditional sensors. Instead of processing other raw data separately and then fusing them at the object level, we propose an end-to-end cross-level fusion concept with Transformer, which integrates highly abstract object list information with raw camera images for 3D object detection. Object lists are fed into a Transformer as denoising queries and propagated together with learnable queries through the latter feature aggregation process. Additionally, a deformable Gaussian mask, derived from the positional and size dimensional priors from the object lists, is explicitly integrated into the Transformer decoder. This directs attention toward the target area of interest and accelerates model training convergence. Furthermore, as there is no public dataset containing object lists as a standalone modality, we propose an approach to generate pseudo object lists from ground-truth bounding boxes by simulating state noise and false positives and negatives. As the first work to conduct cross-level fusion, our approach shows substantial performance improvements over the vision-based baseline on the nuScenes dataset. It demonstrates its generalization capability over diverse noise levels of simulated object lists and real detectors.", "date": "2025-12-14", "categories": []} {"id": "2601.12994", "title": "AsyncBEV: Cross-modal Flow Alignment in Asynchronous 3D Object Detection", "abstract": "In autonomous driving, multi-modal perception tasks like 3D object detection typically rely on well-synchronized sensors, both at training and inference. However, despite the use of hardware- or software-based synchronization algorithms, perfect synchrony is rarely guaranteed: Sensors may operate at different frequencies, and real-world factors such as network latency, hardware failures, or processing bottlenecks often introduce time offsets between sensors. Such asynchrony degrades perception performance, especially for dynamic objects. To address this challenge, we propose AsyncBEV, a trainable lightweight and generic module to improve the robustness of 3D Birds' Eye View (BEV) object detection models against sensor asynchrony. Inspired by scene flow estimation, AsyncBEV first estimates the 2D flow from the BEV features of two different sensor modalities, taking into account the known time offset between these sensor measurements. The predicted feature flow is then used to warp and spatially align the feature maps, which we show can easily be integrated into different current BEV detector architectures (e.g., BEV grid-based and token-based). Extensive experiments demonstrate AsyncBEV improves robustness against both small and large asynchrony between LiDAR or camera sensors in both the token-based CMT and grid-based UniBEV, especially for dynamic objects. We significantly outperform the ego motion compensated CMT and UniBEV baselines, notably by $16.6$ % and $11.9$ % NDS on dynamic objects in the worst-case scenario of a $0.5 s$ time offset. Code will be released upon acceptance.", "date": "2026-01-19", "categories": []} {"id": "2601.13386", "title": "Leveraging Transformer Decoder for Automotive Radar Object Detection", "abstract": "In this paper, we present a Transformer-based architecture for 3D radar object detection that uses a novel Transformer Decoder as the prediction head to directly regress 3D bounding boxes and class scores from radar feature representations. To bridge multi-scale radar features and the decoder, we propose Pyramid Token Fusion (PTF), a lightweight module that converts a feature pyramid into a unified, scale-aware token sequence. By formulating detection as a set prediction problem with learnable object queries and positional encodings, our design models long-range spatial-temporal correlations and cross-feature interactions. This approach eliminates dense proposal generation and heuristic post-processing such as extensive non-maximum suppression (NMS) tuning. We evaluate the proposed framework on the RADDet, where it achieves significant improvements over state-of-the-art radar-only baselines.", "date": "2026-01-19", "categories": []} {"id": "2601.10819", "title": "A Unified 3D Object Perception Framework for Real-Time Outside-In Multi-Camera Systems", "abstract": "Accurate 3D object perception and multi-target multi-camera (MTMC) tracking are fundamental for the digital transformation of industrial infrastructure. However, transitioning \"inside-out\" autonomous driving models to \"outside-in\" static camera networks presents significant challenges due to heterogeneous camera placements and extreme occlusion. In this paper, we present an adapted Sparse4D framework specifically optimized for large-scale infrastructure environments. Our system leverages absolute world-coordinate geometric priors and introduces an occlusion-aware ReID embedding module to maintain identity stability across distributed sensor networks. To bridge the Sim2Real domain gap without manual labeling, we employ a generative data augmentation strategy using the NVIDIA COSMOS framework, creating diverse environmental styles that enhance the model's appearance-invariance. Evaluated on the AI City Challenge 2025 benchmark, our camera-only framework achieves a state-of-the-art HOTA of $45.22$. Furthermore, we address real-time deployment constraints by developing an optimized TensorRT plugin for Multi-Scale Deformable Aggregation (MSDA). Our hardware-accelerated implementation achieves a $2.15\\times$ speedup on modern GPU architectures, enabling a single Blackwell-class GPU to support over 64 concurrent camera streams.", "date": "2026-01-15", "categories": []} {"id": "2512.22972", "title": "Wavelet-based Multi-View Fusion of 4D Radar Tensor and Camera for Robust 3D Object Detection", "abstract": "4D millimeter-wave (mmWave) radar has been widely adopted in autonomous driving and robot perception due to its low cost and all-weather robustness. However, point-cloud-based radar representations suffer from information loss due to multi-stage signal processing, while directly utilizing raw 4D radar tensors incurs prohibitive computational costs. To address these challenges, we propose WRCFormer, a novel 3D object detection framework that efficiently fuses raw 4D radar cubes with camera images via decoupled multi-view radar representations. Our approach introduces two key components: (1) A Wavelet Attention Module embedded in a wavelet-based Feature Pyramid Network (FPN), which enhances the representation of sparse radar signals and image data by capturing joint spatial-frequency features, thereby mitigating information loss while maintaining computational efficiency. (2) A Geometry-guided Progressive Fusion mechanism, a two-stage query-based fusion strategy that progressively aligns multi-view radar and visual features through geometric priors, enabling modality-agnostic and efficient integration without overwhelming computational overhead. Extensive experiments on the K-Radar benchmark show that WRCFormer achieves state-of-the-art performance, surpassing the best existing model by approximately 2.4% in all scenarios and 1.6% in sleet conditions, demonstrating strong robustness in adverse weather.", "date": "2025-12-28", "categories": []} {"id": "2601.09812", "title": "LCF3D: A Robust and Real-Time Late-Cascade Fusion Framework for 3D Object Detection in Autonomous Driving", "abstract": "Accurately localizing 3D objects like pedestrians, cyclists, and other vehicles is essential in Autonomous Driving. To ensure high detection performance, Autonomous Vehicles complement RGB cameras with LiDAR sensors, but effectively combining these data sources for 3D object detection remains challenging. We propose LCF3D, a novel sensor fusion framework that combines a 2D object detector on RGB images with a 3D object detector on LiDAR point clouds. By leveraging multimodal fusion principles, we compensate for inaccuracies in the LiDAR object detection network. Our solution combines two key principles: (i) late fusion, to reduce LiDAR False Positives by matching LiDAR 3D detections with RGB 2D detections and filtering out unmatched LiDAR detections; and (ii) cascade fusion, to recover missed objects from LiDAR by generating new 3D frustum proposals corresponding to unmatched RGB detections. Experiments show that LCF3D is beneficial for domain generalization, as it turns out to be successful in handling different sensor configurations between training and testing domains. LCF3D achieves significant improvements over LiDAR-based methods, particularly for challenging categories like pedestrians and cyclists in the KITTI dataset, as well as motorcycles and bicycles in nuScenes. Code can be downloaded from: https://github.com/CarloSgaravatti/LCF3D.", "date": "2026-01-14", "categories": []} {"id": "2504.18419", "title": "A Multimodal Hybrid Late-Cascade Fusion Network for Enhanced 3D Object Detection", "abstract": "We present a new way to detect 3D objects from multimodal inputs, leveraging both LiDAR and RGB cameras in a hybrid late-cascade scheme, that combines an RGB detection network and a 3D LiDAR detector. We exploit late fusion principles to reduce LiDAR False Positives, matching LiDAR detections with RGB ones by projecting the LiDAR bounding boxes on the image. We rely on cascade fusion principles to recover LiDAR False Negatives leveraging epipolar constraints and frustums generated by RGB detections of separate views. Our solution can be plugged on top of any underlying single-modal detectors, enabling a flexible training process that can take advantage of pre-trained LiDAR and RGB detectors, or train the two branches separately. We evaluate our results on the KITTI object detection benchmark, showing significant performance improvements, especially for the detection of Pedestrians and Cyclists.", "date": "2025-04-25", "categories": []} {"id": "2411.15016", "title": "MSSF: A 4D Radar and Camera Fusion Framework With Multi-Stage Sampling for 3D Object Detection in Autonomous Driving", "abstract": "As one of the automotive sensors that have emerged in recent years, 4D millimeter-wave radar has a higher resolution than conventional 3D radar and provides precise elevation measurements. But its point clouds are still sparse and noisy, making it challenging to meet the requirements of autonomous driving. Camera, as another commonly used sensor, can capture rich semantic information. As a result, the fusion of 4D radar and camera can provide an affordable and robust perception solution for autonomous driving systems. However, previous radar-camera fusion methods have not yet been thoroughly investigated, resulting in a large performance gap compared to LiDAR-based methods. Specifically, they ignore the feature-blurring problem and do not deeply interact with image semantic information. To this end, we present a simple but effective multi-stage sampling fusion (MSSF) network based on 4D radar and camera. On the one hand, we design a fusion block that can deeply interact point cloud features with image features, and can be applied to commonly used single-modal backbones in a plug-and-play manner. The fusion block encompasses two types, namely, simple feature fusion (SFF) and multiscale deformable feature fusion (MSDFF). The SFF is easy to implement, while the MSDFF has stronger fusion abilities. On the other hand, we propose a semantic-guided head to perform foreground-background segmentation on voxels with voxel feature re-weighting, further alleviating the problem of feature blurring. Extensive experiments on the View-of-Delft (VoD) and TJ4DRadset datasets demonstrate the effectiveness of our MSSF. Notably, compared to state-of-the-art methods, MSSF achieves a 7.0% and 4.0% improvement in 3D mean average precision on the VoD and TJ4DRadSet datasets, respectively. It even surpasses classical LiDAR-based methods on the VoD dataset.", "date": "2024-11-22", "categories": []} {"id": "2601.08174", "title": "Towards Cross-Platform Generalization: Domain Adaptive 3D Detection with Augmentation and Pseudo-Labeling", "abstract": "This technical report represents the award-winning solution to the Cross-platform 3D Object Detection task in the RoboSense2025 Challenge. Our approach is built upon PVRCNN++, an efficient 3D object detection framework that effectively integrates point-based and voxel-based features. On top of this foundation, we improve cross-platform generalization by narrowing domain gaps through tailored data augmentation and a self-training strategy with pseudo-labels. These enhancements enabled our approach to secure the 3rd place in the challenge, achieving a 3D AP of 62.67% for the Car category on the phase-1 target domain, and 58.76% and 49.81% for Car and Pedestrian categories respectively on the phase-2 target domain.", "date": "2026-01-13", "categories": []} {"id": "2511.17806", "title": "REXO: Indoor Multi-View Radar Object Detection via 3D Bounding Box Diffusion", "abstract": "Multi-view indoor radar perception has drawn attention due to its cost-effectiveness and low privacy risks. Existing methods often rely on {implicit} cross-view radar feature association, such as proposal pairing in RFMask or query-to-feature cross-attention in RETR, which can lead to ambiguous feature matches and degraded detection in complex indoor scenes. To address these limitations, we propose \\textbf{REXO} (multi-view Radar object dEtection with 3D bounding boX diffusiOn), which lifts the 2D bounding box (BBox) diffusion process of DiffusionDet into the 3D radar space. REXO utilizes these noisy 3D BBoxes to guide an {explicit} cross-view radar feature association, enhancing the cross-view radar-conditioned denoising process. By accounting for prior knowledge that the person is in contact with the ground, REXO reduces the number of diffusion parameters by determining them from this prior. Evaluated on two open indoor radar datasets, our approach surpasses state-of-the-art methods by a margin of +4.22 AP on the HIBER dataset and +11.02 AP on the MMVR dataset. The REXO implementation is available at https://github.com/merlresearch/radar-bbox-diffusion.", "date": "2025-11-21", "categories": []} {"id": "2503.06282", "title": "From Dataset to Real-world: General 3D Object Detection via Generalized Cross-domain Few-shot Learning", "abstract": "LiDAR-based 3D object detection models often struggle to generalize to real-world environments due to limited object diversity in existing datasets. To tackle it, we introduce the first generalized cross-domain few-shot (GCFS) task in 3D object detection, aiming to adapt a source-pretrained model to both common and novel classes in a new domain with only few-shot annotations. We propose a unified framework that learns stable target semantics under limited supervision by bridging 2D open-set semantics with 3D spatial reasoning. Specifically, an image-guided multi-modal fusion injects transferable 2D semantic cues into the 3D pipeline via vision-language models, while a physically-aware box search enhances 2D-to-3D alignment via LiDAR priors. To capture class-specific semantics from sparse data, we further introduce contrastive-enhanced prototype learning, which encodes few-shot instances into discriminative semantic anchors and stabilizes representation learning. Extensive experiments on GCFS benchmarks demonstrate the effectiveness and generality of our approach in realistic deployment settings.", "date": "2025-03-08", "categories": []} {"id": "2601.03617", "title": "Systematic Evaluation of Depth Backbones and Semantic Cues for Monocular Pseudo-LiDAR 3D Detection", "abstract": "Monocular 3D object detection offers a low-cost alternative to LiDAR, yet remains less accurate due to the difficulty of estimating metric depth from a single image. We systematically evaluate how depth backbones and feature engineering affect a monocular Pseudo-LiDAR pipeline on the KITTI validation split. Specifically, we compare NeWCRFs (supervised metric depth) against Depth Anything V2 Metric-Outdoor (Base) under an identical pseudo-LiDAR generation and PointRCNN detection protocol. NeWCRFs yields stronger downstream 3D detection, achieving 10.50\\% AP$_{3D}$ at IoU$=0.7$ on the Moderate split using grayscale intensity (Exp~2). We further test point-cloud augmentations using appearance cues (grayscale intensity) and semantic cues (instance segmentation confidence). Contrary to the expectation that semantics would substantially close the gap, these features provide only marginal gains, and mask-based sampling can degrade performance by removing contextual geometry. Finally, we report a depth-accuracy-versus-distance diagnostic using ground-truth 2D boxes (including Ped/Cyc), highlighting that coarse depth correctness does not fully predict strict 3D IoU. Overall, under an off-the-shelf LiDAR detector, depth-backbone choice and geometric fidelity dominate performance, outweighing secondary feature injection.", "date": "2026-01-07", "categories": []} {"id": "2601.01036", "title": "Mono3DV: Monocular 3D Object Detection with 3D-Aware Bipartite Matching and Variational Query DeNoising", "abstract": "While DETR-like architectures have demonstrated significant potential for monocular 3D object detection, they are often hindered by a critical limitation: the exclusion of 3D attributes from the bipartite matching process. This exclusion arises from the inherent ill-posed nature of 3D estimation from monocular image, which introduces instability during training. Consequently, high-quality 3D predictions can be erroneously suppressed by 2D-only matching criteria, leading to suboptimal results. To address this, we propose Mono3DV, a novel Transformer-based framework. Our approach introduces three key innovations. First, we develop a 3D-Aware Bipartite Matching strategy that directly incorporates 3D geometric information into the matching cost, resolving the misalignment caused by purely 2D criteria. Second, it is important to stabilize the Bipartite Matching to resolve the instability occurring when integrating 3D attributes. Therefore, we propose 3D-DeNoising scheme in the training phase. Finally, recognizing the gradient vanishing issue associated with conventional denoising techniques, we propose a novel Variational Query DeNoising mechanism to overcome this limitation, which significantly enhances model performance. Without leveraging any external data, our method achieves state-of-the-art results on the KITTI 3D object detection benchmark.", "date": "2026-01-03", "categories": []} {"id": "2601.01676", "title": "LabelAny3D: Label Any Object 3D in the Wild", "abstract": "Detecting objects in 3D space from monocular input is crucial for applications ranging from robotics to scene understanding. Despite advanced performance in the indoor and autonomous driving domains, existing monocular 3D detection models struggle with in-the-wild images due to the lack of 3D in-the-wild datasets and the challenges of 3D annotation. We introduce LabelAny3D, an \\emph{analysis-by-synthesis} framework that reconstructs holistic 3D scenes from 2D images to efficiently produce high-quality 3D bounding box annotations. Built on this pipeline, we present COCO3D, a new benchmark for open-vocabulary monocular 3D detection, derived from the MS-COCO dataset and covering a wide range of object categories absent from existing 3D datasets. Experiments show that annotations generated by LabelAny3D improve monocular 3D detection performance across multiple benchmarks, outperforming prior auto-labeling approaches in quality. These results demonstrate the promise of foundation-model-driven annotation for scaling up 3D recognition in realistic, open-world settings.", "date": "2026-01-04", "categories": []} {"id": "2506.06944", "title": "Towards Streaming LiDAR Object Detection with Point Clouds as Egocentric Sequences", "abstract": "Accurate and low-latency 3D object detection is essential for autonomous driving, where safety hinges on both rapid response and reliable perception. While rotating LiDAR sensors are widely adopted for their robustness and fidelity, current detectors face a trade-off: streaming methods process partial polar sectors on the fly for fast updates but suffer from limited visibility, cross-sector dependencies, and distortions from retrofitted Cartesian designs, whereas full-scan methods achieve higher accuracy but are bottlenecked by the inherent latency of a LiDAR revolution. We propose Polar-Fast-Cartesian-Full (PFCF), a hybrid detector that combines fast polar processing for intra-sector feature extraction with accurate Cartesian reasoning for full-scene understanding. Central to PFCF is a custom Mamba SSM-based streaming backbone with dimensionally-decomposed convolutions that avoids distortion-heavy planes, enabling parameter-efficient, translation-invariant, and distortion-robust polar representation learning. Local sector features are extracted via this backbone, then accumulated into a sector feature buffer to enable efficient inter-sector communication through a full-scan backbone. PFCF establishes a new Pareto frontier on the Waymo Open dataset, surpassing prior streaming baselines by 10% mAP and matching full-scan accuracy at twice the update rate. Code is available at \\href{https://github.com/meilongzhang/Polar-Hierarchical-Mamba}{https://github.com/meilongzhang/Polar-Hierarchical-Mamba}.", "date": "2025-06-07", "categories": []} {"id": "2512.24922", "title": "Semi-Supervised Diversity-Aware Domain Adaptation for 3D Object detection", "abstract": "3D object detectors are fundamental components of perception systems in autonomous vehicles. While these detectors achieve remarkable performance on standard autonomous driving benchmarks, they often struggle to generalize across different domains - for instance, a model trained in the U.S. may perform poorly in regions like Asia or Europe. This paper presents a novel lidar domain adaptation method based on neuron activation patterns, demonstrating that state-of-the-art performance can be achieved by annotating only a small, representative, and diverse subset of samples from the target domain if they are correctly selected. The proposed approach requires very small annotation budget and, when combined with post-training techniques inspired by continual learning prevent weight drift from the original model. Empirical evaluation shows that the proposed domain adaptation approach outperforms both linear probing and state-of-the-art domain adaptation techniques.", "date": "2025-12-31", "categories": []} {"id": "2512.22503", "title": "SCAFusion: A Multimodal 3D Detection Framework for Small Object Detection in Lunar Surface Exploration", "abstract": "Reliable and precise detection of small and irregular objects, such as meteor fragments and rocks, is critical for autonomous navigation and operation in lunar surface exploration. Existing multimodal 3D perception methods designed for terrestrial autonomous driving often underperform in off world environments due to poor feature alignment, limited multimodal synergy, and weak small object detection. This paper presents SCAFusion, a multimodal 3D object detection model tailored for lunar robotic missions. Built upon the BEVFusion framework, SCAFusion integrates a Cognitive Adapter for efficient camera backbone tuning, a Contrastive Alignment Module to enhance camera LiDAR feature consistency, a Camera Auxiliary Training Branch to strengthen visual representation, and most importantly, a Section aware Coordinate Attention mechanism explicitly designed to boost the detection performance of small, irregular targets. With negligible increase in parameters and computation, our model achieves 69.7% mAP and 72.1% NDS on the nuScenes validation set, improving the baseline by 5.0% and 2.7%, respectively. In simulated lunar environments built on Isaac Sim, SCAFusion achieves 90.93% mAP, outperforming the baseline by 11.5%, with notable gains in detecting small meteor like obstacles.", "date": "2025-12-27", "categories": []} {"id": "2512.23176", "title": "GVSynergy-Det: Synergistic Gaussian-Voxel Representations for Multi-View 3D Object Detection", "abstract": "Image-based 3D object detection aims to identify and localize objects in 3D space using only RGB images, eliminating the need for expensive depth sensors required by point cloud-based methods. Existing image-based approaches face two critical challenges: methods achieving high accuracy typically require dense 3D supervision, while those operating without such supervision struggle to extract accurate geometry from images alone. In this paper, we present GVSynergy-Det, a novel framework that enhances 3D detection through synergistic Gaussian-Voxel representation learning. Our key insight is that continuous Gaussian and discrete voxel representations capture complementary geometric information: Gaussians excel at modeling fine-grained surface details while voxels provide structured spatial context. We introduce a dual-representation architecture that: 1) adapts generalizable Gaussian Splatting to extract complementary geometric features for detection tasks, and 2) develops a cross-representation enhancement mechanism that enriches voxel features with geometric details from Gaussian fields. Unlike previous methods that either rely on time-consuming per-scene optimization or utilize Gaussian representations solely for depth regularization, our synergistic strategy directly leverages features from both representations through learnable integration, enabling more accurate object localization. Extensive experiments demonstrate that GVSynergy-Det achieves state-of-the-art results on challenging indoor benchmarks, significantly outperforming existing methods on both ScanNetV2 and ARKitScenes datasets, all without requiring any depth or dense 3D geometry supervision (e.g., point clouds or TSDF).", "date": "2025-12-29", "categories": []} {"id": "2512.23635", "title": "Rethinking the Spatio-Temporal Alignment of End-to-End 3D Perception", "abstract": "Spatio-temporal alignment is crucial for temporal modeling of end-to-end (E2E) perception in autonomous driving (AD), providing valuable structural and textural prior information. Existing methods typically rely on the attention mechanism to align objects across frames, simplifying the motion model with a unified explicit physical model (constant velocity, etc.). These approaches prefer semantic features for implicit alignment, challenging the importance of explicit motion modeling in the traditional perception paradigm. However, variations in motion states and object features across categories and frames render this alignment suboptimal. To address this, we propose HAT, a spatio-temporal alignment module that allows each object to adaptively decode the optimal alignment proposal from multiple hypotheses without direct supervision. Specifically, HAT first utilizes multiple explicit motion models to generate spatial anchors and motion-aware feature proposals for historical instances. It then performs multi-hypothesis decoding by incorporating semantic and motion cues embedded in cached object queries, ultimately providing the optimal alignment proposal for the target frame. On nuScenes, HAT consistently improves 3D temporal detectors and trackers across diverse baselines. It achieves state-of-the-art tracking results with 46.0% AMOTA on the test set when paired with the DETR3D detector. In an object-centric E2E AD method, HAT enhances perception accuracy (+1.3% mAP, +3.1% AMOTA) and reduces the collision rate by 32%. When semantics are corrupted (nuScenes-C), the enhancement of motion modeling by HAT enables more robust perception and planning in the E2E AD.", "date": "2025-12-29", "categories": []} {"id": "2512.20217", "title": "LiteFusion: Taming 3D Object Detectors from Vision-Based to Multi-Modal with Minimal Adaptation", "abstract": "3D object detection is fundamental for safe and robust intelligent transportation systems. Current multi-modal 3D object detectors often rely on complex architectures and training strategies to achieve higher detection accuracy. However, these methods heavily rely on the LiDAR sensor so that they suffer from large performance drops when LiDAR is absent, which compromises the robustness and safety of autonomous systems in practical scenarios. Moreover, existing multi-modal detectors face difficulties in deployment on diverse hardware platforms, such as NPUs and FPGAs, due to their reliance on 3D sparse convolution operators, which are primarily optimized for NVIDIA GPUs. To address these challenges, we reconsider the role of LiDAR in the camera-LiDAR fusion paradigm and introduce a novel multi-modal 3D detector, LiteFusion. Instead of treating LiDAR point clouds as an independent modality with a separate feature extraction backbone, LiteFusion utilizes LiDAR data as a complementary source of geometric information to enhance camera-based detection. This straightforward approach completely eliminates the reliance on a 3D backbone, making the method highly deployment-friendly. Specifically, LiteFusion integrates complementary features from LiDAR points into image features within a quaternion space, where the orthogonal constraints are well-preserved during network training. This helps model domain-specific relations across modalities, yielding a compact cross-modal embedding. Experiments on the nuScenes dataset show that LiteFusion improves the baseline vision-based detector by +20.4% mAP and +19.7% NDS with a minimal increase in parameters (1.1%) without using dedicated LiDAR encoders. Notably, even in the absence of LiDAR input, LiteFusion maintains strong results , highlighting its favorable robustness and effectiveness across diverse fusion paradigms and deployment scenarios.", "date": "2025-12-23", "categories": []} {"id": "2512.13982", "title": "FocalComm: Hard Instance-Aware Multi-Agent Perception", "abstract": "Multi-agent collaborative perception (CP) is a promising paradigm for improving autonomous driving safety, particularly for vulnerable road users like pedestrians, via robust 3D perception. However, existing CP approaches often optimize for vehicle detection performance metrics, underperforming on smaller, safety-critical objects such as pedestrians, where detection failures can be catastrophic. Furthermore, previous CP methods rely on full feature exchange rather than communicating only salient features that help reduce false negatives. To this end, we present FocalComm, a novel collaborative perception framework that focuses on exchanging hard-instance-oriented features among connected collaborative agents. FocalComm consists of two key novel designs: (1) a learnable progressive hard instance mining (HIM) module to extract hard instance-oriented features per agent, and (2) a query-based feature-level (intermediate) fusion technique that dynamically weights these identified features during collaboration. We show that FocalComm outperforms state-of-the-art collaborative perception methods on two challenging real-world datasets (V2X-Real and DAIR-V2X) across both vehicle-centric and infrastructure-centric collaborative setups. FocalComm also shows a strong performance gain in pedestrian detection in V2X-Real.", "date": "2025-12-16", "categories": []} {"id": "2512.17620", "title": "StereoMV2D: A Sparse Temporal Stereo-Enhanced Framework for Robust Multi-View 3D Object Detection", "abstract": "Multi-view 3D object detection is a fundamental task in autonomous driving perception, where achieving a balance between detection accuracy and computational efficiency remains crucial. Sparse query-based 3D detectors efficiently aggregate object-relevant features from multi-view images through a set of learnable queries, offering a concise and end-to-end detection paradigm. Building on this foundation, MV2D leverages 2D detection results to provide high-quality object priors for query initialization, enabling higher precision and recall. However, the inherent depth ambiguity in single-frame 2D detections still limits the accuracy of 3D query generation. To address this issue, we propose StereoMV2D, a unified framework that integrates temporal stereo modeling into the 2D detection-guided multi-view 3D detector. By exploiting cross-temporal disparities of the same object across adjacent frames, StereoMV2D enhances depth perception and refines the query priors, while performing all computations efficiently within 2D regions of interest (RoIs). Furthermore, a dynamic confidence gating mechanism adaptively evaluates the reliability of temporal stereo cues through learning statistical patterns derived from the inter-frame matching matrix together with appearance consistency, ensuring robust detection under object appearance and occlusion. Extensive experiments on the nuScenes and Argoverse 2 datasets demonstrate that StereoMV2D achieves superior detection performance without incurring significant computational overhead. Code will be available at https://github.com/Uddd821/StereoMV2D.", "date": "2025-12-19", "categories": []} {"id": "2512.13107", "title": "Diffusion-Based Restoration for Multi-Modal 3D Object Detection in Adverse Weather", "abstract": "Multi-modal 3D object detection is important for reliable perception in robotics and autonomous driving. However, its effectiveness remains limited under adverse weather conditions due to weather-induced distortions and misalignment between different data modalities. In this work, we propose DiffFusion, a novel framework designed to enhance robustness in challenging weather through diffusion-based restoration and adaptive cross-modal fusion. Our key insight is that diffusion models possess strong capabilities for denoising and generating data that can adapt to various weather conditions. Building on this, DiffFusion introduces Diffusion-IR restoring images degraded by weather effects and Point Cloud Restoration (PCR) compensating for corrupted LiDAR data using image object cues. To tackle misalignments between two modalities, we develop Bidirectional Adaptive Fusion and Alignment Module (BAFAM). It enables dynamic multi-modal fusion and bidirectional bird's-eye view (BEV) alignment to maintain consistent spatial correspondence. Extensive experiments on three public datasets show that DiffFusion achieves state-of-the-art robustness under adverse weather while preserving strong clean-data performance. Zero-shot results on the real-world DENSE dataset further validate its generalization. The implementation of our DiffFusion will be released as open-source.", "date": "2025-12-15", "categories": []} {"id": "2512.16077", "title": "Auto-Vocabulary 3D Object Detection", "abstract": "Open-vocabulary 3D object detection methods are able to localize 3D boxes of classes unseen during training. Despite the name, existing methods rely on user-specified classes both at training and inference. We propose to study Auto-Vocabulary 3D Object Detection (AV3DOD), where the classes are automatically generated for the detected objects without any user input. To this end, we introduce Semantic Score (SS) to evaluate the quality of the generated class names. We then develop a novel framework, AV3DOD, which leverages 2D vision-language models (VLMs) to generate rich semantic candidates through image captioning, pseudo 3D box generation, and feature-space semantics expansion. AV3DOD achieves the state-of-the-art (SOTA) performance on both localization (mAP) and semantic quality (SS) on the ScanNetV2 and SUNRGB-D datasets. Notably, it surpasses the SOTA, CoDA, by 3.48 overall mAP and attains a 24.5% relative improvement in SS on ScanNetV2.", "date": "2025-12-18", "categories": []} {"id": "2512.16818", "title": "DenseBEV: Transforming BEV Grid Cells into 3D Objects", "abstract": "In current research, Bird's-Eye-View (BEV)-based transformers are increasingly utilized for multi-camera 3D object detection. Traditional models often employ random queries as anchors, optimizing them successively. Recent advancements complement or replace these random queries with detections from auxiliary networks. We propose a more intuitive and efficient approach by using BEV feature cells directly as anchors. This end-to-end approach leverages the dense grid of BEV queries, considering each cell as a potential object for the final detection task. As a result, we introduce a novel two-stage anchor generation method specifically designed for multi-camera 3D object detection. To address the scaling issues of attention with a large number of queries, we apply BEV-based Non-Maximum Suppression, allowing gradients to flow only through non-suppressed objects. This ensures efficient training without the need for post-processing. By using BEV features from encoders such as BEVFormer directly as object queries, temporal BEV information is inherently embedded. Building on the temporal BEV information already embedded in our object queries, we introduce a hybrid temporal modeling approach by integrating prior detections to further enhance detection performance. Evaluating our method on the nuScenes dataset shows consistent and significant improvements in NDS and mAP over the baseline, even with sparser BEV grids and therefore fewer initial anchors. It is particularly effective for small objects, enhancing pedestrian detection with a 3.8% mAP increase on nuScenes and an 8% increase in LET-mAP on Waymo. Applying our method, named DenseBEV, to the challenging Waymo Open dataset yields state-of-the-art performance, achieving a LET-mAP of 60.7%, surpassing the previous best by 5.4%. Code is available at https://github.com/mdaehl/DenseBEV.", "date": "2025-12-18", "categories": []} {"id": "2512.15581", "title": "IMKD: Intensity-Aware Multi-Level Knowledge Distillation for Camera-Radar Fusion", "abstract": "High-performance Radar-Camera 3D object detection can be achieved by leveraging knowledge distillation without using LiDAR at inference time. However, existing distillation methods typically transfer modality-specific features directly to each sensor, which can distort their unique characteristics and degrade their individual strengths. To address this, we introduce IMKD, a radar-camera fusion framework based on multi-level knowledge distillation that preserves each sensor's intrinsic characteristics while amplifying their complementary strengths. IMKD applies a three-stage, intensity-aware distillation strategy to enrich the fused representation across the architecture: (1) LiDAR-to-Radar intensity-aware feature distillation to enhance radar representations with fine-grained structural cues, (2) LiDAR-to-Fused feature intensity-guided distillation to selectively highlight useful geometry and depth information at the fusion level, fostering complementarity between the modalities rather than forcing them to align, and (3) Camera-Radar intensity-guided fusion mechanism that facilitates effective feature alignment and calibration. Extensive experiments on the nuScenes benchmark show that IMKD reaches 67.0% NDS and 61.0% mAP, outperforming all prior distillation-based radar-camera fusion methods. Our code and models are available at https://github.com/dfki-av/IMKD/.", "date": "2025-12-17", "categories": []} {"id": "2510.19981", "title": "FutrTrack: A Camera-LiDAR Fusion Transformer for 3D Multiple Object Tracking", "abstract": "We propose FutrTrack, a modular camera-LiDAR multi-object tracking framework that builds on existing 3D detectors by introducing a transformer-based smoother and a fusion-driven tracker. Inspired by query-based tracking frameworks, FutrTrack employs a multimodal two-stage transformer refinement and tracking pipeline. Our fusion tracker integrates bounding boxes with multimodal bird's-eye-view (BEV) fusion features from multiple cameras and LiDAR without the need for an explicit motion model. The tracker assigns and propagates identities across frames, leveraging both geometric and semantic cues for robust re-identification under occlusion and viewpoint changes. Prior to tracking, we refine sequences of bounding boxes with a temporal smoother over a moving window to refine trajectories, reduce jitter, and improve spatial consistency. Evaluated on nuScenes and KITTI, FutrTrack demonstrates that query-based transformer tracking methods benefit significantly from multimodal sensor features compared with previous single-sensor approaches. With an aMOTA of 74.7 on the nuScenes test set, FutrTrack achieves strong performance on 3D MOT benchmarks, reducing identity switches while maintaining competitive accuracy. Our approach provides an efficient framework for improving transformer-based trackers to compete with other neural-network-based methods even with limited data and without pretraining.", "date": "2025-10-22", "categories": []} {"id": "2512.11926", "title": "TransBridge: Boost 3D Object Detection by Scene-Level Completion with Transformer Decoder", "abstract": "3D object detection is essential in autonomous driving, providing vital information about moving objects and obstacles. Detecting objects in distant regions with only a few LiDAR points is still a challenge, and numerous strategies have been developed to address point cloud sparsity through densification.This paper presents a joint completion and detection framework that improves the detection feature in sparse areas while maintaining costs unchanged. Specifically, we propose TransBridge, a novel transformer-based up-sampling block that fuses the features from the detection and completion networks.The detection network can benefit from acquiring implicit completion features derived from the completion network. Additionally, we design the Dynamic-Static Reconstruction (DSRecon) module to produce dense LiDAR data for the completion network, meeting the requirement for dense point cloud ground truth.Furthermore, we employ the transformer mechanism to establish connections between channels and spatial relations, resulting in a high-resolution feature map used for completion purposes.Extensive experiments on the nuScenes and Waymo datasets demonstrate the effectiveness of the proposed framework.The results show that our framework consistently improves end-to-end 3D object detection, with the mean average precision (mAP) ranging from 0.7 to 1.5 across multiple methods, indicating its generalization ability. For the two-stage detection framework, it also boosts the mAP up to 5.78 points.", "date": "2025-12-12", "categories": []} {"id": "2512.08223", "title": "SOP^2: Transfer Learning with Scene-Oriented Prompt Pool on 3D Object Detection", "abstract": "With the rise of Large Language Models (LLMs) such as GPT-3, these models exhibit strong generalization capabilities. Through transfer learning techniques such as fine-tuning and prompt tuning, they can be adapted to various downstream tasks with minimal parameter adjustments. This approach is particularly common in the field of Natural Language Processing (NLP). This paper aims to explore the effectiveness of common prompt tuning methods in 3D object detection. We investigate whether a model trained on the large-scale Waymo dataset can serve as a foundation model and adapt to other scenarios within the 3D object detection field. This paper sequentially examines the impact of prompt tokens and prompt generators, and further proposes a Scene-Oriented Prompt Pool (\\textbf{SOP$^2$}). We demonstrate the effectiveness of prompt pools in 3D object detection, with the goal of inspiring future researchers to delve deeper into the potential of prompts in the 3D field.", "date": "2025-12-09", "categories": []} {"id": "2512.08247", "title": "Distilling Future Temporal Knowledge with Masked Feature Reconstruction for 3D Object Detection", "abstract": "Camera-based temporal 3D object detection has shown impressive results in autonomous driving, with offline models improving accuracy by using future frames. Knowledge distillation (KD) can be an appealing framework for transferring rich information from offline models to online models. However, existing KD methods overlook future frames, as they mainly focus on spatial feature distillation under strict frame alignment or on temporal relational distillation, thereby making it challenging for online models to effectively learn future knowledge. To this end, we propose a sparse query-based approach, Future Temporal Knowledge Distillation (FTKD), which effectively transfers future frame knowledge from an offline teacher model to an online student model. Specifically, we present a future-aware feature reconstruction strategy to encourage the student model to capture future features without strict frame alignment. In addition, we further introduce future-guided logit distillation to leverage the teacher's stable foreground and background context. FTKD is applied to two high-performing 3D object detection baselines, achieving up to 1.3 mAP and 1.3 NDS gains on the nuScenes dataset, as well as the most accurate velocity estimation, without increasing inference cost.", "date": "2025-12-09", "categories": []} {"id": "2404.01064", "title": "Roadside Monocular 3D Detection Prompted by 2D Detection", "abstract": "Roadside monocular 3D detection requires detecting objects of predefined classes in an RGB frame and predicting their 3D attributes, such as bird's-eye-view (BEV) locations. It has broad applications in traffic control, vehicle-vehicle communication, and vehicle-infrastructure cooperative perception. To address this task, we introduce Promptable 3D Detector (Pro3D), a novel detector design that leverages 2D detections as prompts. We build our Pro3D upon two key insights. First, compared to a typical 3D detector, a 2D detector is ``easier'' to train due to fewer loss terms and performs significantly better at localizing objects w.r.t 2D metrics. Second, once 2D detections precisely locate objects in the image, a 3D detector can focus on lifting these detections into 3D BEV, especially when fixed camera pose or scene geometry provide an informative prior. To encode and incorporate 2D detections, we explore three methods: (a) concatenating features from both 2D and 3D detectors, (b) attentively fusing 2D and 3D detector features, and (c) encoding properties of predicted 2D bounding boxes \\{$x$, $y$, width, height, label\\} and attentively fusing them with the 3D detector feature. Interestingly, the third method significantly outperforms the others, underscoring the effectiveness of 2D detections as prompts that offer precise object targets and allow the 3D detector to focus on lifting them into 3D. Pro3D is adaptable for use with a wide range of 2D and 3D detectors with minimal modifications. Comprehensive experiments demonstrate that our Pro3D significantly enhances existing methods, achieving state-of-the-art results on two contemporary benchmarks.", "date": "2024-04-01", "categories": []} {"id": "2409.14985", "title": "Image-Guided Semantic Pseudo-LiDAR Point Generation for 3D Object Detection", "abstract": "In autonomous driving scenarios, accurate perception is becoming an even more critical task for safe navigation. While LiDAR provides precise spatial data, its inherent sparsity makes it difficult to detect small or distant objects. Existing methods try to address this by generating additional points within a Region of Interest (RoI), but relying on LiDAR alone often leads to false positives and a failure to recover meaningful structures. To address these limitations, we propose Image-Guided Semantic Pseudo-LiDAR Point Generation model, called ImagePG, a novel framework that leverages rich RGB image features to generate dense and semantically meaningful 3D points. Our framework includes an Image-Guided RoI Points Generation (IG-RPG) module, which creates pseudo-points guided by image features, and an Image-Aware Occupancy Prediction Network (I-OPN), which provides spatial priors to guide point placement. A multi-stage refinement (MR) module further enhances point quality and detection robustness. To the best of our knowledge, ImagePG is the first method to directly leverage image features for point generation. Extensive experiments on the KITTI and Waymo datasets demonstrate that ImagePG significantly improves the detection of small and distant objects like pedestrians and cyclists, reducing false positives by nearly 50%. On the KITTI benchmark, our framework improves mAP by +1.38%p (car), +7.91%p (pedestrian), and +5.21%p (cyclist) on the test set over the baseline, achieving state-of-the-art cyclist performance on the KITTI leaderboard. The code is available at: https://github.com/MS-LIMA/ImagePG", "date": "2024-09-23", "categories": []} {"id": "2507.17665", "title": "Perspective-Invariant 3D Object Detection", "abstract": "With the rise of robotics, LiDAR-based 3D object detection has garnered significant attention in both academia and industry. However, existing datasets and methods predominantly focus on vehicle-mounted platforms, leaving other autonomous platforms underexplored. To bridge this gap, we introduce Pi3DET, the first benchmark featuring LiDAR data and 3D bounding box annotations collected from multiple platforms: vehicle, quadruped, and drone, thereby facilitating research in 3D object detection for non-vehicle platforms as well as cross-platform 3D detection. Based on Pi3DET, we propose a novel cross-platform adaptation framework that transfers knowledge from the well-studied vehicle platform to other platforms. This framework achieves perspective-invariant 3D detection through robust alignment at both geometric and feature levels. Additionally, we establish a benchmark to evaluate the resilience and robustness of current 3D detectors in cross-platform scenarios, providing valuable insights for developing adaptive 3D perception systems. Extensive experiments validate the effectiveness of our approach on challenging cross-platform tasks, demonstrating substantial gains over existing adaptation methods. We hope this work paves the way for generalizable and unified 3D perception systems across diverse and complex environments. Our Pi3DET dataset, cross-platform benchmark suite, and annotation toolkit have been made publicly available.", "date": "2025-07-23", "categories": []} {"id": "2512.05698", "title": "OWL: Unsupervised 3D Object Detection by Occupancy Guided Warm-up and Large Model Priors Reasoning", "abstract": "Unsupervised 3D object detection leverages heuristic algorithms to discover potential objects, offering a promising route to reduce annotation costs in autonomous driving. Existing approaches mainly generate pseudo labels and refine them through self-training iterations. However, these pseudo-labels are often incorrect at the beginning of training, resulting in misleading the optimization process. Moreover, effectively filtering and refining them remains a critical challenge. In this paper, we propose OWL for unsupervised 3D object detection by occupancy guided warm-up and large-model priors reasoning. OWL first employs an Occupancy Guided Warm-up (OGW) strategy to initialize the backbone weight with spatial perception capabilities, mitigating the interference of incorrect pseudo-labels on network convergence. Furthermore, OWL introduces an Instance-Cued Reasoning (ICR) module that leverages the prior knowledge of large models to assess pseudo-label quality, enabling precise filtering and refinement. Finally, we design a Weight-adapted Self-training (WAS) strategy to dynamically re-weight pseudo-labels, improving the performance through self-training. Extensive experiments on Waymo Open Dataset (WOD) and KITTI demonstrate that OWL outperforms state-of-the-art unsupervised methods by over 15.0% mAP, revealing the effectiveness of our method.", "date": "2025-12-05", "categories": []} {"id": "2309.07104", "title": "Polygon Intersection-over-Union Loss for Viewpoint-Agnostic Monocular 3D Vehicle Detection", "abstract": "Monocular 3D object detection is a challenging task because depth information is difficult to obtain from 2D images. A subset of viewpoint-agnostic monocular 3D detection methods also do not explicitly leverage scene homography or geometry during training, meaning that a model trained thusly can detect objects in images from arbitrary viewpoints. Such works predict the projections of the 3D bounding boxes on the image plane to estimate the location of the 3D boxes, but these projections are not rectangular so the calculation of IoU between these projected polygons is not straightforward. This work proposes an efficient, fully differentiable algorithm for the calculation of IoU between two convex polygons, which can be utilized to compute the IoU between two 3D bounding box footprints viewed from an arbitrary angle. We test the performance of the proposed polygon IoU loss (PIoU loss) on three state-of-the-art viewpoint-agnostic 3D detection models. Experiments demonstrate that the proposed PIoU loss converges faster than L1 loss and that in 3D detection models, a combination of PIoU loss and L1 loss gives better results than L1 loss alone (+1.64% AP70 for MonoCon on cars, +0.18% AP70 for RTM3D on cars, and +0.83%/+2.46% AP50/AP25 for MonoRCNN on cyclists).", "date": "2023-09-13", "categories": []} {"id": "2405.16848", "title": "A re-calibration method for object detection with multi-modal alignment bias in autonomous driving", "abstract": "Multi-modal object detection in autonomous driving has achieved great breakthroughs due to the usage of fusing complementary information from different sensors. The calibration in fusion between sensors such as LiDAR and camera was always supposed to be precise in previous work. However, in reality, calibration matrices are fixed when the vehicles leave the factory, but mechanical vibration, road bumps, and data lags may cause calibration bias. As there is relatively limited research on the impact of calibration on fusion detection performance, multi-sensor detection methods with flexible calibration dependency have remained a key objective. In this paper, we systematically evaluate the sensitivity of the SOTA EPNet++ detection framework and prove that even slight bias on calibration can reduce the performance seriously. To address this vulnerability, we propose a re-calibration model to re-calibrate the misalignment in detection tasks. This model integrates LiDAR point cloud, camera image, and initial calibration matrix as inputs, generating re-calibrated bias through semantic segmentation guidance and a tailored loss function design. The re-calibration model can operate with existing detection algorithms, enhancing both robustness against calibration bias and overall object detection performance. Our approach establishes a foundational methodology for maintaining reliability in multi-modal perception systems under real-world calibration uncertainties.", "date": "2024-05-27", "categories": []} {"id": "2412.07768", "title": "Test-time Correction: An Online 3D Detection System via Visual Prompting", "abstract": "This paper introduces Test-time Correction (TTC), an online 3D detection system designed to rectify test-time errors using various auxiliary feedback, aiming to enhance the safety of deployed autonomous driving systems. Unlike conventional offline 3D detectors that remain fixed during inference, TTC enables immediate online error correction without retraining, allowing autonomous vehicles to adapt to new scenarios and reduce deployment risks. To achieve this, we equip existing 3D detectors with an Online Adapter (OA) module -- a prompt-driven query generator for real-time correction. At the core of OA module are visual prompts: image-based descriptions of objects of interest derived from auxiliary feedback such as mismatches with 2D detections, road descriptions, or user clicks. These visual prompts, collected from risky objects during inference, are maintained in a visual prompt buffer to enable continuous correction in future frames. By leveraging this mechanism, TTC consistently detects risky objects, achieving reliable, adaptive, and versatile driving autonomy. Extensive experiments show that TTC significantly improves instant error rectification over frozen 3D detectors, even under limited labels, zero-shot settings, and adverse conditions. We hope this work inspires future research on post-deployment online rectification systems for autonomous driving.", "date": "2024-12-10", "categories": []} {"id": "2504.07958", "title": "Detect Anything 3D in the Wild", "abstract": "Despite the success of deep learning in close-set 3D object detection, existing approaches struggle with zero-shot generalization to novel objects and camera configurations. We introduce DetAny3D, a promptable 3D detection foundation model capable of detecting any novel object under arbitrary camera configurations using only monocular inputs. Training a foundation model for 3D detection is fundamentally constrained by the limited availability of annotated 3D data, which motivates DetAny3D to leverage the rich prior knowledge embedded in extensively pre-trained 2D foundation models to compensate for this scarcity. To effectively transfer 2D knowledge to 3D, DetAny3D incorporates two core modules: the 2D Aggregator, which aligns features from different 2D foundation models, and the 3D Interpreter with Zero-Embedding Mapping, which stabilizes early training in 2D-to-3D knowledge transfer. Experimental results validate the strong generalization of our DetAny3D, which not only achieves state-of-the-art performance on unseen categories and novel camera configurations, but also surpasses most competitors on in-domain data. DetAny3D sheds light on the potential of the 3D foundation model for diverse applications in real-world scenarios, e.g., rare object detection in autonomous driving, and demonstrates promise for further exploration of 3D-centric tasks in open-world settings. More visualization results can be found at our code repository.", "date": "2025-04-10", "categories": []} {"id": "2512.02972", "title": "BEVDilation: LiDAR-Centric Multi-Modal Fusion for 3D Object Detection", "abstract": "Integrating LiDAR and camera information in the bird's eye view (BEV) representation has demonstrated its effectiveness in 3D object detection. However, because of the fundamental disparity in geometric accuracy between these sensors, indiscriminate fusion in previous methods often leads to degraded performance. In this paper, we propose BEVDilation, a novel LiDAR-centric framework that prioritizes LiDAR information in the fusion. By formulating image BEV features as implicit guidance rather than naive concatenation, our strategy effectively alleviates the spatial misalignment caused by image depth estimation errors. Furthermore, the image guidance can effectively help the LiDAR-centric paradigm to address the sparsity and semantic limitations of point clouds. Specifically, we propose a Sparse Voxel Dilation Block that mitigates the inherent point sparsity by densifying foreground voxels through image priors. Moreover, we introduce a Semantic-Guided BEV Dilation Block to enhance the LiDAR feature diffusion processing with image semantic guidance and long-range context capture. On the challenging nuScenes benchmark, BEVDilation achieves better performance than state-of-the-art methods while maintaining competitive computational efficiency. Importantly, our LiDAR-centric strategy demonstrates greater robustness to depth noise compared to naive fusion. The source code is available at https://github.com/gwenzhang/BEVDilation.", "date": "2025-12-02", "categories": []} {"id": "2512.00060", "title": "PEFT-DML: Parameter-Efficient Fine-Tuning Deep Metric Learning for Robust Multi-Modal 3D Object Detection in Autonomous Driving", "abstract": "This study introduces PEFT-DML, a parameter-efficient deep metric learning framework for robust multi-modal 3D object detection in autonomous driving. Unlike conventional models that assume fixed sensor availability, PEFT-DML maps diverse modalities (LiDAR, radar, camera, IMU, GNSS) into a shared latent space, enabling reliable detection even under sensor dropout or unseen modality class combinations. By integrating Low-Rank Adaptation (LoRA) and adapter layers, PEFT-DML achieves significant training efficiency while enhancing robustness to fast motion, weather variability, and domain shifts. Experiments on benchmarks nuScenes demonstrate superior accuracy.", "date": "2025-11-23", "categories": []} {"id": "2512.01178", "title": "VSRD++: Autolabeling for 3D Object Detection via Instance-Aware Volumetric Silhouette Rendering", "abstract": "Monocular 3D object detection is a fundamental yet challenging task in 3D scene understanding. Existing approaches heavily depend on supervised learning with extensive 3D annotations, which are often acquired from LiDAR point clouds through labor-intensive labeling processes. To tackle this problem, we propose VSRD++, a novel weakly supervised framework for monocular 3D object detection that eliminates the reliance on 3D annotations and leverages neural-field-based volumetric rendering with weak 2D supervision. VSRD++ consists of a two-stage pipeline: multi-view 3D autolabeling and subsequent monocular 3D detector training. In the multi-view autolabeling stage, object surfaces are represented as signed distance fields (SDFs) and rendered as instance masks via the proposed instance-aware volumetric silhouette rendering. To optimize 3D bounding boxes, we decompose each instance's SDF into a cuboid SDF and a residual distance field (RDF) that captures deviations from the cuboid. To address the geometry inconsistency commonly observed in volume rendering methods applied to dynamic objects, we model the dynamic objects by including velocity into bounding box attributes as well as assigning confidence to each pseudo-label. Moreover, we also employ a 3D attribute initialization module to initialize the dynamic bounding box parameters. In the monocular 3D object detection phase, the optimized 3D bounding boxes serve as pseudo labels for training monocular 3D object detectors. Extensive experiments on the KITTI-360 dataset demonstrate that VSRD++ significantly outperforms existing weakly supervised approaches for monocular 3D object detection on both static and dynamic scenes. Code is available at https://github.com/Magicboomliu/VSRD_plus_plus", "date": "2025-12-01", "categories": []} {"id": "2512.01352", "title": "OpenBox: Annotate Any Bounding Boxes in 3D", "abstract": "Unsupervised and open-vocabulary 3D object detection has recently gained attention, particularly in autonomous driving, where reducing annotation costs and recognizing unseen objects are critical for both safety and scalability. However, most existing approaches uniformly annotate 3D bounding boxes, ignore objects' physical states, and require multiple self-training iterations for annotation refinement, resulting in suboptimal quality and substantial computational overhead. To address these challenges, we propose OpenBox, a two-stage automatic annotation pipeline that leverages a 2D vision foundation model. In the first stage, OpenBox associates instance-level cues from 2D images processed by a vision foundation model with the corresponding 3D point clouds via cross-modal instance alignment. In the second stage, it categorizes instances by rigidity and motion state, then generates adaptive bounding boxes with class-specific size statistics. As a result, OpenBox produces high-quality 3D bounding box annotations without requiring self-training. Experiments on the Waymo Open Dataset, the Lyft Level 5 Perception dataset, and the nuScenes dataset demonstrate improved accuracy and efficiency over baselines.", "date": "2025-12-01", "categories": []} {"id": "2511.22404", "title": "UAV-MM3D: A Large-Scale Synthetic Benchmark for 3D Perception of Unmanned Aerial Vehicles with Multi-Modal Data", "abstract": "Accurate perception of UAVs in complex low-altitude environments is critical for airspace security and related intelligent systems. Developing reliable solutions requires large-scale, accurately annotated, and multimodal data. However, real-world UAV data collection faces inherent constraints due to airspace regulations, privacy concerns, and environmental variability, while manual annotation of 3D poses and cross-modal correspondences is time-consuming and costly. To overcome these challenges, we introduce UAV-MM3D, a high-fidelity multimodal synthetic dataset for low-altitude UAV perception and motion understanding. It comprises 400K synchronized frames across diverse scenes (urban areas, suburbs, forests, coastal regions) and weather conditions (clear, cloudy, rainy, foggy), featuring multiple UAV models (micro, small, medium-sized) and five modalities - RGB, IR, LiDAR, Radar, and DVS (Dynamic Vision Sensor). Each frame provides 2D/3D bounding boxes, 6-DoF poses, and instance-level annotations, enabling core tasks related to UAVs such as 3D detection, pose estimation, target tracking, and short-term trajectory forecasting. We further propose LGFusionNet, a LiDAR-guided multimodal fusion baseline, and a dedicated UAV trajectory prediction baseline to facilitate benchmarking. With its controllable simulation environment, comprehensive scenario coverage, and rich annotations, UAV3D offers a public benchmark for advancing 3D perception of UAVs.", "date": "2025-11-27", "categories": []} {"id": "2411.16833", "title": "Open Vocabulary Monocular 3D Object Detection", "abstract": "We propose and study open-vocabulary monocular 3D detection, a novel task that aims to detect objects of any categores in metric 3D space from a single RGB image. Existing 3D object detectors either rely on costly sensors such as LiDAR or multi-view setups, or remain confined to closed vocabularies settings with limited categories, restricting their applicability. We identify two key challenges in this new setting. First, the scarcity of 3D bounding box annotations limits the ability to train generalizable models. To reduce dependence on 3D supervision, we propose a framework that effectively integrates pretrained 2D and 3D vision foundation models. Second, missing labels and semantic ambiguities (\\eg, table vs. desk) in existing datasets hinder reliable evaluation. To address this, we design a novel metric that captures model performance while mitigating annotation issues. Our approach achieves state-of-the-art results in zero-shot 3D detection of novel categories as well as in-domain detection on seen classes. We hope our method provides a strong baseline and our evaluation protocol establishes a reliable benchmark for future research.", "date": "2024-11-25", "categories": []} {"id": "2511.20253", "title": "Zoo3D: Zero-Shot 3D Object Detection at Scene Level", "abstract": "3D object detection is fundamental for spatial understanding. Real-world environments demand models capable of recognizing diverse, previously unseen objects, which remains a major limitation of closed-set methods. Existing open-vocabulary 3D detectors relax annotation requirements but still depend on training scenes, either as point clouds or images. We take this a step further by introducing Zoo3D, the first training-free 3D object detection framework. Our method constructs 3D bounding boxes via graph clustering of 2D instance masks, then assigns semantic labels using a novel open-vocabulary module with best-view selection and view-consensus mask generation. Zoo3D operates in two modes: the zero-shot Zoo3D$_0$, which requires no training at all, and the self-supervised Zoo3D$_1$, which refines 3D box prediction by training a class-agnostic detector on Zoo3D$_0$-generated pseudo labels. Furthermore, we extend Zoo3D beyond point clouds to work directly with posed and even unposed images. Across ScanNet200 and ARKitScenes benchmarks, both Zoo3D$_0$ and Zoo3D$_1$ achieve state-of-the-art results in open-vocabulary 3D object detection. Remarkably, our zero-shot Zoo3D$_0$ outperforms all existing self-supervised methods, hence demonstrating the power and adaptability of training-free, off-the-shelf approaches for real-world 3D understanding. Code is available at https://github.com/col14m/zoo3d .", "date": "2025-11-25", "categories": []} {"id": "2409.08840", "title": "Directed-CP: Directed Collaborative Perception for Connected and Autonomous Vehicles via Proactive Attention", "abstract": "Collaborative perception (CP) leverages visual data from connected and autonomous vehicles (CAV) to enhance an ego vehicle's field of view (FoV). Despite recent progress, current CP methods expand the ego vehicle's 360-degree perceptual range almost equally, which faces two key challenges. Firstly, in areas with uneven traffic distribution, focusing on directions with little traffic offers limited benefits. Secondly, under limited communication budgets, allocating excessive bandwidth to less critical directions lowers the perception accuracy in more vital areas. To address these issues, we propose Direct-CP, a proactive and direction-aware CP system aiming at improving CP in specific directions. Our key idea is to enable an ego vehicle to proactively signal its interested directions and readjust its attention to enhance local directional CP performance. To achieve this, we first propose an RSU-aided direction masking mechanism that assists an ego vehicle in identifying vital directions. Additionally, we design a direction-aware selective attention module to wisely aggregate pertinent features based on ego vehicle's directional priorities, communication budget, and the positional data of CAVs. Moreover, we introduce a direction-weighted detection loss (DWLoss) to capture the divergence between directional CP outcomes and the ground truth, facilitating effective model training. Extensive experiments on the V2X-Sim 2.0 dataset demonstrate that our approach achieves 19.8\\% higher local perception accuracy in interested directions and 2.5\\% higher overall perception accuracy than the state-of-the-art methods in collaborative 3D object detection tasks. Codes are available at https://github.com/yihangtao/Directed-CP.git.", "date": "2024-09-13", "categories": []} {"id": "2511.17619", "title": "Rethinking the Encoding and Annotating of 3D Bounding Box: Corner-Aware 3D Object Detection from Point Clouds", "abstract": "Center-aligned regression remains dominant in LiDAR-based 3D object detection, yet it suffers from fundamental instability: object centers often fall in sparse or empty regions of the bird's-eye-view (BEV) due to the front-surface-biased nature of LiDAR point clouds, leading to noisy and inaccurate bounding box predictions. To circumvent this limitation, we revisit bounding box representation and propose corner-aligned regression, which shifts the prediction target from unstable centers to geometrically informative corners that reside in dense, observable regions. Leveraging the inherent geometric constraints among corners and image 2D boxes, partial parameters of 3D bounding boxes can be recovered from corner annotations, enabling a weakly supervised paradigm without requiring complete 3D labels. We design a simple yet effective corner-aware detection head that can be plugged into existing detectors. Experiments on KITTI show our method improves performance by 3.5% AP over center-based baseline, and achieves 83% of fully supervised accuracy using only BEV corner clicks, demonstrating the effectiveness of our corner-aware regression strategy.", "date": "2025-11-18", "categories": []} {"id": "2511.18695", "title": "Exploring Surround-View Fisheye Camera 3D Object Detection", "abstract": "In this work, we explore the technical feasibility of implementing end-to-end 3D object detection (3DOD) with surround-view fisheye camera system. Specifically, we first investigate the performance drop incurred when transferring classic pinhole-based 3D object detectors to fisheye imagery. To mitigate this, we then develop two methods that incorporate the unique geometry of fisheye images into mainstream detection frameworks: one based on the bird's-eye-view (BEV) paradigm, named FisheyeBEVDet, and the other on the query-based paradigm, named FisheyePETR. Both methods adopt spherical spatial representations to effectively capture fisheye geometry. In light of the lack of dedicated evaluation benchmarks, we release Fisheye3DOD, a new open dataset synthesized using CARLA and featuring both standard pinhole and fisheye camera arrays. Experiments on Fisheye3DOD show that our fisheye-compatible modeling improves accuracy by up to 6.2% over baseline methods.", "date": "2025-11-24", "categories": []} {"id": "2511.18788", "title": "StereoDETR: Stereo-based Transformer for 3D Object Detection", "abstract": "Compared to monocular 3D object detection, stereo-based 3D methods offer significantly higher accuracy but still suffer from high computational overhead and latency. The state-of-the-art stereo 3D detection method achieves twice the accuracy of monocular approaches, yet its inference speed is only half as fast. In this paper, we propose StereoDETR, an efficient stereo 3D object detection framework based on DETR. StereoDETR consists of two branches: a monocular DETR branch and a stereo branch. The DETR branch is built upon 2D DETR with additional channels for predicting object scale, orientation, and sampling points. The stereo branch leverages low-cost multi-scale disparity features to predict object-level depth maps. These two branches are coupled solely through a differentiable depth sampling strategy. To handle occlusion, we introduce a constrained supervision strategy for sampling points without requiring extra annotations. StereoDETR achieves real-time inference and is the first stereo-based method to surpass monocular approaches in speed. It also achieves competitive accuracy on the public KITTI benchmark, setting new state-of-the-art results on pedestrian and cyclist subsets. The code is available at https://github.com/shiyi-mu/StereoDETR-OPEN.", "date": "2025-11-24", "categories": []} {"id": "2511.18814", "title": "DetAny4D: Detect Anything 4D Temporally in a Streaming RGB Video", "abstract": "Reliable 4D object detection, which refers to 3D object detection in streaming video, is crucial for perceiving and understanding the real world. Existing open-set 4D object detection methods typically make predictions on a frame-by-frame basis without modeling temporal consistency, or rely on complex multi-stage pipelines that are prone to error propagation across cascaded stages. Progress in this area has been hindered by the lack of large-scale datasets that capture continuous reliable 3D bounding box (b-box) annotations. To overcome these challenges, we first introduce DA4D, a large-scale 4D detection dataset containing over 280k sequences with high-quality b-box annotations collected under diverse conditions. Building on DA4D, we propose DetAny4D, an open-set end-to-end framework that predicts 3D b-boxes directly from sequential inputs. DetAny4D fuses multi-modal features from pre-trained foundational models and designs a geometry-aware spatiotemporal decoder to effectively capture both spatial and temporal dynamics. Furthermore, it adopts a multi-task learning architecture coupled with a dedicated training strategy to maintain global consistency across sequences of varying lengths. Extensive experiments show that DetAny4D achieves competitive detection accuracy and significantly improves temporal stability, effectively addressing long-standing issues of jitter and inconsistency in 4D object detection. Data and code will be released upon acceptance.", "date": "2025-11-24", "categories": []} {"id": "2511.19057", "title": "LAA3D: A Benchmark of Detecting and Tracking Low-Altitude Aircraft in 3D Space", "abstract": "Perception of Low-Altitude Aircraft (LAA) in 3D space enables precise 3D object localization and behavior understanding. However, datasets tailored for 3D LAA perception remain scarce. To address this gap, we present LAA3D, a large-scale dataset designed to advance 3D detection and tracking of low-altitude aerial vehicles. LAA3D contains 15,000 real images and 600,000 synthetic frames, captured across diverse scenarios, including urban and suburban environments. It covers multiple aerial object categories, including electric Vertical Take-Off and Landing (eVTOL) aircraft, Micro Aerial Vehicles (MAVs), and Helicopters. Each instance is annotated with 3D bounding box, class label, and instance identity, supporting tasks such as 3D object detection, 3D multi-object tracking (MOT), and 6-DoF pose estimation. Besides, we establish the LAA3D Benchmark, integrating multiple tasks and methods with unified evaluation protocols for comparison. Furthermore, we propose MonoLAA, a monocular 3D detection baseline, achieving robust 3D localization from zoom cameras with varying focal lengths. Models pretrained on synthetic images transfer effectively to real-world data with fine-tuning, demonstrating strong sim-to-real generalization. Our LAA3D provides a comprehensive foundation for future research in low-altitude 3D object perception.", "date": "2025-11-24", "categories": []} {"id": "2511.19301", "title": "IDEAL-M3D: Instance Diversity-Enriched Active Learning for Monocular 3D Detection", "abstract": "Monocular 3D detection relies on just a single camera and is therefore easy to deploy. Yet, achieving reliable 3D understanding from monocular images requires substantial annotation, and 3D labels are especially costly. To maximize performance under constrained labeling budgets, it is essential to prioritize annotating samples expected to deliver the largest performance gains. This prioritization is the focus of active learning. Curiously, we observed two significant limitations in active learning algorithms for 3D monocular object detection. First, previous approaches select entire images, which is inefficient, as non-informative instances contained in the same image also need to be labeled. Secondly, existing methods rely on uncertainty-based selection, which in monocular 3D object detection creates a bias toward depth ambiguity. Consequently, distant objects are selected, while nearby objects are overlooked.\n To address these limitations, we propose IDEAL-M3D, the first instance-level pipeline for monocular 3D detection. For the first time, we demonstrate that an explicitly diverse, fast-to-train ensemble improves diversity-driven active learning for monocular 3D. We induce diversity with heterogeneous backbones and task-agnostic features, loss weight perturbation, and time-dependent bagging. IDEAL-M3D shows superior performance and significant resource savings: with just 60% of the annotations, we achieve similar or better AP3D on KITTI validation and test set results compared to training the same detector on the whole dataset.", "date": "2025-11-24", "categories": []} {"id": "2511.16140", "title": "Real-Time 3D Object Detection with Inference-Aligned Learning", "abstract": "Real-time 3D object detection from point clouds is essential for dynamic scene understanding in applications such as augmented reality, robotics and navigation. We introduce a novel Spatial-prioritized and Rank-aware 3D object detection (SR3D) framework for indoor point clouds, to bridge the gap between how detectors are trained and how they are evaluated. This gap stems from the lack of spatial reliability and ranking awareness during training, which conflicts with the ranking-based prediction selection used as inference. Such a training-inference gap hampers the model's ability to learn representations aligned with inference-time behavior. To address the limitation, SR3D consists of two components tailored to the spatial nature of point clouds during training: a novel spatial-prioritized optimal transport assignment that dynamically emphasizes well-located and spatially reliable samples, and a rank-aware adaptive self-distillation scheme that adaptively injects ranking perception via a self-distillation paradigm. Extensive experiments on ScanNet V2 and SUN RGB-D show that SR3D effectively bridges the training-inference gap and significantly outperforms prior methods in accuracy while maintaining real-time speed.", "date": "2025-11-20", "categories": []} {"id": "2502.15488", "title": "FQ-PETR: Fully Quantized Position Embedding Transformation for Multi-View 3D Object Detection", "abstract": "Camera-based multi-view 3D detection is crucial for autonomous driving. PETR and its variants (PETRs) excel in benchmarks but face deployment challenges due to high computational cost and memory footprint. Quantization is an effective technique for compressing deep neural networks by reducing the bit width of weights and activations. However, directly applying existing quantization methods to PETRs leads to severe accuracy degradation. This issue primarily arises from two key challenges: (1) significant magnitude disparity between multi-modal features-specifically, image features and camera-ray positional embeddings (PE), and (2) the inefficiency and approximation error of quantizing non-linear operators, which commonly rely on hardware-unfriendly computations. In this paper, we propose FQ-PETR, a fully quantized framework for PETRs, featuring three key innovations: (1) Quantization-Friendly LiDAR-ray Position Embedding (QFPE): Replacing multi-point sampling with LiDAR-prior-guided single-point sampling and anchor-based embedding eliminates problematic non-linearities (e.g., inverse-sigmoid) and aligns PE scale with image features, preserving accuracy. (2) Dual-Lookup Table (DULUT): This algorithm approximates complex non-linear functions using two cascaded linear LUTs, achieving high fidelity with minimal entries and no specialized hardware. (3) Quantization After Numerical Stabilization (QANS): Performing quantization after softmax numerical stabilization mitigates attention distortion from large inputs. On PETRs (e.g. PETR, StreamPETR, PETRv2, MV2d), FQ-PETR under W8A8 achieves near-floating-point accuracy (1% degradation) while reducing latency by up to 75%, significantly outperforming existing PTQ and QAT baselines.", "date": "2025-02-21", "categories": []} {"id": "2503.02687", "title": "Class-Aware PillarMix: Can Mixed Sample Data Augmentation Enhance 3D Object Detection with Radar Point Clouds?", "abstract": "Due to the significant effort required for data collection and annotation in 3D perception tasks, mixed sample data augmentation (MSDA) has been widely studied to generate diverse training samples by mixing existing data. Recently, many MSDA techniques have been developed for point clouds, but they mainly target LiDAR data, leaving their application to radar point clouds largely unexplored. In this paper, we examine the feasibility of applying existing MSDA methods to radar point clouds and identify several challenges in adapting these techniques. These obstacles stem from the radar's irregular angular distribution, deviations from a single-sensor polar layout in multi-radar setups, and point sparsity. To address these issues, we propose Class-Aware PillarMix (CAPMix), a novel MSDA approach that applies MixUp at the pillar level in 3D point clouds, guided by class labels. Unlike methods that rely a single mix ratio to the entire sample, CAPMix assigns an independent ratio to each pillar, boosting sample diversity. To account for the density of different classes, we use class-specific distributions: for dense objects (e.g., large vehicles), we skew ratios to favor points from another sample, while for sparse objects (e.g., pedestrians), we sample more points from the original. This class-aware mixing retains critical details and enriches each sample with new information, ultimately generating more diverse training data. Experimental results demonstrate that our method not only significantly boosts performance but also outperforms existing MSDA approaches across two datasets (Bosch Street and K-Radar). We believe that this straightforward yet effective approach will spark further investigation into MSDA techniques for radar data.", "date": "2025-03-04", "categories": []} {"id": "2504.00559", "title": "AttentiveGRU: Recurrent Spatio-Temporal Modeling for Advanced Radar-Based BEV Object Detection", "abstract": "Bird's-eye view (BEV) object detection has become important for advanced automotive 3D radar-based perception systems. However, the inherently sparse and non-deterministic nature of radar data limits the effectiveness of traditional single-frame BEV paradigms. In this paper, we addresses this limitation by introducing AttentiveGRU, a novel attention-based recurrent approach tailored for radar constraints, which extracts individualized spatio-temporal context for objects by dynamically identifying and fusing temporally correlated structures across present and memory states. By leveraging the consistency of object's latent representation over time, our approach exploits temporal relations to enrich feature representations for both stationary and moving objects, thereby enhancing detection performance and eliminating the need for externally providing or estimating any information about ego vehicle motion. Our experimental results on the public nuScenes dataset show a significant increase in mAP for the car category by 21% over the best radar-only submission. Further evaluations on an additional dataset demonstrate notable improvements in object detection capabilities, underscoring the applicability and effectiveness of our method.", "date": "2025-04-01", "categories": []} {"id": "2503.12093", "title": "SFMNet: Sparse Focal Modulation for 3D Object Detection", "abstract": "We propose SFMNet, a novel 3D sparse detector that combines the efficiency of sparse convolutions with the ability to model long-range dependencies. While traditional sparse convolution techniques efficiently capture local structures, they struggle with modeling long-range relationships. However, capturing long-range dependencies is fundamental for 3D object detection. In contrast, transformers are designed to capture these long-range dependencies through attention mechanisms. But, they come with high computational costs, due to their quadratic query-key-value interactions. Furthermore, directly applying attention to non-empty voxels is inefficient due to the sparse nature of 3D scenes. Our SFMNet is built on a novel Sparse Focal Modulation (SFM) module, which integrates short- and long-range contexts with linear complexity by leveraging a new hierarchical sparse convolution design. This approach enables SFMNet to achieve high detection performance with improved efficiency, making it well-suited for large-scale LiDAR scenes. We show that our detector achieves state-of-the-art performance on autonomous driving datasets.", "date": "2025-03-15", "categories": []} {"id": "2509.08374", "title": "InsFusion: Rethink Instance-level LiDAR-Camera Fusion for 3D Object Detection", "abstract": "Three-dimensional Object Detection from multi-view cameras and LiDAR is a crucial component for autonomous driving and smart transportation. However, in the process of basic feature extraction, perspective transformation, and feature fusion, noise and error will gradually accumulate. To address this issue, we propose InsFusion, which can extract proposals from both raw and fused features and utilizes these proposals to query the raw features, thereby mitigating the impact of accumulated errors. Additionally, by incorporating attention mechanisms applied to the raw features, it thereby mitigates the impact of accumulated errors. Experiments on the nuScenes dataset demonstrate that InsFusion is compatible with various advanced baseline methods and delivers new state-of-the-art performance for 3D object detection.", "date": "2025-09-10", "categories": []} {"id": "2511.13138", "title": "WinMamba: Multi-Scale Shifted Windows in State Space Model for 3D Object Detection", "abstract": "3D object detection is critical for autonomous driving, yet it remains fundamentally challenging to simultaneously maximize computational efficiency and capture long-range spatial dependencies. We observed that Mamba-based models, with their linear state-space design, capture long-range dependencies at lower cost, offering a promising balance between efficiency and accuracy. However, existing methods rely on axis-aligned scanning within a fixed window, inevitably discarding spatial information. To address this problem, we propose WinMamba, a novel Mamba-based 3D feature-encoding backbone composed of stacked WinMamba blocks. To enhance the backbone with robust multi-scale representation, the WinMamba block incorporates a window-scale-adaptive module that compensates voxel features across varying resolutions during sampling. Meanwhile, to obtain rich contextual cues within the linear state space, we equip the WinMamba layer with a learnable positional encoding and a window-shift strategy. Extensive experiments on the KITTI and Waymo datasets demonstrate that WinMamba significantly outperforms the baseline. Ablation studies further validate the individual contributions of the WSF and AWF modules in improving detection accuracy. The code will be made publicly available.", "date": "2025-11-17", "categories": []} {"id": "2511.13195", "title": "Difficulty-Aware Label-Guided Denoising for Monocular 3D Object Detection", "abstract": "Monocular 3D object detection is a cost-effective solution for applications like autonomous driving and robotics, but remains fundamentally ill-posed due to inherently ambiguous depth cues. Recent DETR-based methods attempt to mitigate this through global attention and auxiliary depth prediction, yet they still struggle with inaccurate depth estimates. Moreover, these methods often overlook instance-level detection difficulty, such as occlusion, distance, and truncation, leading to suboptimal detection performance. We propose MonoDLGD, a novel Difficulty-Aware Label-Guided Denoising framework that adaptively perturbs and reconstructs ground-truth labels based on detection uncertainty. Specifically, MonoDLGD applies stronger perturbations to easier instances and weaker ones into harder cases, and then reconstructs them to effectively provide explicit geometric supervision. By jointly optimizing label reconstruction and 3D object detection, MonoDLGD encourages geometry-aware representation learning and improves robustness to varying levels of object complexity. Extensive experiments on the KITTI benchmark demonstrate that MonoDLGD achieves state-of-the-art performance across all difficulty levels.", "date": "2025-11-17", "categories": []} {"id": "2508.01562", "title": "Adaptive LiDAR Scanning: Harnessing Temporal Cues for Efficient 3D Object Detection via Multi-Modal Fusion", "abstract": "Multi-sensor fusion using LiDAR and RGB cameras significantly enhances 3D object detection task. However, conventional LiDAR sensors perform dense, stateless scans, ignoring the strong temporal continuity in real-world scenes. This leads to substantial sensing redundancy and excessive power consumption, limiting their practicality on resource-constrained platforms. To address this inefficiency, we propose a predictive, history-aware adaptive scanning framework that anticipates informative regions of interest (ROI) based on past observations. Our approach introduces a lightweight predictor network that distills historical spatial and temporal contexts into refined query embeddings. These embeddings guide a differentiable Mask Generator network, which leverages Gumbel-Softmax sampling to produce binary masks identifying critical ROIs for the upcoming frame. Our method significantly reduces unnecessary data acquisition by concentrating dense LiDAR scanning only within these ROIs and sparsely sampling elsewhere. Experiments on nuScenes and Lyft benchmarks demonstrate that our adaptive scanning strategy reduces LiDAR energy consumption by over 65% while maintaining competitive or even superior 3D object detection performance compared to traditional LiDAR-camera fusion methods with dense LiDAR scanning.", "date": "2025-08-03", "categories": []} {"id": "2511.09347", "title": "FQ-PETR: Fully Quantized Position Embedding Transformation for Multi-View 3D Object Detection", "abstract": "Camera-based multi-view 3D detection is crucial for autonomous driving. PETR and its variants (PETRs) excel in benchmarks but face deployment challenges due to high computational cost and memory footprint. Quantization is an effective technique for compressing deep neural networks by reducing the bit width of weights and activations. However, directly applying existing quantization methods to PETRs leads to severe accuracy degradation. This issue primarily arises from two key challenges: (1) significant magnitude disparity between multi-modal features-specifically, image features and camera-ray positional embeddings (PE), and (2) the inefficiency and approximation error of quantizing non-linear operators, which commonly rely on hardware-unfriendly computations. In this paper, we propose FQ-PETR, a fully quantized framework for PETRs, featuring three key innovations: (1) Quantization-Friendly LiDAR-ray Position Embedding (QFPE): Replacing multi-point sampling with LiDAR-prior-guided single-point sampling and anchor-based embedding eliminates problematic non-linearities (e.g., inverse-sigmoid) and aligns PE scale with image features, preserving accuracy. (2) Dual-Lookup Table (DULUT): This algorithm approximates complex non-linear functions using two cascaded linear LUTs, achieving high fidelity with minimal entries and no specialized hardware. (3) Quantization After Numerical Stabilization (QANS): Performing quantization after softmax numerical stabilization mitigates attention distortion from large inputs. On PETRs (e.g. PETR, StreamPETR, PETRv2, MV2d), FQ-PETR under W8A8 achieves near-floating-point accuracy (1% degradation) while reducing latency by up to 75%, significantly outperforming existing PTQ and QAT baselines.", "date": "2025-11-12", "categories": []} {"id": "2511.10046", "title": "FreDFT: Frequency Domain Fusion Transformer for Visible-Infrared Object Detection", "abstract": "Visible-infrared object detection has gained sufficient attention due to its detection performance in low light, fog, and rain conditions. However, visible and infrared modalities captured by different sensors exist the information imbalance problem in complex scenarios, which can cause inadequate cross-modal fusion, resulting in degraded detection performance. \\textcolor{red}{Furthermore, most existing methods use transformers in the spatial domain to capture complementary features, ignoring the advantages of developing frequency domain transformers to mine complementary information.} To solve these weaknesses, we propose a frequency domain fusion transformer, called FreDFT, for visible-infrared object detection. The proposed approach employs a novel multimodal frequency domain attention (MFDA) to mine complementary information between modalities and a frequency domain feed-forward layer (FDFFL) via a mixed-scale frequency feature fusion strategy is designed to better enhance multimodal features. To eliminate the imbalance of multimodal information, a cross-modal global modeling module (CGMM) is constructed to perform pixel-wise inter-modal feature interaction in a spatial and channel manner. Moreover, a local feature enhancement module (LFEM) is developed to strengthen multimodal local feature representation and promote multimodal feature fusion by using various convolution layers and applying a channel shuffle. Extensive experimental results have verified that our proposed FreDFT achieves excellent performance on multiple public datasets compared with other state-of-the-art methods. The code of our FreDFT is linked at https://github.com/WenCongWu/FreDFT.", "date": "2025-11-13", "categories": []} {"id": "2412.02054", "title": "Redundant Queries in DETR-Based 3D Detection Methods: Unnecessary and Prunable", "abstract": "Query-based models are extensively used in 3D object detection tasks, with a wide range of pre-trained checkpoints readily available online. However, despite their popularity, these models often require an excessive number of object queries, far surpassing the actual number of objects to detect. The redundant queries result in unnecessary computational and memory costs. In this paper, we find that not all queries contribute equally -- a significant portion of queries have a much smaller impact compared to others. Based on this observation, we propose an embarrassingly simple approach called Gradually Pruning Queries (GPQ), which prunes queries incrementally based on their classification scores. A key advantage of GPQ is that it requires no additional learnable parameters. It is straightforward to implement in any query-based method, as it can be seamlessly integrated as a fine-tuning step using an existing checkpoint after training. With GPQ, users can easily generate multiple models with fewer queries, starting from a checkpoint with an excessive number of queries. Experiments on various advanced 3D detectors show that GPQ effectively reduces redundant queries while maintaining performance. Using our method, model inference on desktop GPUs can be accelerated by up to 1.35x. Moreover, after deployment on edge devices, it achieves up to a 67.86% reduction in FLOPs and a 65.16% decrease in inference time. The code will be available at https://github.com/iseri27/Gpq.", "date": "2024-12-03", "categories": []} {"id": "2511.09999", "title": "MOBA: A Material-Oriented Backdoor Attack against LiDAR-based 3D Object Detection Systems", "abstract": "LiDAR-based 3D object detection is widely used in safety-critical systems. However, these systems remain vulnerable to backdoor attacks that embed hidden malicious behaviors during training. A key limitation of existing backdoor attacks is their lack of physical realizability, primarily due to the digital-to-physical domain gap. Digital triggers often fail in real-world settings because they overlook material-dependent LiDAR reflection properties. On the other hand, physically constructed triggers are often unoptimized, leading to low effectiveness or easy detectability.This paper introduces Material-Oriented Backdoor Attack (MOBA), a novel framework that bridges the digital-physical gap by explicitly modeling the material properties of real-world triggers. MOBA tackles two key challenges in physical backdoor design: 1) robustness of the trigger material under diverse environmental conditions, 2) alignment between the physical trigger's behavior and its digital simulation. First, we propose a systematic approach to selecting robust trigger materials, identifying titanium dioxide (TiO_2) for its high diffuse reflectivity and environmental resilience. Second, to ensure the digital trigger accurately mimics the physical behavior of the material-based trigger, we develop a novel simulation pipeline that features: (1) an angle-independent approximation of the Oren-Nayar BRDF model to generate realistic LiDAR intensities, and (2) a distance-aware scaling mechanism to maintain spatial consistency across varying depths. We conduct extensive experiments on state-of-the-art LiDAR-based and Camera-LiDAR fusion models, showing that MOBA achieves a 93.50% attack success rate, outperforming prior methods by over 41%. Our work reveals a new class of physically realizable threats and underscores the urgent need for defenses that account for material-level properties in real-world environments.", "date": "2025-11-13", "categories": []} {"id": "2511.10035", "title": "DGFusion: Dual-guided Fusion for Robust Multi-Modal 3D Object Detection", "abstract": "As a critical task in autonomous driving perception systems, 3D object detection is used to identify and track key objects, such as vehicles and pedestrians. However, detecting distant, small, or occluded objects (hard instances) remains a challenge, which directly compromises the safety of autonomous driving systems. We observe that existing multi-modal 3D object detection methods often follow a single-guided paradigm, failing to account for the differences in information density of hard instances between modalities. In this work, we propose DGFusion, based on the Dual-guided paradigm, which fully inherits the advantages of the Point-guide-Image paradigm and integrates the Image-guide-Point paradigm to address the limitations of the single paradigms. The core of DGFusion, the Difficulty-aware Instance Pair Matcher (DIPM), performs instance-level feature matching based on difficulty to generate easy and hard instance pairs, while the Dual-guided Modules exploit the advantages of both pair types to enable effective multi-modal feature fusion. Experimental results demonstrate that our DGFusion outperforms the baseline methods, with respective improvements of +1.0\\% mAP, +0.8\\% NDS, and +1.3\\% average recall on nuScenes. Extensive experiments demonstrate consistent robustness gains for hard instance detection across ego-distance, size, visibility, and small-scale training scenarios.", "date": "2025-11-13", "categories": []} {"id": "2511.07862", "title": "MonoCLUE : Object-Aware Clustering Enhances Monocular 3D Object Detection", "abstract": "Monocular 3D object detection offers a cost-effective solution for autonomous driving but suffers from ill-posed depth and limited field of view. These constraints cause a lack of geometric cues and reduced accuracy in occluded or truncated scenes. While recent approaches incorporate additional depth information to address geometric ambiguity, they overlook the visual cues crucial for robust recognition. We propose MonoCLUE, which enhances monocular 3D detection by leveraging both local clustering and generalized scene memory of visual features. First, we perform K-means clustering on visual features to capture distinct object-level appearance parts (e.g., bonnet, car roof), improving detection of partially visible objects. The clustered features are propagated across regions to capture objects with similar appearances. Second, we construct a generalized scene memory by aggregating clustered features across images, providing consistent representations that generalize across scenes. This improves object-level feature consistency, enabling stable detection across varying environments. Lastly, we integrate both local cluster features and generalized scene memory into object queries, guiding attention toward informative regions. Exploiting a unified local clustering and generalized scene memory strategy, MonoCLUE enables robust monocular 3D detection under occlusion and limited visibility, achieving state-of-the-art performance on the KITTI benchmark.", "date": "2025-11-11", "categories": []} {"id": "2511.07966", "title": "Multi-Modal Assistance for Unsupervised Domain Adaptation on Point Cloud 3D Object Detection", "abstract": "Unsupervised domain adaptation for LiDAR-based 3D object detection (3D UDA) based on the teacher-student architecture with pseudo labels has achieved notable improvements in recent years. Although it is quite popular to collect point clouds and images simultaneously, little attention has been paid to the usefulness of image data in 3D UDA when training the models. In this paper, we propose an approach named MMAssist that improves the performance of 3D UDA with multi-modal assistance. A method is designed to align 3D features between the source domain and the target domain by using image and text features as bridges. More specifically, we project the ground truth labels or pseudo labels to the images to get a set of 2D bounding boxes. For each 2D box, we extract its image feature from a pre-trained vision backbone. A large vision-language model (LVLM) is adopted to extract the box's text description, and a pre-trained text encoder is used to obtain its text feature. During the training of the model in the source domain and the student model in the target domain, we align the 3D features of the predicted boxes with their corresponding image and text features, and the 3D features and the aligned features are fused with learned weights for the final prediction. The features between the student branch and the teacher branch in the target domain are aligned as well. To enhance the pseudo labels, we use an off-the-shelf 2D object detector to generate 2D bounding boxes from images and estimate their corresponding 3D boxes with the aid of point cloud, and these 3D boxes are combined with the pseudo labels generated by the teacher model. Experimental results show that our approach achieves promising performance compared with state-of-the-art methods in three domain adaptation tasks on three popular 3D object detection datasets. The code is available at https://github.com/liangp/MMAssist.", "date": "2025-11-11", "categories": []} {"id": "2412.20665", "title": "SM3Det: A Unified Model for Multi-Modal Remote Sensing Object Detection", "abstract": "With the rapid advancement of remote sensing technology, high-resolution multi-modal imagery is now more widely accessible. Conventional Object detection models are trained on a single dataset, often restricted to a specific imaging modality and annotation format. However, such an approach overlooks the valuable shared knowledge across multi-modalities and limits the model's applicability in more versatile scenarios. This paper introduces a new task called Multi-Modal Datasets and Multi-Task Object Detection (M2Det) for remote sensing, designed to accurately detect horizontal or oriented objects from any sensor modality. This task poses challenges due to 1) the trade-offs involved in managing multi-modal modelling and 2) the complexities of multi-task optimization. To address these, we establish a benchmark dataset and propose a unified model, SM3Det (Single Model for Multi-Modal datasets and Multi-Task object Detection). SM3Det leverages a grid-level sparse MoE backbone to enable joint knowledge learning while preserving distinct feature representations for different modalities. Furthermore, it integrates a consistency and synchronization optimization strategy using dynamic learning rate adjustment, allowing it to effectively handle varying levels of learning difficulty across modalities and tasks. Extensive experiments demonstrate SM3Det's effectiveness and generalizability, consistently outperforming specialized models on individual datasets. The code is available at https://github.com/zcablii/SM3Det.", "date": "2024-12-30", "categories": []} {"id": "2507.19856", "title": "RaGS: Unleashing 3D Gaussian Splatting from 4D Radar and Monocular Cues for 3D Object Detection", "abstract": "4D millimeter-wave radar is a promising sensing modality for autonomous driving, yet effective 3D object detection from 4D radar and monocular images remains challenging. Existing fusion approaches either rely on instance proposals lacking global context or dense BEV grids constrained by rigid structures, lacking a flexible and adaptive representation for diverse scenes. To address this, we propose RaGS, the first framework that leverages 3D Gaussian Splatting (GS) to fuse 4D radar and monocular cues for 3D object detection. 3D GS models the scene as a continuous field of Gaussians, enabling dynamic resource allocation to foreground objects while maintaining flexibility and efficiency. Moreover, the velocity dimension of 4D radar provides motion cues that help anchor and refine the spatial distribution of Gaussians. Specifically, RaGS adopts a cascaded pipeline to construct and progressively refine the Gaussian field. It begins with Frustum-based Localization Initiation (FLI), which unprojects foreground pixels to initialize coarse Gaussian centers. Then, Iterative Multimodal Aggregation (IMA) explicitly exploits image semantics and implicitly integrates 4D radar velocity geometry to refine the Gaussians within regions of interest. Finally, Multi-level Gaussian Fusion (MGF) renders the Gaussian field into hierarchical BEV features for 3D object detection. By dynamically focusing on sparse and informative regions, RaGS achieves object-centric precision and comprehensive scene perception. Extensive experiments on View-of-Delft, TJ4DRadSet, and OmniHD-Scenes demonstrate its robustness and SOTA performance. Code will be released.", "date": "2025-07-26", "categories": []} {"id": "2511.07106", "title": "HENet++: Hybrid Encoding and Multi-task Learning for 3D Perception and End-to-end Autonomous Driving", "abstract": "Three-dimensional feature extraction is a critical component of autonomous driving systems, where perception tasks such as 3D object detection, bird's-eye-view (BEV) semantic segmentation, and occupancy prediction serve as important constraints on 3D features. While large image encoders, high-resolution images, and long-term temporal inputs can significantly enhance feature quality and deliver remarkable performance gains, these techniques are often incompatible in both training and inference due to computational resource constraints. Moreover, different tasks favor distinct feature representations, making it difficult for a single model to perform end-to-end inference across multiple tasks while maintaining accuracy comparable to that of single-task models. To alleviate these issues, we present the HENet and HENet++ framework for multi-task 3D perception and end-to-end autonomous driving. Specifically, we propose a hybrid image encoding network that uses a large image encoder for short-term frames and a small one for long-term frames. Furthermore, our framework simultaneously extracts both dense and sparse features, providing more suitable representations for different tasks, reducing cumulative errors, and delivering more comprehensive information to the planning module. The proposed architecture maintains compatibility with various existing 3D feature extraction methods and supports multimodal inputs. HENet++ achieves state-of-the-art end-to-end multi-task 3D perception results on the nuScenes benchmark, while also attaining the lowest collision rate on the nuScenes end-to-end autonomous driving benchmark.", "date": "2025-11-10", "categories": []} {"id": "2511.04347", "title": "Evaluating the Impact of Weather-Induced Sensor Occlusion on BEVFusion for 3D Object Detection", "abstract": "Accurate 3D object detection is essential for automated vehicles to navigate safely in complex real-world environments. Bird's Eye View (BEV) representations, which project multi-sensor data into a top-down spatial format, have emerged as a powerful approach for robust perception. Although BEV-based fusion architectures have demonstrated strong performance through multimodal integration, the effects of sensor occlusions, caused by environmental conditions such as fog, haze, or physical obstructions, on 3D detection accuracy remain underexplored. In this work, we investigate the impact of occlusions on both camera and Light Detection and Ranging (LiDAR) outputs using the BEVFusion architecture, evaluated on the nuScenes dataset. Detection performance is measured using mean Average Precision (mAP) and the nuScenes Detection Score (NDS). Our results show that moderate camera occlusions lead to a 41.3% drop in mAP (from 35.6% to 20.9%) when detection is based only on the camera. On the other hand, LiDAR sharply drops in performance only under heavy occlusion, with mAP falling by 47.3% (from 64.7% to 34.1%), with a severe impact on long-range detection. In fused settings, the effect depends on which sensor is occluded: occluding the camera leads to a minor 4.1% drop (from 68.5% to 65.7%), while occluding LiDAR results in a larger 26.8% drop (to 50.1%), revealing the model's stronger reliance on LiDAR for the task of 3D object detection. Our results highlight the need for future research into occlusion-aware evaluation methods and improved sensor fusion techniques that can maintain detection accuracy in the presence of partial sensor failure or degradation due to adverse environmental conditions.", "date": "2025-11-06", "categories": []} {"id": "2506.04789", "title": "Object-X: Learning to Reconstruct Multi-Modal 3D Object Representations", "abstract": "Learning effective multi-modal 3D representations of objects is essential for numerous applications, such as augmented reality and robotics. Existing methods often rely on task-specific embeddings that are tailored either for semantic understanding or geometric reconstruction. As a result, these embeddings typically cannot be decoded into explicit geometry and simultaneously reused across tasks. In this paper, we propose Object-X, a versatile multi-modal object representation framework capable of encoding rich object embeddings (e.g. images, point cloud, text) and decoding them back into detailed geometric and visual reconstructions. Object-X operates by geometrically grounding the captured modalities in a 3D voxel grid and learning an unstructured embedding fusing the information from the voxels with the object attributes. The learned embedding enables 3D Gaussian Splatting-based object reconstruction, while also supporting a range of downstream tasks, including scene alignment, single-image 3D object reconstruction, and localization. Evaluations on two challenging real-world datasets demonstrate that Object-X produces high-fidelity novel-view synthesis comparable to standard 3D Gaussian Splatting, while significantly improving geometric accuracy. Moreover, Object-X achieves competitive performance with specialized methods in scene alignment and localization. Critically, our object-centric descriptors require 3-4 orders of magnitude less storage compared to traditional image- or point cloud-based approaches, establishing Object-X as a scalable and highly practical solution for multi-modal 3D scene representation.", "date": "2025-06-05", "categories": []} {"id": "2510.15991", "title": "CrossRay3D: Geometry and Distribution Guidance for Efficient Multimodal 3D Detection", "abstract": "The sparse cross-modality detector offers more advantages than its counterpart, the Bird's-Eye-View (BEV) detector, particularly in terms of adaptability for downstream tasks and computational cost savings. However, existing sparse detectors overlook the quality of token representation, leaving it with a sub-optimal foreground quality and limited performance. In this paper, we identify that the geometric structure preserved and the class distribution are the key to improving the performance of the sparse detector, and propose a Sparse Selector (SS). The core module of SS is Ray-Aware Supervision (RAS), which preserves rich geometric information during the training stage, and Class-Balanced Supervision, which adaptively reweights the salience of class semantics, ensuring that tokens associated with small objects are retained during token sampling. Thereby, outperforming other sparse multi-modal detectors in the representation of tokens. Additionally, we design Ray Positional Encoding (Ray PE) to address the distribution differences between the LiDAR modality and the image. Finally, we integrate the aforementioned module into an end-to-end sparse multi-modality detector, dubbed CrossRay3D. Experiments show that, on the challenging nuScenes benchmark, CrossRay3D achieves state-of-the-art performance with 72.4 mAP and 74.7 NDS, while running 1.84 faster than other leading methods. Moreover, CrossRay3D demonstrates strong robustness even in scenarios where LiDAR or camera data are partially or entirely missing.", "date": "2025-10-14", "categories": []} {"id": "2511.02293", "title": "3D Point Cloud Object Detection on Edge Devices for Split Computing", "abstract": "The field of autonomous driving technology is rapidly advancing, with deep learning being a key component. Particularly in the field of sensing, 3D point cloud data collected by LiDAR is utilized to run deep neural network models for 3D object detection. However, these state-of-the-art models are complex, leading to longer processing times and increased power consumption on edge devices. The objective of this study is to address these issues by leveraging Split Computing, a distributed machine learning inference method. Split Computing aims to lessen the computational burden on edge devices, thereby reducing processing time and power consumption. Furthermore, it minimizes the risk of data breaches by only transmitting intermediate data from the deep neural network model. Experimental results show that splitting after voxelization reduces the inference time by 70.8% and the edge device execution time by 90.0%. When splitting within the network, the inference time is reduced by up to 57.1%, and the edge device execution time is reduced by up to 69.5%.", "date": "2025-11-04", "categories": []} {"id": "2510.27166", "title": "M^3Detection: Multi-Frame Multi-Level Feature Fusion for Multi-Modal 3D Object Detection with Camera and 4D Imaging Radar", "abstract": "Recent advances in 4D imaging radar have enabled robust perception in adverse weather, while camera sensors provide dense semantic information. Fusing the these complementary modalities has great potential for cost-effective 3D perception. However, most existing camera-radar fusion methods are limited to single-frame inputs, capturing only a partial view of the scene. The incomplete scene information, compounded by image degradation and 4D radar sparsity, hinders overall detection performance. In contrast, multi-frame fusion offers richer spatiotemporal information but faces two challenges: achieving robust and effective object feature fusion across frames and modalities, and mitigating the computational cost of redundant feature extraction. Consequently, we propose M^3Detection, a unified multi-frame 3D object detection framework that performs multi-level feature fusion on multi-modal data from camera and 4D imaging radar. Our framework leverages intermediate features from the baseline detector and employs the tracker to produce reference trajectories, improving computational efficiency and providing richer information for second-stage. In the second stage, we design a global-level inter-object feature aggregation module guided by radar information to align global features across candidate proposals and a local-level inter-grid feature aggregation module that expands local features along the reference trajectories to enhance fine-grained object representation. The aggregated features are then processed by a trajectory-level multi-frame spatiotemporal reasoning module to encode cross-frame interactions and enhance temporal representation. Extensive experiments on the VoD and TJ4DRadSet datasets demonstrate that M^3Detection achieves state-of-the-art 3D detection performance, validating its effectiveness in multi-frame detection with camera-4D imaging radar fusion.", "date": "2025-10-31", "categories": []} {"id": "2510.24688", "title": "MIC-BEV: Multi-Infrastructure Camera Bird's-Eye-View Transformer with Relation-Aware Fusion for 3D Object Detection", "abstract": "Infrastructure-based perception plays a crucial role in intelligent transportation systems, offering global situational awareness and enabling cooperative autonomy. However, existing camera-based detection models often underperform in such scenarios due to challenges such as multi-view infrastructure setup, diverse camera configurations, degraded visual inputs, and various road layouts. We introduce MIC-BEV, a Transformer-based bird's-eye-view (BEV) perception framework for infrastructure-based multi-camera 3D object detection. MIC-BEV flexibly supports a variable number of cameras with heterogeneous intrinsic and extrinsic parameters and demonstrates strong robustness under sensor degradation. The proposed graph-enhanced fusion module in MIC-BEV integrates multi-view image features into the BEV space by exploiting geometric relationships between cameras and BEV cells alongside latent visual cues. To support training and evaluation, we introduce M2I, a synthetic dataset for infrastructure-based object detection, featuring diverse camera configurations, road layouts, and environmental conditions. Extensive experiments on both M2I and the real-world dataset RoScenes demonstrate that MIC-BEV achieves state-of-the-art performance in 3D object detection. It also remains robust under challenging conditions, including extreme weather and sensor degradation. These results highlight the potential of MIC-BEV for real-world deployment. The dataset and source code are available at: https://github.com/HandsomeYun/MIC-BEV.", "date": "2025-10-28", "categories": []} {"id": "2510.23144", "title": "DQ3D: Depth-guided Query for Transformer-Based 3D Object Detection in Traffic Scenarios", "abstract": "3D object detection from multi-view images in traffic scenarios has garnered significant attention in recent years. Many existing approaches rely on object queries that are generated from 3D reference points to localize objects. However, a limitation of these methods is that some reference points are often far from the target object, which can lead to false positive detections. In this paper, we propose a depth-guided query generator for 3D object detection (DQ3D) that leverages depth information and 2D detections to ensure that reference points are sampled from the surface or interior of the object. Furthermore, to address partially occluded objects in current frame, we introduce a hybrid attention mechanism that fuses historical detection results with depth-guided queries, thereby forming hybrid queries. Evaluation on the nuScenes dataset demonstrates that our method outperforms the baseline by 6.3\\% in terms of mean Average Precision (mAP) and 4.3\\% in the NuScenes Detection Score (NDS).", "date": "2025-10-27", "categories": []} {"id": "2510.23151", "title": "AG-Fusion: adaptive gated multimodal fusion for 3d object detection in complex scenes", "abstract": "Multimodal camera-LiDAR fusion technology has found extensive application in 3D object detection, demonstrating encouraging performance. However, existing methods exhibit significant performance degradation in challenging scenarios characterized by sensor degradation or environmental disturbances. We propose a novel Adaptive Gated Fusion (AG-Fusion) approach that selectively integrates cross-modal knowledge by identifying reliable patterns for robust detection in complex scenes. Specifically, we first project features from each modality into a unified BEV space and enhance them using a window-based attention mechanism. Subsequently, an adaptive gated fusion module based on cross-modal attention is designed to integrate these features into reliable BEV representations robust to challenging environments. Furthermore, we construct a new dataset named Excavator3D (E3D) focusing on challenging excavator operation scenarios to benchmark performance in complex conditions. Our method not only achieves competitive performance on the standard KITTI dataset with 93.92% accuracy, but also significantly outperforms the baseline by 24.88% on the challenging E3D dataset, demonstrating superior robustness to unreliable modal information in complex industrial scenes.", "date": "2025-10-27", "categories": []} {"id": "2508.00744", "title": "Rethinking Backbone Design for Lightweight 3D Object Detection in LiDAR", "abstract": "Recent advancements in LiDAR-based 3D object detection have significantly accelerated progress toward the realization of fully autonomous driving in real-world environments. Despite achieving high detection performance, most of the approaches still rely on a VGG-based or ResNet-based backbone for feature exploration, which increases the model complexity. Lightweight backbone design is well-explored for 2D object detection, but research on 3D object detection still remains limited. In this work, we introduce Dense Backbone, a lightweight backbone that combines the benefits of high processing speed, lightweight architecture, and robust detection accuracy. We adapt multiple SoTA 3d object detectors, such as PillarNet, with our backbone and show that with our backbone, these models retain most of their detection capability at a significantly reduced computational cost. To our knowledge, this is the first dense-layer-based backbone tailored specifically for 3D object detection from point cloud data. DensePillarNet, our adaptation of PillarNet, achieves a 29% reduction in model parameters and a 28% reduction in latency with just a 2% drop in detection accuracy on the nuScenes test set. Furthermore, Dense Backbone's plug-and-play design allows straightforward integration into existing architectures, requiring no modifications to other network components.", "date": "2025-08-01", "categories": []} {"id": "2510.19215", "title": "SFGFusion: Surface Fitting Guided 3D Object Detection with 4D Radar and Camera Fusion", "abstract": "3D object detection is essential for autonomous driving. As an emerging sensor, 4D imaging radar offers advantages as low cost, long-range detection, and accurate velocity measurement, making it highly suitable for object detection. However, its sparse point clouds and low resolution limit object geometric representation and hinder multi-modal fusion. In this study, we introduce SFGFusion, a novel camera-4D imaging radar detection network guided by surface fitting. By estimating quadratic surface parameters of objects from image and radar data, the explicit surface fitting model enhances spatial representation and cross-modal interaction, enabling more reliable prediction of fine-grained dense depth. The predicted depth serves two purposes: 1) in an image branch to guide the transformation of image features from perspective view (PV) to a unified bird's-eye view (BEV) for multi-modal fusion, improving spatial mapping accuracy; and 2) in a surface pseudo-point branch to generate dense pseudo-point cloud, mitigating the radar point sparsity. The original radar point cloud is also encoded in a separate radar branch. These two point cloud branches adopt a pillar-based method and subsequently transform the features into the BEV space. Finally, a standard 2D backbone and detection head are used to predict object labels and bounding boxes from BEV features. Experimental results show that SFGFusion effectively fuses camera and 4D radar features, achieving superior performance on the TJ4DRadSet and view-of-delft (VoD) object detection benchmarks.", "date": "2025-10-22", "categories": []} {"id": "2510.18539", "title": "GBlobs: Local LiDAR Geometry for Improved Sensor Placement Generalization", "abstract": "This technical report outlines the top-ranking solution for RoboSense 2025: Track 3, achieving state-of-the-art performance on 3D object detection under various sensor placements. Our submission utilizes GBlobs, a local point cloud feature descriptor specifically designed to enhance model generalization across diverse LiDAR configurations. Current LiDAR-based 3D detectors often suffer from a \\enquote{geometric shortcut} when trained on conventional global features (\\ie, absolute Cartesian coordinates). This introduces a position bias that causes models to primarily rely on absolute object position rather than distinguishing shape and appearance characteristics. Although effective for in-domain data, this shortcut severely limits generalization when encountering different point distributions, such as those resulting from varying sensor placements. By using GBlobs as network input features, we effectively circumvent this geometric shortcut, compelling the network to learn robust, object-centric representations. This approach significantly enhances the model's ability to generalize, resulting in the exceptional performance demonstrated in this challenge.", "date": "2025-10-21", "categories": []} {"id": "2510.17686", "title": "Towards 3D Objectness Learning in an Open World", "abstract": "Recent advancements in 3D object detection and novel category detection have made significant progress, yet research on learning generalized 3D objectness remains insufficient. In this paper, we delve into learning open-world 3D objectness, which focuses on detecting all objects in a 3D scene, including novel objects unseen during training. Traditional closed-set 3D detectors struggle to generalize to open-world scenarios, while directly incorporating 3D open-vocabulary models for open-world ability struggles with vocabulary expansion and semantic overlap. To achieve generalized 3D object discovery, We propose OP3Det, a class-agnostic Open-World Prompt-free 3D Detector to detect any objects within 3D scenes without relying on hand-crafted text prompts. We introduce the strong generalization and zero-shot capabilities of 2D foundation models, utilizing both 2D semantic priors and 3D geometric priors for class-agnostic proposals to broaden 3D object discovery. Then, by integrating complementary information from point cloud and RGB image in the cross-modal mixture of experts, OP3Det dynamically routes uni-modal and multi-modal features to learn generalized 3D objectness. Extensive experiments demonstrate the extraordinary performance of OP3Det, which significantly surpasses existing open-world 3D detectors by up to 16.0% in AR and achieves a 13.5% improvement compared to closed-world 3D detectors.", "date": "2025-10-20", "categories": []} {"id": "2503.12009", "title": "UniMamba: Unified Spatial-Channel Representation Learning with Group-Efficient Mamba for LiDAR-based 3D Object Detection", "abstract": "Recent advances in LiDAR 3D detection have demonstrated the effectiveness of Transformer-based frameworks in capturing the global dependencies from point cloud spaces, which serialize the 3D voxels into the flattened 1D sequence for iterative self-attention. However, the spatial structure of 3D voxels will be inevitably destroyed during the serialization process. Besides, due to the considerable number of 3D voxels and quadratic complexity of Transformers, multiple sequences are grouped before feeding to Transformers, leading to a limited receptive field. Inspired by the impressive performance of State Space Models (SSM) achieved in the field of 2D vision tasks, in this paper, we propose a novel Unified Mamba (UniMamba), which seamlessly integrates the merits of 3D convolution and SSM in a concise multi-head manner, aiming to perform \"local and global\" spatial context aggregation efficiently and simultaneously. Specifically, a UniMamba block is designed which mainly consists of spatial locality modeling, complementary Z-order serialization and local-global sequential aggregator. The spatial locality modeling module integrates 3D submanifold convolution to capture the dynamic spatial position embedding before serialization. Then the efficient Z-order curve is adopted for serialization both horizontally and vertically. Furthermore, the local-global sequential aggregator adopts the channel grouping strategy to efficiently encode both \"local and global\" spatial inter-dependencies using multi-head SSM. Additionally, an encoder-decoder architecture with stacked UniMamba blocks is formed to facilitate multi-scale spatial learning hierarchically. Extensive experiments are conducted on three popular datasets: nuScenes, Waymo and Argoverse 2. Particularly, our UniMamba achieves 70.2 mAP on the nuScenes dataset.", "date": "2025-03-15", "categories": []} {"id": "2510.15385", "title": "FreqPDE: Rethinking Positional Depth Embedding for Multi-View 3D Object Detection Transformers", "abstract": "Detecting 3D objects accurately from multi-view 2D images is a challenging yet essential task in the field of autonomous driving. Current methods resort to integrating depth prediction to recover the spatial information for object query decoding, which necessitates explicit supervision from LiDAR points during the training phase. However, the predicted depth quality is still unsatisfactory such as depth discontinuity of object boundaries and indistinction of small objects, which are mainly caused by the sparse supervision of projected points and the use of high-level image features for depth prediction. Besides, cross-view consistency and scale invariance are also overlooked in previous methods. In this paper, we introduce Frequency-aware Positional Depth Embedding (FreqPDE) to equip 2D image features with spatial information for 3D detection transformer decoder, which can be obtained through three main modules. Specifically, the Frequency-aware Spatial Pyramid Encoder (FSPE) constructs a feature pyramid by combining high-frequency edge clues and low-frequency semantics from different levels respectively. Then the Cross-view Scale-invariant Depth Predictor (CSDP) estimates the pixel-level depth distribution with cross-view and efficient channel attention mechanism. Finally, the Positional Depth Encoder (PDE) combines the 2D image features and 3D position embeddings to generate the 3D depth-aware features for query decoding. Additionally, hybrid depth supervision is adopted for complementary depth learning from both metric and distribution aspects. Extensive experiments conducted on the nuScenes dataset demonstrate the effectiveness and superiority of our proposed method.", "date": "2025-10-17", "categories": []} {"id": "2510.10287", "title": "Bridging Perspectives: Foundation Model Guided BEV Maps for 3D Object Detection and Tracking", "abstract": "Camera-based 3D object detection and tracking are essential for perception in autonomous driving. Current state-of-the-art approaches often rely exclusively on either perspective-view (PV) or bird's-eye-view (BEV) features, limiting their ability to leverage both fine-grained object details and spatially structured scene representations. In this work, we propose DualViewDistill, a hybrid detection and tracking framework that incorporates both PV and BEV camera image features to leverage their complementary strengths. Our approach introduces BEV maps guided by foundation models, leveraging descriptive DINOv2 features that are distilled into BEV representations through a novel distillation process. By integrating PV features with BEV maps enriched with semantic and geometric features from DINOv2, our model leverages this hybrid representation via deformable aggregation to enhance 3D object detection and tracking. Extensive experiments on the nuScenes and Argoverse 2 benchmarks demonstrate that DualViewDistill achieves state-of-the-art performance. The results showcase the potential of foundation model BEV maps to enable more reliable perception for autonomous driving. We make the code and pre-trained models available at https://dualviewdistill.cs.uni-freiburg.de .", "date": "2025-10-11", "categories": []} {"id": "2510.11632", "title": "NV3D: Leveraging Spatial Shape Through Normal Vector-based 3D Object Detection", "abstract": "Recent studies in 3D object detection for autonomous vehicles aim to enrich features through the utilization of multi-modal setups or the extraction of local patterns within LiDAR point clouds. However, multi-modal methods face significant challenges in feature alignment, and gaining features locally can be oversimplified for complex 3D object detection tasks. In this paper, we propose a novel model, NV3D, which utilizes local features acquired from voxel neighbors, as normal vectors computed per voxel basis using K-nearest neighbors (KNN) and principal component analysis (PCA). This informative feature enables NV3D to determine the relationship between the surface and pertinent target entities, including cars, pedestrians, or cyclists. During the normal vector extraction process, NV3D offers two distinct sampling strategies: normal vector density-based sampling and FOV-aware bin-based sampling, allowing elimination of up to 55% of data while maintaining performance. In addition, we applied element-wise attention fusion, which accepts voxel features as the query and value and normal vector features as the key, similar to the attention mechanism. Our method is trained on the KITTI dataset and has demonstrated superior performance in car and cyclist detection owing to their spatial shapes. In the validation set, NV3D without sampling achieves 86.60% and 80.18% mean Average Precision (mAP), greater than the baseline Voxel R-CNN by 2.61% and 4.23% mAP, respectively. With both samplings, NV3D achieves 85.54% mAP in car detection, exceeding the baseline by 1.56% mAP, despite roughly 55% of voxels being filtered out.", "date": "2025-10-13", "categories": []} {"id": "2510.08017", "title": "RayFusion: Ray Fusion Enhanced Collaborative Visual Perception", "abstract": "Collaborative visual perception methods have gained widespread attention in the autonomous driving community in recent years due to their ability to address sensor limitation problems. However, the absence of explicit depth information often makes it difficult for camera-based perception systems, e.g., 3D object detection, to generate accurate predictions. To alleviate the ambiguity in depth estimation, we propose RayFusion, a ray-based fusion method for collaborative visual perception. Using ray occupancy information from collaborators, RayFusion reduces redundancy and false positive predictions along camera rays, enhancing the detection performance of purely camera-based collaborative perception systems. Comprehensive experiments show that our method consistently outperforms existing state-of-the-art models, substantially advancing the performance of collaborative visual perception. The code is available at https://github.com/wangsh0111/RayFusion.", "date": "2025-10-09", "categories": []} {"id": "2501.04969", "title": "Self-Supervised Representation Learning with Joint Embedding Predictive Architecture for Automotive LiDAR Object Detection", "abstract": "Recently, self-supervised representation learning relying on vast amounts of unlabeled data has been explored as a pre-training method for autonomous driving. However, directly applying popular contrastive or generative methods to this problem is insufficient and may even lead to negative transfer. In this paper, we present AD-L-JEPA, a novel self-supervised pre-training framework with a joint embedding predictive architecture (JEPA) for automotive LiDAR object detection. Unlike existing methods, AD-L-JEPA is neither generative nor contrastive. Instead of explicitly generating masked regions, our method predicts Bird's-Eye-View embeddings to capture the diverse nature of driving scenes. Furthermore, our approach eliminates the need to manually form contrastive pairs by employing explicit variance regularization to avoid representation collapse. Experimental results demonstrate consistent improvements on the LiDAR 3D object detection downstream task across the KITTI3D, Waymo, and ONCE datasets, while reducing GPU hours by 1.9x-2.7x and GPU memory by 2.8x-4x compared with the state-of-the-art method Occupancy-MAE. Notably, on the largest ONCE dataset, pre-training on 100K frames yields a 1.61 mAP gain, better than all other methods pre-trained on either 100K or 500K frames, and pre-training on 500K frames yields a 2.98 mAP gain, better than all other methods pre-trained on either 500K or 1M frames. AD-L-JEPA constitutes the first JEPA-based pre-training method for autonomous driving. It offers better quality, faster, and more GPU-memory-efficient self-supervised representation learning. The source code of AD-L-JEPA is ready to be released.", "date": "2025-01-09", "categories": []} {"id": "2510.01829", "title": "Calibrating the Full Predictive Class Distribution of 3D Object Detectors for Autonomous Driving", "abstract": "In autonomous systems, precise object detection and uncertainty estimation are critical for self-aware and safe operation. This work addresses confidence calibration for the classification task of 3D object detectors. We argue that it is necessary to regard the calibration of the full predictive confidence distribution over all classes and deduce a metric which captures the calibration of dominant and secondary class predictions. We propose two auxiliary regularizing loss terms which introduce either calibration of the dominant prediction or the full prediction vector as a training goal. We evaluate a range of post-hoc and train-time methods for CenterPoint, PillarNet and DSVT-Pillar and find that combining our loss term, which regularizes for calibration of the full class prediction, and isotonic regression lead to the best calibration of CenterPoint and PillarNet with respect to both dominant and secondary class predictions. We further find that DSVT-Pillar can not be jointly calibrated for dominant and secondary predictions using the same method.", "date": "2025-10-02", "categories": []} {"id": "2509.15935", "title": "PAN: Pillars-Attention-Based Network for 3D Object Detection", "abstract": "Camera-radar fusion offers a robust and low-cost alternative to Camera-lidar fusion for the 3D object detection task in real-time under adverse weather and lighting conditions. However, currently, in the literature, it is possible to find few works focusing on this modality and, most importantly, developing new architectures to explore the advantages of the radar point cloud, such as accurate distance estimation and speed information. Therefore, this work presents a novel and efficient 3D object detection algorithm using cameras and radars in the bird's-eye-view (BEV). Our algorithm exploits the advantages of radar before fusing the features into a detection head. A new backbone is introduced, which maps the radar pillar features into an embedded dimension. A self-attention mechanism allows the backbone to model the dependencies between the radar points. We are using a simplified convolutional layer to replace the FPN-based convolutional layers used in the PointPillars-based architectures with the main goal of reducing inference time. Our results show that with this modification, our approach achieves the new state-of-the-art in the 3D object detection problem, reaching 58.2 of the NDS metric for the use of ResNet-50, while also setting a new benchmark for inference time on the nuScenes dataset for the same category.", "date": "2025-09-19", "categories": []} {"id": "2404.15224", "title": "Deep Models for Multi-View 3D Object Recognition: A Review", "abstract": "Human decision-making often relies on visual information from multiple perspectives or views. In contrast, machine learning-based object recognition utilizes information from a single image of the object. However, the information conveyed by a single image may not be sufficient for accurate decision-making, particularly in complex recognition problems. The utilization of multi-view 3D representations for object recognition has thus far demonstrated the most promising results for achieving state-of-the-art performance. This review paper comprehensively covers recent progress in multi-view 3D object recognition methods for 3D classification and retrieval tasks. Specifically, we focus on deep learning-based and transformer-based techniques, as they are widely utilized and have achieved state-of-the-art performance. We provide detailed information about existing deep learning-based and transformer-based multi-view 3D object recognition models, including the most commonly used 3D datasets, camera configurations and number of views, view selection strategies, pre-trained CNN architectures, fusion strategies, and recognition performance on 3D classification and 3D retrieval tasks. Additionally, we examine various computer vision applications that use multi-view classification. Finally, we highlight key findings and future directions for developing multi-view 3D object recognition methods to provide readers with a comprehensive understanding of the field.", "date": "2024-04-23", "categories": []} {"id": "2509.23700", "title": "INSTINCT: Instance-Level Interaction Architecture for Query-Based Collaborative Perception", "abstract": "Collaborative perception systems overcome single-vehicle limitations in long-range detection and occlusion scenarios by integrating multi-agent sensory data, improving accuracy and safety. However, frequent cooperative interactions and real-time requirements impose stringent bandwidth constraints. Previous works proves that query-based instance-level interaction reduces bandwidth demands and manual priors, however, LiDAR-focused implementations in collaborative perception remain underdeveloped, with performance still trailing state-of-the-art approaches. To bridge this gap, we propose INSTINCT (INSTance-level INteraCtion ArchiTecture), a novel collaborative perception framework featuring three core components: 1) a quality-aware filtering mechanism for high-quality instance feature selection; 2) a dual-branch detection routing scheme to decouple collaboration-irrelevant and collaboration-relevant instances; and 3) a Cross Agent Local Instance Fusion module to aggregate local hybrid instance features. Additionally, we enhance the ground truth (GT) sampling technique to facilitate training with diverse hybrid instance features. Extensive experiments across multiple datasets demonstrate that INSTINCT achieves superior performance. Specifically, our method achieves an improvement in accuracy 13.23%/33.08% in DAIR-V2X and V2V4Real while reducing the communication bandwidth to 1/281 and 1/264 compared to state-of-the-art methods. The code is available at https://github.com/CrazyShout/INSTINCT.", "date": "2025-09-28", "categories": []} {"id": "2509.24802", "title": "TACO-Net: Topological Signatures Triumph in 3D Object Classification", "abstract": "3D object classification is a crucial problem due to its significant practical relevance in many fields, including computer vision, robotics, and autonomous driving. Although deep learning methods applied to point clouds sampled on CAD models of the objects and/or captured by LiDAR or RGBD cameras have achieved remarkable success in recent years, achieving high classification accuracy remains a challenging problem due to the unordered point clouds and their irregularity and noise. To this end, we propose a novel state-of-the-art (SOTA) 3D object classification technique that combines topological data analysis with various image filtration techniques to classify objects when they are represented using point clouds. We transform every point cloud into a voxelized binary 3D image to extract distinguishing topological features. Next, we train a lightweight one-dimensional Convolutional Neural Network (1D CNN) using the extracted feature set from the training dataset. Our framework, TACO-Net, sets a new state-of-the-art by achieving $99.05\\%$ and $99.52\\%$ accuracy on the widely used synthetic benchmarks ModelNet40 and ModelNet10, and further demonstrates its robustness on the large-scale real-world OmniObject3D dataset. When tested with ten different kinds of corrupted ModelNet40 inputs, the proposed TACO-Net demonstrates strong resiliency overall.", "date": "2025-09-29", "categories": []} {"id": "2404.09931", "title": "Zero-shot detection of buildings in mobile LiDAR using Language Vision Model", "abstract": "Recent advances have demonstrated that Language Vision Models (LVMs) surpass the existing State-of-the-Art (SOTA) in two-dimensional (2D) computer vision tasks, motivating attempts to apply LVMs to three-dimensional (3D) data. While LVMs are efficient and effective in addressing various downstream 2D vision tasks without training, they face significant challenges when it comes to point clouds, a representative format for representing 3D data. It is more difficult to extract features from 3D data and there are challenges due to large data sizes and the cost of the collection and labelling, resulting in a notably limited availability of datasets. Moreover, constructing LVMs for point clouds is even more challenging due to the requirements for large amounts of data and training time. To address these issues, our research aims to 1) apply the Grounded SAM through Spherical Projection to transfer 3D to 2D, and 2) experiment with synthetic data to evaluate its effectiveness in bridging the gap between synthetic and real-world data domains. Our approach exhibited high performance with an accuracy of 0.96, an IoU of 0.85, precision of 0.92, recall of 0.91, and an F1 score of 0.92, confirming its potential. However, challenges such as occlusion problems and pixel-level overlaps of multi-label points during spherical image generation remain to be addressed in future studies.", "date": "2024-04-15", "categories": []} {"id": "2509.21388", "title": "TUN3D: Towards Real-World Scene Understanding from Unposed Images", "abstract": "Layout estimation and 3D object detection are two fundamental tasks in indoor scene understanding. When combined, they enable the creation of a compact yet semantically rich spatial representation of a scene. Existing approaches typically rely on point cloud input, which poses a major limitation since most consumer cameras lack depth sensors and visual-only data remains far more common. We address this issue with TUN3D, the first method that tackles joint layout estimation and 3D object detection in real scans, given multi-view images as input, and does not require ground-truth camera poses or depth supervision. Our approach builds on a lightweight sparse-convolutional backbone and employs two dedicated heads: one for 3D object detection and one for layout estimation, leveraging a novel and effective parametric wall representation. Extensive experiments show that TUN3D achieves state-of-the-art performance across three challenging scene understanding benchmarks: (i) using ground-truth point clouds, (ii) using posed images, and (iii) using unposed images. While performing on par with specialized 3D object detection methods, TUN3D significantly advances layout estimation, setting a new benchmark in holistic indoor scene understanding. Code is available at https://github.com/col14m/tun3d .", "date": "2025-09-23", "categories": []} {"id": "2411.19860", "title": "SpaRC: Sparse Radar-Camera Fusion for 3D Object Detection", "abstract": "In this work, we present SpaRC, a novel Sparse fusion transformer for 3D perception that integrates multi-view image semantics with Radar and Camera point features. The fusion of radar and camera modalities has emerged as an efficient perception paradigm for autonomous driving systems. While conventional approaches utilize dense Bird's Eye View (BEV)-based architectures for depth estimation, contemporary query-based transformers excel in camera-only detection through object-centric methodology. However, these query-based approaches exhibit limitations in false positive detections and localization precision due to implicit depth modeling. We address these challenges through three key contributions: (1) sparse frustum fusion (SFF) for cross-modal feature alignment, (2) range-adaptive radar aggregation (RAR) for precise object localization, and (3) local self-attention (LSA) for focused query aggregation. In contrast to existing methods requiring computationally intensive BEV-grid rendering, SpaRC operates directly on encoded point features, yielding substantial improvements in efficiency and accuracy. Empirical evaluations on the nuScenes and TruckScenes benchmarks demonstrate that SpaRC significantly outperforms existing dense BEV-based and sparse query-based detectors. Our method achieves state-of-the-art performance metrics of 67.1 NDS and 63.1 AMOTA. The code and pretrained models are available at https://github.com/phi-wol/sparc.", "date": "2024-11-29", "categories": []} {"id": "2509.18613", "title": "MLF-4DRCNet: Multi-Level Fusion with 4D Radar and Camera for 3D Object Detection in Autonomous Driving", "abstract": "The emerging 4D millimeter-wave radar, measuring the range, azimuth, elevation, and Doppler velocity of objects, is recognized for its cost-effectiveness and robustness in autonomous driving. Nevertheless, its point clouds exhibit significant sparsity and noise, restricting its standalone application in 3D object detection. Recent 4D radar-camera fusion methods have provided effective perception. Most existing approaches, however, adopt explicit Bird's-Eye-View fusion paradigms originally designed for LiDAR-camera fusion, neglecting radar's inherent drawbacks. Specifically, they overlook the sparse and incomplete geometry of radar point clouds and restrict fusion to coarse scene-level integration. To address these problems, we propose MLF-4DRCNet, a novel two-stage framework for 3D object detection via multi-level fusion of 4D radar and camera images. Our model incorporates the point-, scene-, and proposal-level multi-modal information, enabling comprehensive feature representation. It comprises three crucial components: the Enhanced Radar Point Encoder (ERPE) module, the Hierarchical Scene Fusion Pooling (HSFP) module, and the Proposal-Level Fusion Enhancement (PLFE) module. Operating at the point-level, ERPE densities radar point clouds with 2D image instances and encodes them into voxels via the proposed Triple-Attention Voxel Feature Encoder. HSFP dynamically integrates multi-scale voxel features with 2D image features using deformable attention to capture scene context and adopts pooling to the fused features. PLFE refines region proposals by fusing image features, and further integrates with the pooled features from HSFP. Experimental results on the View-of-Delft (VoD) and TJ4DRadSet datasets demonstrate that MLF-4DRCNet achieves the state-of-the-art performance. Notably, it attains performance comparable to LiDAR-based models on the VoD dataset.", "date": "2025-09-23", "categories": []} {"id": "2509.17462", "title": "MAESTRO: Task-Relevant Optimization via Adaptive Feature Enhancement and Suppression for Multi-task 3D Perception", "abstract": "The goal of multi-task learning is to learn to conduct multiple tasks simultaneously based on a shared data representation. While this approach can improve learning efficiency, it may also cause performance degradation due to task conflicts that arise when optimizing the model for different objectives. To address this challenge, we introduce MAESTRO, a structured framework designed to generate task-specific features and mitigate feature interference in multi-task 3D perception, including 3D object detection, bird's-eye view (BEV) map segmentation, and 3D occupancy prediction. MAESTRO comprises three components: the Class-wise Prototype Generator (CPG), the Task-Specific Feature Generator (TSFG), and the Scene Prototype Aggregator (SPA). CPG groups class categories into foreground and background groups and generates group-wise prototypes. The foreground and background prototypes are assigned to the 3D object detection task and the map segmentation task, respectively, while both are assigned to the 3D occupancy prediction task. TSFG leverages these prototype groups to retain task-relevant features while suppressing irrelevant features, thereby enhancing the performance for each task. SPA enhances the prototype groups assigned for 3D occupancy prediction by utilizing the information produced by the 3D object detection head and the map segmentation head. Extensive experiments on the nuScenes and Occ3D benchmarks demonstrate that MAESTRO consistently outperforms existing methods across 3D object detection, BEV map segmentation, and 3D occupancy prediction tasks.", "date": "2025-09-22", "categories": []} {"id": "2509.17712", "title": "RCTDistill: Cross-Modal Knowledge Distillation Framework for Radar-Camera 3D Object Detection with Temporal Fusion", "abstract": "Radar-camera fusion methods have emerged as a cost-effective approach for 3D object detection but still lag behind LiDAR-based methods in performance. Recent works have focused on employing temporal fusion and Knowledge Distillation (KD) strategies to overcome these limitations. However, existing approaches have not sufficiently accounted for uncertainties arising from object motion or sensor-specific errors inherent in radar and camera modalities. In this work, we propose RCTDistill, a novel cross-modal KD method based on temporal fusion, comprising three key modules: Range-Azimuth Knowledge Distillation (RAKD), Temporal Knowledge Distillation (TKD), and Region-Decoupled Knowledge Distillation (RDKD). RAKD is designed to consider the inherent errors in the range and azimuth directions, enabling effective knowledge transfer from LiDAR features to refine inaccurate BEV representations. TKD mitigates temporal misalignment caused by dynamic objects by aligning historical radar-camera BEV features with current LiDAR representations. RDKD enhances feature discrimination by distilling relational knowledge from the teacher model, allowing the student to differentiate foreground and background features. RCTDistill achieves state-of-the-art radar-camera fusion performance on both the nuScenes and View-of-Delft (VoD) datasets, with the fastest inference speed of 26.2 FPS.", "date": "2025-09-22", "categories": []} {"id": "2509.15924", "title": "Sparse Multiview Open-Vocabulary 3D Detection", "abstract": "The ability to interpret and comprehend a 3D scene is essential for many vision and robotics systems. In numerous applications, this involves 3D object detection, i.e.~identifying the location and dimensions of objects belonging to a specific category, typically represented as bounding boxes. This has traditionally been solved by training to detect a fixed set of categories, which limits its use. In this work, we investigate open-vocabulary 3D object detection in the challenging yet practical sparse-view setting, where only a limited number of posed RGB images are available as input. Our approach is training-free, relying on pre-trained, off-the-shelf 2D foundation models instead of employing computationally expensive 3D feature fusion or requiring 3D-specific learning. By lifting 2D detections and directly optimizing 3D proposals for featuremetric consistency across views, we fully leverage the extensive training data available in 2D compared to 3D. Through standard benchmarks, we demonstrate that this simple pipeline establishes a powerful baseline, performing competitively with state-of-the-art techniques in densely sampled scenarios while significantly outperforming them in the sparse-view setting.", "date": "2025-09-19", "categories": []} {"id": "2509.14151", "title": "BEVUDA++: Geometric-aware Unsupervised Domain Adaptation for Multi-View 3D Object Detection", "abstract": "Vision-centric Bird's Eye View (BEV) perception holds considerable promise for autonomous driving. Recent studies have prioritized efficiency or accuracy enhancements, yet the issue of domain shift has been overlooked, leading to substantial performance degradation upon transfer. We identify major domain gaps in real-world cross-domain scenarios and initiate the first effort to address the Domain Adaptation (DA) challenge in multi-view 3D object detection for BEV perception. Given the complexity of BEV perception approaches with their multiple components, domain shift accumulation across multi-geometric spaces (e.g., 2D, 3D Voxel, BEV) poses a significant challenge for BEV domain adaptation. In this paper, we introduce an innovative geometric-aware teacher-student framework, BEVUDA++, to diminish this issue, comprising a Reliable Depth Teacher (RDT) and a Geometric Consistent Student (GCS) model. Specifically, RDT effectively blends target LiDAR with dependable depth predictions to generate depth-aware information based on uncertainty estimation, enhancing the extraction of Voxel and BEV features that are essential for understanding the target domain. To collaboratively reduce the domain shift, GCS maps features from multiple spaces into a unified geometric embedding space, thereby narrowing the gap in data distribution between the two domains. Additionally, we introduce a novel Uncertainty-guided Exponential Moving Average (UEMA) to further reduce error accumulation due to domain shifts informed by previously obtained uncertainty guidance. To demonstrate the superiority of our proposed method, we execute comprehensive experiments in four cross-domain scenarios, securing state-of-the-art performance in BEV 3D object detection tasks, e.g., 12.9\\% NDS and 9.5\\% mAP enhancement on Day-Night adaptation.", "date": "2025-09-17", "categories": []} {"id": "2312.10986", "title": "Long-Tailed 3D Detection via Multi-Modal Fusion", "abstract": "Contemporary autonomous vehicle (AV) benchmarks have advanced techniques for training 3D detectors. While class labels naturally follow a long-tailed distribution in the real world, existing benchmarks only focus on a few common classes (e.g., pedestrian and car) and neglect many rare but crucial classes (e.g., emergency vehicle and stroller). However, AVs must reliably detect both common and rare classes for safe operation in the open world. We address this challenge by formally studying the problem of Long-Tailed 3D Detection (LT3D), which evaluates all annotated classes, including those in-the-tail. We address LT3D with hierarchical losses that promote feature sharing across classes, and introduce diagnostic metrics that award partial credit to \"reasonable\" mistakes with respect to the semantic hierarchy. Further, we point out that rare-class accuracy is particularly improved via multi-modal late fusion (MMLF) of independently trained uni-modal LiDAR and RGB detectors. Such an MMLF framework allows us to leverage large-scale uni-modal datasets (with more examples for rare classes) to train better uni-modal detectors. Finally, we examine three critical components of our simple MMLF approach from first principles: whether to train 2D or 3D RGB detectors for fusion, whether to match RGB and LiDAR detections in 3D or the projected 2D image plane, and how to fuse matched detections. Extensive experiments reveal that 2D RGB detectors achieve better recognition accuracy for rare classes than 3D RGB detectors, matching on the 2D image plane mitigates depth estimation errors for better matching, and score calibration and probabilistic fusion notably improves the final performance further. Our MMLF significantly outperforms prior work for LT3D, particularly improving on the six rarest classes from 12.8 to 20.0 mAP! Our code and models are available on our project page.", "date": "2023-12-18", "categories": []} {"id": "2410.23767", "title": "HD-OOD3D: Supervised and Unsupervised Out-of-Distribution object detection in LiDAR data", "abstract": "Autonomous systems rely on accurate 3D object detection from LiDAR data, yet most detectors are limited to a predefined set of known classes, making them vulnerable to unexpected out-of-distribution (OOD) objects. In this work, we present HD-OOD3D, a novel two-stage method for detecting unknown objects. We demonstrate the superiority of two-stage approaches over single-stage methods, achieving more robust detection of unknown objects while addressing key challenges in the evaluation protocol. Furthermore, we conduct an in-depth analysis of the standard evaluation protocol for OOD detection, revealing the critical impact of hyperparameter choices. To address the challenge of scaling the learning of unknown objects, we explore unsupervised training strategies to generate pseudo-labels for unknowns. Among the different approaches evaluated, our experiments show that top-5 auto-labelling offers more promising performance compared to simple resizing techniques.", "date": "2024-10-31", "categories": []} {"id": "2509.07507", "title": "MVAT: Multi-View Aware Teacher for Weakly Supervised 3D Object Detection", "abstract": "Annotating 3D data remains a costly bottleneck for 3D object detection, motivating the development of weakly supervised annotation methods that rely on more accessible 2D box annotations. However, relying solely on 2D boxes introduces projection ambiguities since a single 2D box can correspond to multiple valid 3D poses. Furthermore, partial object visibility under a single viewpoint setting makes accurate 3D box estimation difficult. We propose MVAT, a novel framework that leverages temporal multi-view present in sequential data to address these challenges. Our approach aggregates object-centric point clouds across time to build 3D object representations as dense and complete as possible. A Teacher-Student distillation paradigm is employed: The Teacher network learns from single viewpoints but targets are derived from temporally aggregated static objects. Then the Teacher generates high quality pseudo-labels that the Student learns to predict from a single viewpoint for both static and moving objects. The whole framework incorporates a multi-view 2D projection loss to enforce consistency between predicted 3D boxes and all available 2D annotations. Experiments on the nuScenes and Waymo Open datasets demonstrate that MVAT achieves state-of-the-art performance for weakly supervised 3D object detection, significantly narrowing the gap with fully supervised methods without requiring any 3D box annotations. % \\footnote{Code available upon acceptance} Our code is available in our public repository (\\href{https://github.com/CEA-LIST/MVAT}{code}).", "date": "2025-09-09", "categories": []} {"id": "2509.05780", "title": "3DPillars: Pillar-based two-stage 3D object detection", "abstract": "PointPillars is the fastest 3D object detector that exploits pseudo image representations to encode features for 3D objects in a scene. Albeit efficient, PointPillars is typically outperformed by state-of-the-art 3D detection methods due to the following limitations: 1) The pseudo image representations fail to preserve precise 3D structures, and 2) they make it difficult to adopt a two-stage detection pipeline using 3D object proposals that typically shows better performance than a single-stage approach. We introduce in this paper the first two-stage 3D detection framework exploiting pseudo image representations, narrowing the performance gaps between PointPillars and state-of-the-art methods, while retaining its efficiency. Our framework consists of two novel components that overcome the aforementioned limitations of PointPillars: First, we introduce a new CNN architecture, dubbed 3DPillars, that enables learning 3D voxel-based features from the pseudo image representation efficiently using 2D convolutions. The basic idea behind 3DPillars is that 3D features from voxels can be viewed as a stack of pseudo images. To implement this idea, we propose a separable voxel feature module that extracts voxel-based features without using 3D convolutions. Second, we introduce an RoI head with a sparse scene context feature module that aggregates multi-scale features from 3DPillars to obtain a sparse scene feature. This enables adopting a two-stage pipeline effectively, and fully leveraging contextual information of a scene to refine 3D object proposals. Experimental results on the KITTI and Waymo Open datasets demonstrate the effectiveness and efficiency of our approach, achieving a good compromise in terms of speed and accuracy.", "date": "2025-09-06", "categories": []} {"id": "2509.05785", "title": "CRAB: Camera-Radar Fusion for Reducing Depth Ambiguity in Backward Projection based View Transformation", "abstract": "Recently, camera-radar fusion-based 3D object detection methods in bird's eye view (BEV) have gained attention due to the complementary characteristics and cost-effectiveness of these sensors. Previous approaches using forward projection struggle with sparse BEV feature generation, while those employing backward projection overlook depth ambiguity, leading to false positives. In this paper, to address the aforementioned limitations, we propose a novel camera-radar fusion-based 3D object detection and segmentation model named CRAB (Camera-Radar fusion for reducing depth Ambiguity in Backward projection-based view transformation), using a backward projection that leverages radar to mitigate depth ambiguity. During the view transformation, CRAB aggregates perspective view image context features into BEV queries. It improves depth distinction among queries along the same ray by combining the dense but unreliable depth distribution from images with the sparse yet precise depth information from radar occupancy. We further introduce spatial cross-attention with a feature map containing radar context information to enhance the comprehension of the 3D scene. When evaluated on the nuScenes open dataset, our proposed approach achieves a state-of-the-art performance among backward projection-based camera-radar fusion methods with 62.4\\% NDS and 54.0\\% mAP in 3D object detection.", "date": "2025-09-06", "categories": []} {"id": "2509.05954", "title": "StripDet: Strip Attention-Based Lightweight 3D Object Detection from Point Cloud", "abstract": "The deployment of high-accuracy 3D object detection models from point cloud remains a significant challenge due to their substantial computational and memory requirements. To address this, we introduce StripDet, a novel lightweight framework designed for on-device efficiency. First, we propose the novel Strip Attention Block (SAB), a highly efficient module designed to capture long-range spatial dependencies. By decomposing standard 2D convolutions into asymmetric strip convolutions, SAB efficiently extracts directional features while reducing computational complexity from quadratic to linear. Second, we design a hardware-friendly hierarchical backbone that integrates SAB with depthwise separable convolutions and a simple multiscale fusion strategy, achieving end-to-end efficiency. Extensive experiments on the KITTI dataset validate StripDet's superiority. With only 0.65M parameters, our model achieves a 79.97% mAP for car detection, surpassing the baseline PointPillars with a 7x parameter reduction. Furthermore, StripDet outperforms recent lightweight and knowledge distillation-based methods, achieving a superior accuracy-efficiency trade-off while establishing itself as a practical solution for real-world 3D detection on edge devices.", "date": "2025-09-07", "categories": []} {"id": "2509.05999", "title": "S-LAM3D: Segmentation-Guided Monocular 3D Object Detection via Feature Space Fusion", "abstract": "Monocular 3D Object Detection represents a challenging Computer Vision task due to the nature of the input used, which is a single 2D image, lacking in any depth cues and placing the depth estimation problem as an ill-posed one. Existing solutions leverage the information extracted from the input by using Convolutional Neural Networks or Transformer architectures as feature extraction backbones, followed by specific detection heads for 3D parameters prediction. In this paper, we introduce a decoupled strategy based on injecting precomputed segmentation information priors and fusing them directly into the feature space for guiding the detection, without expanding the detection model or jointly learning the priors. The focus is on evaluating the impact of additional segmentation information on existing detection pipelines without adding additional prediction branches. The proposed method is evaluated on the KITTI 3D Object Detection Benchmark, outperforming the equivalent architecture that relies only on RGB image features for small objects in the scene: pedestrians and cyclists, and proving that understanding the input data can balance the need for additional sensors or training data.", "date": "2025-09-07", "categories": []} {"id": "2507.23567", "title": "3D-MOOD: Lifting 2D to 3D for Monocular Open-Set Object Detection", "abstract": "Monocular 3D object detection is valuable for various applications such as robotics and AR/VR. Existing methods are confined to closed-set settings, where the training and testing sets consist of the same scenes and/or object categories. However, real-world applications often introduce new environments and novel object categories, posing a challenge to these methods. In this paper, we address monocular 3D object detection in an open-set setting and introduce the first end-to-end 3D Monocular Open-set Object Detector (3D-MOOD). We propose to lift the open-set 2D detection into 3D space through our designed 3D bounding box head, enabling end-to-end joint training for both 2D and 3D tasks to yield better overall performance. We condition the object queries with geometry prior and overcome the generalization for 3D estimation across diverse scenes. To further improve performance, we design the canonical image space for more efficient cross-dataset training. We evaluate 3D-MOOD on both closed-set settings (Omni3D) and open-set settings (Omni3D to Argoverse 2, ScanNet), and achieve new state-of-the-art results. Code and models are available at royyang0714.github.io/3D-MOOD.", "date": "2025-07-31", "categories": []} {"id": "2509.04711", "title": "Domain Adaptation for Different Sensor Configurations in 3D Object Detection", "abstract": "Recent advances in autonomous driving have underscored the importance of accurate 3D object detection, with LiDAR playing a central role due to its robustness under diverse visibility conditions. However, different vehicle platforms often deploy distinct sensor configurations, causing performance degradation when models trained on one configuration are applied to another because of shifts in the point cloud distribution. Prior work on multi-dataset training and domain adaptation for 3D object detection has largely addressed environmental domain gaps and density variation within a single LiDAR; in contrast, the domain gap for different sensor configurations remains largely unexplored. In this work, we address domain adaptation across different sensor configurations in 3D object detection. We propose two techniques: Downstream Fine-tuning (dataset-specific fine-tuning after multi-dataset training) and Partial Layer Fine-tuning (updating only a subset of layers to improve cross-configuration generalization). Using paired datasets collected in the same geographic region with multiple sensor configurations, we show that joint training with Downstream Fine-tuning and Partial Layer Fine-tuning consistently outperforms naive joint training for each configuration. Our findings provide a practical and scalable solution for adapting 3D object detection models to the diverse vehicle platforms.", "date": "2025-09-04", "categories": []} {"id": "2301.05711", "title": "OA-DET3D: Embedding Object Awareness as a General Plug-in for Multi-Camera 3D Object Detection", "abstract": "The recent advance in multi-camera 3D object detection is featured by bird's-eye view (BEV) representation or object queries. However, the ill-posed transformation from image-plane view to 3D space inevitably causes feature clutter and distortion, making the objects blur into the background. To this end, we explore how to incorporate supplementary cues for differentiating objects in the transformed feature representation. Formally, we introduce OA-DET3D, a general plug-in module that improves 3D object detection by bringing object awareness into a variety of existing 3D object detection pipelines. Specifically, OA-DET3D boosts the representation of objects by leveraging object-centric depth information and foreground pseudo points. First, we use object-level supervision from the properties of each 3D bounding box to guide the network in learning the depth distribution. Next, we select foreground pixels using a 2D object detector and project them into 3D space for pseudo-voxel feature encoding. Finally, the object-aware depth features and pseudo-voxel features are incorporated into the BEV representation or query feature from the baseline model with a deformable attention mechanism. We conduct extensive experiments on the nuScenes dataset and Argoverse 2 dataset to validate the merits of OA-DET3D. Our method achieves consistent improvements over the BEV-based baselines in terms of both average precision and comprehensive detection score.", "date": "2023-01-13", "categories": []} {"id": "2504.18325", "title": "Depth3DLane: Monocular 3D Lane Detection via Depth Prior Distillation", "abstract": "Monocular 3D lane detection is challenging due to the difficulty in capturing depth information from single-camera images. A common strategy involves transforming front-view (FV) images into bird's-eye-view (BEV) space through inverse perspective mapping (IPM), facilitating lane detection using BEV features. However, IPM's flat-ground assumption and loss of contextual information lead to inaccuracies in reconstructing 3D information, especially height. In this paper, we introduce a BEV-based framework to address these limitations and improve 3D lane detection accuracy. Our approach incorporates a Hierarchical Depth-Aware Head that provides multi-scale depth features, mitigating the flat-ground assumption by enhancing spatial awareness across varying depths. Additionally, we leverage Depth Prior Distillation to transfer semantic depth knowledge from a teacher model, capturing richer structural and contextual information for complex lane structures. To further refine lane continuity and ensure smooth lane reconstruction, we introduce a Conditional Random Field module that enforces spatial coherence in lane predictions. Extensive experiments validate that our method achieves state-of-the-art performance in terms of z-axis error and outperforms other methods in the field in overall performance. The code is released at: https://anonymous.4open.science/r/Depth3DLane-DCDD.", "date": "2025-04-25", "categories": []} {"id": "2508.17439", "title": "Investigating Domain Gaps for Indoor 3D Object Detection", "abstract": "As a fundamental task for indoor scene understanding, 3D object detection has been extensively studied, and the accuracy on indoor point cloud data has been substantially improved. However, existing researches have been conducted on limited datasets, where the training and testing sets share the same distribution. In this paper, we consider the task of adapting indoor 3D object detectors from one dataset to another, presenting a comprehensive benchmark with ScanNet, SUN RGB-D and 3D Front datasets, as well as our newly proposed large-scale datasets ProcTHOR-OD and ProcFront generated by a 3D simulator. Since indoor point cloud datasets are collected and constructed in different ways, the object detectors are likely to overfit to specific factors within each dataset, such as point cloud quality, bounding box layout and instance features. We conduct experiments across datasets on different adaptation scenarios including synthetic-to-real adaptation, point cloud quality adaptation, layout adaptation and instance feature adaptation, analyzing the impact of different domain gaps on 3D object detectors. We also introduce several approaches to improve adaptation performances, providing baselines for domain adaptive indoor 3D object detection, hoping that future works may propose detectors with stronger generalization ability across domains. Our project homepage can be found in https://jeremyzhao1998.github.io/DAVoteNet-release/.", "date": "2025-08-24", "categories": []} {"id": "2508.20488", "title": "Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts", "abstract": "Accurate monocular 3D object detection (M3OD) is pivotal for safety-critical applications like autonomous driving, yet its reliability deteriorates significantly under real-world domain shifts caused by environmental or sensor variations. To address these shifts, Test-Time Adaptation (TTA) methods have emerged, enabling models to adapt to target distributions during inference. While prior TTA approaches recognize the positive correlation between low uncertainty and high generalization ability, they fail to address the dual uncertainty inherent to M3OD: semantic uncertainty (ambiguous class predictions) and geometric uncertainty (unstable spatial localization). To bridge this gap, we propose Dual Uncertainty Optimization (DUO), the first TTA framework designed to jointly minimize both uncertainties for robust M3OD. Through a convex optimization lens, we introduce an innovative convex structure of the focal loss and further derive a novel unsupervised version, enabling label-agnostic uncertainty weighting and balanced learning for high-uncertainty objects. In parallel, we design a semantic-aware normal field constraint that preserves geometric coherence in regions with clear semantic cues, reducing uncertainty from the unstable 3D representation. This dual-branch mechanism forms a complementary loop: enhanced spatial perception improves semantic classification, and robust semantic predictions further refine spatial understanding. Extensive experiments demonstrate the superiority of DUO over existing methods across various datasets and domain shift types.", "date": "2025-08-28", "categories": []} {"id": "2508.20530", "title": "Enhancing Pseudo-Boxes via Data-Level LiDAR-Camera Fusion for Unsupervised 3D Object Detection", "abstract": "Existing LiDAR-based 3D object detectors typically rely on manually annotated labels for training to achieve good performance. However, obtaining high-quality 3D labels is time-consuming and labor-intensive. To address this issue, recent works explore unsupervised 3D object detection by introducing RGB images as an auxiliary modal to assist pseudo-box generation. However, these methods simply integrate pseudo-boxes generated by LiDAR point clouds and RGB images. Yet, such a label-level fusion strategy brings limited improvements to the quality of pseudo-boxes, as it overlooks the complementary nature in terms of LiDAR and RGB image data. To overcome the above limitations, we propose a novel data-level fusion framework that integrates RGB images and LiDAR data at an early stage. Specifically, we utilize vision foundation models for instance segmentation and depth estimation on images and introduce a bi-directional fusion method, where real points acquire category labels from the 2D space, while 2D pixels are projected onto 3D to enhance real point density. To mitigate noise from depth and segmentation estimations, we propose a local and global filtering method, which applies local radius filtering to suppress depth estimation errors and global statistical filtering to remove segmentation-induced outliers. Furthermore, we propose a data-level fusion based dynamic self-evolution strategy, which iteratively refines pseudo-boxes under a dense representation, significantly improving localization accuracy. Extensive experiments on the nuScenes dataset demonstrate that the detector trained by our method significantly outperforms that trained by previous state-of-the-art methods with 28.4$\\%$ mAP on the nuScenes validation benchmark.", "date": "2025-08-28", "categories": []} {"id": "2506.14243", "title": "Cross-Modal Geometric Hierarchy Fusion: An Implicit-Submap Driven Framework for Resilient 3D Place Recognition", "abstract": "LiDAR-based place recognition serves as a crucial enabler for long-term autonomy in robotics and autonomous driving systems. Yet, prevailing methodologies relying on handcrafted feature extraction face dual challenges: (1) Inconsistent point cloud density, induced by ego-motion dynamics and environmental disturbances during repeated traversals, leads to descriptor instability, and (2) Representation fragility stems from reliance on single-level geometric abstractions that lack discriminative power in structurally complex scenarios. To address these limitations, we propose a novel framework that redefines 3D place recognition through density-agnostic geometric reasoning. Specifically, we introduce an implicit 3D representation based on elastic points, which is immune to the interference of original scene point cloud density and achieves the characteristic of uniform distribution. Subsequently, we derive the occupancy grid and normal vector information of the scene from this implicit representation. Finally, with the aid of these two types of information, we obtain descriptors that fuse geometric information from both bird's-eye view (capturing macro-level spatial layouts) and 3D segment (encoding micro-scale surface geometries) perspectives. We conducted extensive experiments on numerous datasets (KITTI, KITTI-360, MulRan, NCLT) across diverse environments. The experimental results demonstrate that our method achieves state-of-the-art performance. Moreover, our approach strikes an optimal balance between accuracy, runtime, and memory optimization for historical maps, showcasing excellent Resilient and scalability. Our code will be open-sourced in the future.", "date": "2025-06-17", "categories": []} {"id": "2507.16224", "title": "LDRFusion: A LiDAR-Dominant multimodal refinement framework for 3D object detection", "abstract": "Existing LiDAR-Camera fusion methods have achieved strong results in 3D object detection. To address the sparsity of point clouds, previous approaches typically construct spatial pseudo point clouds via depth completion as auxiliary input and adopts a proposal-refinement framework to generate detection results. However, introducing pseudo points inevitably brings noise, potentially resulting in inaccurate predictions. Considering the differing roles and reliability levels of each modality, we propose LDRFusion, a novel Lidar-dominant two-stage refinement framework for multi-sensor fusion. The first stage soley relies on LiDAR to produce accurately localized proposals, followed by a second stage where pseudo point clouds are incorporated to detect challenging instances. The instance-level results from both stages are subsequently merged. To further enhance the representation of local structures in pseudo point clouds, we present a hierarchical pseudo point residual encoding module, which encodes neighborhood sets using both feature and positional residuals. Experiments on the KITTI dataset demonstrate that our framework consistently achieves strong performance across multiple categories and difficulty levels.", "date": "2025-07-22", "categories": []} {"id": "2508.03252", "title": "Robust Single-Stage Fully Sparse 3D Object Detection via Detachable Latent Diffusion", "abstract": "Denoising Diffusion Probabilistic Models (DDPMs) have shown success in robust 3D object detection tasks. Existing methods often rely on the score matching from 3D boxes or pre-trained diffusion priors. However, they typically require multi-step iterations in inference, which limits efficiency. To address this, we propose a Robust single-stage fully Sparse 3D object Detection Network with a Detachable Latent Framework (DLF) of DDPMs, named RSDNet. Specifically, RSDNet learns the denoising process in latent feature spaces through lightweight denoising networks like multi-level denoising autoencoders (DAEs). This enables RSDNet to effectively understand scene distributions under multi-level perturbations, achieving robust and reliable detection. Meanwhile, we reformulate the noising and denoising mechanisms of DDPMs, enabling DLF to construct multi-type and multi-level noise samples and targets, enhancing RSDNet robustness to multiple perturbations. Furthermore, a semantic-geometric conditional guidance is introduced to perceive the object boundaries and shapes, alleviating the center feature missing problem in sparse representations, enabling RSDNet to perform in a fully sparse detection pipeline. Moreover, the detachable denoising network design of DLF enables RSDNet to perform single-step detection in inference, further enhancing detection efficiency. Extensive experiments on public benchmarks show that RSDNet can outperform existing methods, achieving state-of-the-art detection.", "date": "2025-08-05", "categories": []} {"id": "2508.19593", "title": "Generalizing Monocular 3D Object Detection", "abstract": "Monocular 3D object detection (Mono3D) is a fundamental computer vision task that estimates an object's class, 3D position, dimensions, and orientation from a single image. Its applications, including autonomous driving, augmented reality, and robotics, critically rely on accurate 3D environmental understanding. This thesis addresses the challenge of generalizing Mono3D models to diverse scenarios, including occlusions, datasets, object sizes, and camera parameters. To enhance occlusion robustness, we propose a mathematically differentiable NMS (GrooMeD-NMS). To improve generalization to new datasets, we explore depth equivariant (DEVIANT) backbones. We address the issue of large object detection, demonstrating that it's not solely a data imbalance or receptive field problem but also a noise sensitivity issue. To mitigate this, we introduce a segmentation-based approach in bird's-eye view with dice loss (SeaBird). Finally, we mathematically analyze the extrapolation of Mono3D models to unseen camera heights and improve Mono3D generalization in such out-of-distribution settings.", "date": "2025-08-27", "categories": []} {"id": "2508.20063", "title": "OpenM3D: Open Vocabulary Multi-view Indoor 3D Object Detection without Human Annotations", "abstract": "Open-vocabulary (OV) 3D object detection is an emerging field, yet its exploration through image-based methods remains limited compared to 3D point cloud-based methods. We introduce OpenM3D, a novel open-vocabulary multi-view indoor 3D object detector trained without human annotations. In particular, OpenM3D is a single-stage detector adapting the 2D-induced voxel features from the ImGeoNet model. To support OV, it is jointly trained with a class-agnostic 3D localization loss requiring high-quality 3D pseudo boxes and a voxel-semantic alignment loss requiring diverse pre-trained CLIP features. We follow the training setting of OV-3DET where posed RGB-D images are given but no human annotations of 3D boxes or classes are available. We propose a 3D Pseudo Box Generation method using a graph embedding technique that combines 2D segments into coherent 3D structures. Our pseudo-boxes achieve higher precision and recall than other methods, including the method proposed in OV-3DET. We further sample diverse CLIP features from 2D segments associated with each coherent 3D structure to align with the corresponding voxel feature. The key to training a highly accurate single-stage detector requires both losses to be learned toward high-quality targets. At inference, OpenM3D, a highly efficient detector, requires only multi-view images for input and demonstrates superior accuracy and speed (0.3 sec. per scene) on ScanNet200 and ARKitScenes indoor benchmarks compared to existing methods. We outperform a strong two-stage method that leverages our class-agnostic detector with a ViT CLIP-based OV classifier and a baseline incorporating multi-view depth estimator on both accuracy and speed.", "date": "2025-08-27", "categories": []} {"id": "2508.18293", "title": "Towards Training-Free Underwater 3D Object Detection from Sonar Point Clouds: A Comparison of Traditional and Deep Learning Approaches", "abstract": "Underwater 3D object detection remains one of the most challenging frontiers in computer vision, where traditional approaches struggle with the harsh acoustic environment and scarcity of training data. While deep learning has revolutionized terrestrial 3D detection, its application underwater faces a critical bottleneck: obtaining sufficient annotated sonar data is prohibitively expensive and logistically complex, often requiring specialized vessels, expert surveyors, and favorable weather conditions. This work addresses a fundamental question: Can we achieve reliable underwater 3D object detection without real-world training data? We tackle this challenge by developing and comparing two paradigms for training-free detection of artificial structures in multibeam echo-sounder point clouds. Our dual approach combines a physics-based sonar simulation pipeline that generates synthetic training data for state-of-the-art neural networks, with a robust model-based template matching system that leverages geometric priors of target objects. Evaluation on real bathymetry surveys from the Baltic Sea reveals surprising insights: while neural networks trained on synthetic data achieve 98% mean Average Precision (mAP) on simulated scenes, they drop to 40% mAP on real sonar data due to domain shift. Conversely, our template matching approach maintains 83% mAP on real data without requiring any training, demonstrating remarkable robustness to acoustic noise and environmental variations. Our findings challenge conventional wisdom about data-hungry deep learning in underwater domains and establish the first large-scale benchmark for training-free underwater 3D detection. This work opens new possibilities for autonomous underwater vehicle navigation, marine archaeology, and offshore infrastructure monitoring in data-scarce environments where traditional machine learning approaches fail.", "date": "2025-08-22", "categories": []} {"id": "2411.08402", "title": "V2X-R: Cooperative LiDAR-4D Radar Fusion with Denoising Diffusion for 3D Object Detection", "abstract": "Current Vehicle-to-Everything (V2X) systems have significantly enhanced 3D object detection using LiDAR and camera data. However, these methods suffer from performance degradation in adverse weather conditions. The weather-robust 4D radar provides Doppler and additional geometric information, raising the possibility of addressing this challenge. To this end, we present V2X-R, the first simulated V2X dataset incorporating LiDAR, camera, and 4D radar. V2X-R contains 12,079 scenarios with 37,727 frames of LiDAR and 4D radar point clouds, 150,908 images, and 170,859 annotated 3D vehicle bounding boxes. Subsequently, we propose a novel cooperative LiDAR-4D radar fusion pipeline for 3D object detection and implement it with various fusion strategies. To achieve weather-robust detection, we additionally propose a Multi-modal Denoising Diffusion (MDD) module in our fusion pipeline. MDD utilizes weather-robust 4D radar feature as a condition to prompt the diffusion model to denoise noisy LiDAR features. Experiments show that our LiDAR-4D radar fusion pipeline demonstrates superior performance in the V2X-R dataset. Over and above this, our MDD module further improved the performance of basic fusion model by up to 5.73%/6.70% in foggy/snowy conditions with barely disrupting normal performance. The dataset and code will be publicly available at: https://github.com/ylwhxht/V2X-R.", "date": "2024-11-13", "categories": []} {"id": "2506.15610", "title": "BoxFusion: Reconstruction-Free Open-Vocabulary 3D Object Detection via Real-Time Multi-View Box Fusion", "abstract": "Open-vocabulary 3D object detection has gained significant interest due to its critical applications in autonomous driving and embodied AI. Existing detection methods, whether offline or online, typically rely on dense point cloud reconstruction, which imposes substantial computational overhead and memory constraints, hindering real-time deployment in downstream tasks. To address this, we propose a novel reconstruction-free online framework tailored for memory-efficient and real-time 3D detection. Specifically, given streaming posed RGB-D video input, we leverage Cubify Anything as a pre-trained visual foundation model (VFM) for single-view 3D object detection by bounding boxes, coupled with CLIP to capture open-vocabulary semantics of detected objects. To fuse all detected bounding boxes across different views into a unified one, we employ an association module for correspondences of multi-views and an optimization module to fuse the 3D bounding boxes of the same instance predicted in multi-views. The association module utilizes 3D Non-Maximum Suppression (NMS) and a box correspondence matching module, while the optimization module uses an IoU-guided efficient random optimization technique based on particle filtering to enforce multi-view consistency of the 3D bounding boxes while minimizing computational complexity. Extensive experiments on ScanNetV2 and CA-1M datasets demonstrate that our method achieves state-of-the-art performance among online methods. Benefiting from this novel reconstruction-free paradigm for 3D object detection, our method exhibits great generalization abilities in various scenarios, enabling real-time perception even in environments exceeding 1000 square meters.", "date": "2025-06-18", "categories": []} {"id": "2508.16812", "title": "Towards Open-Vocabulary Multimodal 3D Object Detection with Attributes", "abstract": "3D object detection plays a crucial role in autonomous systems, yet existing methods are limited by closed-set assumptions and struggle to recognize novel objects and their attributes in real-world scenarios. We propose OVODA, a novel framework enabling both open-vocabulary 3D object and attribute detection with no need to know the novel class anchor size. OVODA uses foundation models to bridge the semantic gap between 3D features and texts while jointly detecting attributes, e.g., spatial relationships, motion states, etc. To facilitate such research direction, we propose OVAD, a new dataset that supplements existing 3D object detection benchmarks with comprehensive attribute annotations. OVODA incorporates several key innovations, including foundation model feature concatenation, prompt tuning strategies, and specialized techniques for attribute detection, including perspective-specified prompts and horizontal flip augmentation. Our results on both the nuScenes and Argoverse 2 datasets show that under the condition of no given anchor sizes of novel classes, OVODA outperforms the state-of-the-art methods in open-vocabulary 3D object detection while successfully recognizing object attributes. Our OVAD dataset is released here: https://doi.org/10.5281/zenodo.16904069 .", "date": "2025-08-22", "categories": []} {"id": "2508.16408", "title": "SAMFusion: Sensor-Adaptive Multimodal Fusion for 3D Object Detection in Adverse Weather", "abstract": "Multimodal sensor fusion is an essential capability for autonomous robots, enabling object detection and decision-making in the presence of failing or uncertain inputs. While recent fusion methods excel in normal environmental conditions, these approaches fail in adverse weather, e.g., heavy fog, snow, or obstructions due to soiling. We introduce a novel multi-sensor fusion approach tailored to adverse weather conditions. In addition to fusing RGB and LiDAR sensors, which are employed in recent autonomous driving literature, our sensor fusion stack is also capable of learning from NIR gated camera and radar modalities to tackle low light and inclement weather. We fuse multimodal sensor data through attentive, depth-based blending schemes, with learned refinement on the Bird's Eye View (BEV) plane to combine image and range features effectively. Our detections are predicted by a transformer decoder that weighs modalities based on distance and visibility. We demonstrate that our method improves the reliability of multimodal sensor fusion in autonomous vehicles under challenging weather conditions, bridging the gap between ideal conditions and real-world edge cases. Our approach improves average precision by 17.2 AP compared to the next best method for vulnerable pedestrians in long distances and challenging foggy scenes. Our project page is available at https://light.princeton.edu/samfusion/", "date": "2025-08-22", "categories": []} {"id": "2508.15353", "title": "RCDINO: Enhancing Radar-Camera 3D Object Detection with DINOv2 Semantic Features", "abstract": "Three-dimensional object detection is essential for autonomous driving and robotics, relying on effective fusion of multimodal data from cameras and radar. This work proposes RCDINO, a multimodal transformer-based model that enhances visual backbone features by fusing them with semantically rich representations from the pretrained DINOv2 foundation model. This approach enriches visual representations and improves the model's detection performance while preserving compatibility with the baseline architecture. Experiments on the nuScenes dataset demonstrate that RCDINO achieves state-of-the-art performance among radar-camera models, with 56.4 NDS and 48.1 mAP. Our implementation is available at https://github.com/OlgaMatykina/RCDINO.", "date": "2025-08-21", "categories": []} {"id": "2307.00347", "title": "Spatial-Temporal Graph Enhanced DETR Towards Multi-Frame 3D Object Detection", "abstract": "The Detection Transformer (DETR) has revolutionized the design of CNN-based object detection systems, showcasing impressive performance. However, its potential in the domain of multi-frame 3D object detection remains largely unexplored. In this paper, we present STEMD, a novel end-to-end framework that enhances the DETR-like paradigm for multi-frame 3D object detection by addressing three key aspects specifically tailored for this task. First, to model the inter-object spatial interaction and complex temporal dependencies, we introduce the spatial-temporal graph attention network, which represents queries as nodes in a graph and enables effective modeling of object interactions within a social context. To solve the problem of missing hard cases in the proposed output of the encoder in the current frame, we incorporate the output of the previous frame to initialize the query input of the decoder. Finally, it poses a challenge for the network to distinguish between the positive query and other highly similar queries that are not the best match. And similar queries are insufficiently suppressed and turn into redundant prediction boxes. To address this issue, our proposed IoU regularization term encourages similar queries to be distinct during the refinement. Through extensive experiments, we demonstrate the effectiveness of our approach in handling challenging scenarios, while incurring only a minor additional computational overhead. The code is publicly available at https://github.com/Eaphan/STEMD.", "date": "2023-07-01", "categories": []} {"id": "2405.16873", "title": "ContrastAlign: Toward Robust BEV Feature Alignment via Contrastive Learning for Multi-Modal 3D Object Detection", "abstract": "In the field of 3D object detection tasks, fusing heterogeneous features from LiDAR and camera sensors into a unified Bird's Eye View (BEV) representation is a widely adopted paradigm. However, existing methods often suffer from imprecise sensor calibration, leading to feature misalignment in LiDAR-camera BEV fusion. Moreover, such inaccuracies cause errors in depth estimation for the camera branch, aggravating misalignment between LiDAR and camera BEV features. In this work, we propose a novel ContrastAlign approach that utilizes contrastive learning to enhance the alignment of heterogeneous modalities, thereby improving the robustness of the fusion process. Specifically, our approach comprises three key components: (1) the L-Instance module, which extracts LiDAR instance features within the LiDAR BEV features; (2) the C-Instance module, which predicts camera instance features through Region of Interest (RoI) pooling on the camera BEV features; (3) the InstanceFusion module, which employs contrastive learning to generate consistent instance features across heFterogeneous modalities. Subsequently, we use graph matching to calculate the similarity between the neighboring camera instance features and the similarity instance features to complete the alignment of instance features. Our method achieves SOTA performance, with an mAP of 71.5%, surpassing GraphBEV by 1.4% on the nuScenes val set. Importantly, our method excels BEVFusion under conditions with spatial & temporal misalignment noise, improving mAP by 1.4% and 11.1% on nuScenes dataset. Notably, on the Argoverse2 dataset, ContrastAlign outperforms GraphBEV by 1.0% in mAP, indicating that the farther the distance, the more severe the feature misalignment and the more effective.", "date": "2024-05-27", "categories": []} {"id": "2508.10935", "title": "HQ-OV3D: A High Box Quality Open-World 3D Detection Framework based on Diffision Model", "abstract": "Traditional closed-set 3D detection frameworks fail to meet the demands of open-world applications like autonomous driving. Existing open-vocabulary 3D detection methods typically adopt a two-stage pipeline consisting of pseudo-label generation followed by semantic alignment. While vision-language models (VLMs) recently have dramatically improved the semantic accuracy of pseudo-labels, their geometric quality, particularly bounding box precision, remains commonly neglected. To address this issue, we propose a High Box Quality Open-Vocabulary 3D Detection (HQ-OV3D) framework, dedicated to generate and refine high-quality pseudo-labels for open-vocabulary classes. The framework comprises two key components: an Intra-Modality Cross-Validated (IMCV) Proposal Generator that utilizes cross-modality geometric consistency to generate high-quality initial 3D proposals, and an Annotated-Class Assisted (ACA) Denoiser that progressively refines 3D proposals by leveraging geometric priors from annotated categories through a DDIM-based denoising mechanism. Compared to the state-of-the-art method, training with pseudo-labels generated by our approach achieves a 7.37% improvement in mAP on novel classes, demonstrating the superior quality of the pseudo-labels produced by our framework. HQ-OV3D can serve not only as a strong standalone open-vocabulary 3D detector but also as a plug-in high-quality pseudo-label generator for existing open-vocabulary detection or annotation pipelines.", "date": "2025-08-12", "categories": []} {"id": "2508.11951", "title": "Transferable Class Statistics and Multi-scale Feature Approximation for 3D Object Detection", "abstract": "This paper investigates multi-scale feature approximation and transferable features for object detection from point clouds. Multi-scale features are critical for object detection from point clouds. However, multi-scale feature learning usually involves multiple neighborhood searches and scale-aware layers, which can hinder efforts to achieve lightweight models and may not be conducive to research constrained by limited computational resources. This paper approximates point-based multi-scale features from a single neighborhood based on knowledge distillation. To compensate for the loss of constructive diversity in a single neighborhood, this paper designs a transferable feature embedding mechanism. Specifically, class-aware statistics are employed as transferable features given the small computational cost. In addition, this paper introduces the central weighted intersection over union for localization to alleviate the misalignment brought by the center offset in optimization. Note that the method presented in this paper saves computational costs. Extensive experiments on public datasets demonstrate the effectiveness of the proposed method.", "date": "2025-08-16", "categories": []} {"id": "2508.12330", "title": "DoppDrive: Doppler-Driven Temporal Aggregation for Improved Radar Object Detection", "abstract": "Radar-based object detection is essential for autonomous driving due to radar's long detection range. However, the sparsity of radar point clouds, especially at long range, poses challenges for accurate detection. Existing methods increase point density through temporal aggregation with ego-motion compensation, but this approach introduces scatter from dynamic objects, degrading detection performance. We propose DoppDrive, a novel Doppler-Driven temporal aggregation method that enhances radar point cloud density while minimizing scatter. Points from previous frames are shifted radially according to their dynamic Doppler component to eliminate radial scatter, with each point assigned a unique aggregation duration based on its Doppler and angle to minimize tangential scatter. DoppDrive is a point cloud density enhancement step applied before detection, compatible with any detector, and we demonstrate that it significantly improves object detection performance across various detectors and datasets.", "date": "2025-08-17", "categories": []} {"id": "2508.12684", "title": "Refine-and-Contrast: Adaptive Instance-Aware BEV Representations for Multi-UAV Collaborative Object Detection", "abstract": "Multi-UAV collaborative 3D detection enables accurate and robust perception by fusing multi-view observations from aerial platforms, offering significant advantages in coverage and occlusion handling, while posing new challenges for computation on resource-constrained UAV platforms. In this paper, we present AdaBEV, a novel framework that learns adaptive instance-aware BEV representations through a refine-and-contrast paradigm. Unlike existing methods that treat all BEV grids equally, AdaBEV introduces a Box-Guided Refinement Module (BG-RM) and an Instance-Background Contrastive Learning (IBCL) to enhance semantic awareness and feature discriminability. BG-RM refines only BEV grids associated with foreground instances using 2D supervision and spatial subdivision, while IBCL promotes stronger separation between foreground and background features via contrastive learning in BEV space. Extensive experiments on the Air-Co-Pred dataset demonstrate that AdaBEV achieves superior accuracy-computation trade-offs across model scales, outperforming other state-of-the-art methods at low resolutions and approaching upper bound performance while maintaining low-resolution BEV inputs and negligible overhead.", "date": "2025-08-18", "categories": []} {"id": "2508.12695", "title": "Neural Rendering for Sensor Adaptation in 3D Object Detection", "abstract": "Autonomous vehicles often have varying camera sensor setups, which is inevitable due to restricted placement options for different vehicle types. Training a perception model on one particular setup and evaluating it on a new, different sensor setup reveals the so-called cross-sensor domain gap, typically leading to a degradation in accuracy. In this paper, we investigate the impact of the cross-sensor domain gap on state-of-the-art 3D object detectors. To this end, we introduce CamShift, a dataset inspired by nuScenes and created in CARLA to specifically simulate the domain gap between subcompact vehicles and sport utility vehicles (SUVs). Using CamShift, we demonstrate significant cross-sensor performance degradation, identify robustness dependencies on model architecture, and propose a data-driven solution to mitigate the effect. On the one hand, we show that model architectures based on a dense Bird's Eye View (BEV) representation with backward projection, such as BEVFormer, are the most robust against varying sensor configurations. On the other hand, we propose a novel data-driven sensor adaptation pipeline based on neural rendering, which can transform entire datasets to match different camera sensor setups. Applying this approach improves performance across all investigated 3D object detectors, mitigating the cross-sensor domain gap by a large margin and reducing the need for new data collection by enabling efficient data reusability across vehicles with different sensor setups. The CamShift dataset and the sensor adaptation benchmark are available at https://dmholtz.github.io/camshift/.", "date": "2025-08-18", "categories": []} {"id": "2508.12917", "title": "CMF-IoU: Multi-Stage Cross-Modal Fusion 3D Object Detection with IoU Joint Prediction", "abstract": "Multi-modal methods based on camera and LiDAR sensors have garnered significant attention in the field of 3D detection. However, many prevalent works focus on single or partial stage fusion, leading to insufficient feature extraction and suboptimal performance. In this paper, we introduce a multi-stage cross-modal fusion 3D detection framework, termed CMF-IOU, to effectively address the challenge of aligning 3D spatial and 2D semantic information. Specifically, we first project the pixel information into 3D space via a depth completion network to get the pseudo points, which unifies the representation of the LiDAR and camera information. Then, a bilateral cross-view enhancement 3D backbone is designed to encode LiDAR points and pseudo points. The first sparse-to-distant (S2D) branch utilizes an encoder-decoder structure to reinforce the representation of sparse LiDAR points. The second residual view consistency (ResVC) branch is proposed to mitigate the influence of inaccurate pseudo points via both the 3D and 2D convolution processes. Subsequently, we introduce an iterative voxel-point aware fine grained pooling module, which captures the spatial information from LiDAR points and textural information from pseudo points in the proposal refinement stage. To achieve more precise refinement during iteration, an intersection over union (IoU) joint prediction branch integrated with a novel proposals generation technique is designed to preserve the bounding boxes with both high IoU and classification scores. Extensive experiments show the superior performance of our method on the KITTI, nuScenes and Waymo datasets.", "date": "2025-08-18", "categories": []} {"id": "2508.11185", "title": "CHARM3R: Towards Unseen Camera Height Robust Monocular 3D Detector", "abstract": "Monocular 3D object detectors, while effective on data from one ego camera height, struggle with unseen or out-of-distribution camera heights. Existing methods often rely on Plucker embeddings, image transformations or data augmentation. This paper takes a step towards this understudied problem by first investigating the impact of camera height variations on state-of-the-art (SoTA) Mono3D models. With a systematic analysis on the extended CARLA dataset with multiple camera heights, we observe that depth estimation is a primary factor influencing performance under height variations. We mathematically prove and also empirically observe consistent negative and positive trends in mean depth error of regressed and ground-based depth models, respectively, under camera height changes. To mitigate this, we propose Camera Height Robust Monocular 3D Detector (CHARM3R), which averages both depth estimates within the model. CHARM3R improves generalization to unseen camera heights by more than $45\\%$, achieving SoTA performance on the CARLA dataset. Codes and Models at https://github.com/abhi1kumar/CHARM3R", "date": "2025-08-15", "categories": []} {"id": "2504.19256", "title": "LM-MCVT: A Lightweight Multi-modal Multi-view Convolutional-Vision Transformer Approach for 3D Object Recognition", "abstract": "In human-centered environments such as restaurants, homes, and warehouses, robots often face challenges in accurately recognizing 3D objects. These challenges stem from the complexity and variability of these environments, including diverse object shapes. In this paper, we propose a novel Lightweight Multi-modal Multi-view Convolutional-Vision Transformer network (LM-MCVT) to enhance 3D object recognition in robotic applications. Our approach leverages the Globally Entropy-based Embeddings Fusion (GEEF) method to integrate multi-views efficiently. The LM-MCVT architecture incorporates pre- and mid-level convolutional encoders and local and global transformers to enhance feature extraction and recognition accuracy. We evaluate our method on the synthetic ModelNet40 dataset and achieve a recognition accuracy of 95.6% using a four-view setup, surpassing existing state-of-the-art methods. To further validate its effectiveness, we conduct 5-fold cross-validation on the real-world OmniObject3D dataset using the same configuration. Results consistently show superior performance, demonstrating the method's robustness in 3D object recognition across synthetic and real-world 3D data.", "date": "2025-04-27", "categories": []} {"id": "2508.07089", "title": "ForeSight: Multi-View Streaming Joint Object Detection and Trajectory Forecasting", "abstract": "We introduce ForeSight, a novel joint detection and forecasting framework for vision-based 3D perception in autonomous vehicles. Traditional approaches treat detection and forecasting as separate sequential tasks, limiting their ability to leverage temporal cues. ForeSight addresses this limitation with a multi-task streaming and bidirectional learning approach, allowing detection and forecasting to share query memory and propagate information seamlessly. The forecast-aware detection transformer enhances spatial reasoning by integrating trajectory predictions from a multiple hypothesis forecast memory queue, while the streaming forecast transformer improves temporal consistency using past forecasts and refined detections. Unlike tracking-based methods, ForeSight eliminates the need for explicit object association, reducing error propagation with a tracking-free model that efficiently scales across multi-frame sequences. Experiments on the nuScenes dataset show that ForeSight achieves state-of-the-art performance, achieving an EPA of 54.9%, surpassing previous methods by 9.3%, while also attaining the best mAP and minADE among multi-view detection and forecasting models.", "date": "2025-08-09", "categories": []} {"id": "2508.07092", "title": "Communication-Efficient Multi-Agent 3D Detection via Hybrid Collaboration", "abstract": "Collaborative 3D detection can substantially boost detection performance by allowing agents to exchange complementary information. It inherently results in a fundamental trade-off between detection performance and communication bandwidth. To tackle this bottleneck issue, we propose a novel hybrid collaboration that adaptively integrates two types of communication messages: perceptual outputs, which are compact, and raw observations, which offer richer information. This approach focuses on two key aspects: i) integrating complementary information from two message types and ii) prioritizing the most critical data within each type. By adaptively selecting the most critical set of messages, it ensures optimal perceptual information and adaptability, effectively meeting the demands of diverse communication scenarios.Building on this hybrid collaboration, we present \\texttt{HyComm}, a communication-efficient LiDAR-based collaborative 3D detection system. \\texttt{HyComm} boasts two main benefits: i) it facilitates adaptable compression rates for messages, addressing various communication requirements, and ii) it uses standardized data formats for messages. This ensures they are independent of specific detection models, fostering adaptability across different agent configurations. To evaluate HyComm, we conduct experiments on both real-world and simulation datasets: DAIR-V2X and OPV2V. HyComm consistently outperforms previous methods and achieves a superior performance-bandwidth trade-off regardless of whether agents use the same or varied detection models. It achieves a lower communication volume of more than 2,006$\\times$ and still outperforms Where2comm on DAIR-V2X in terms of AP50. The related code will be released.", "date": "2025-08-09", "categories": []} {"id": "2508.04702", "title": "BEVCon: Advancing Bird's Eye View Perception with Contrastive Learning", "abstract": "We present BEVCon, a simple yet effective contrastive learning framework designed to improve Bird's Eye View (BEV) perception in autonomous driving. BEV perception offers a top-down-view representation of the surrounding environment, making it crucial for 3D object detection, segmentation, and trajectory prediction tasks. While prior work has primarily focused on enhancing BEV encoders and task-specific heads, we address the underexplored potential of representation learning in BEV models. BEVCon introduces two contrastive learning modules: an instance feature contrast module for refining BEV features and a perspective view contrast module that enhances the image backbone. The dense contrastive learning designed on top of detection losses leads to improved feature representations across both the BEV encoder and the backbone. Extensive experiments on the nuScenes dataset demonstrate that BEVCon achieves consistent performance gains, achieving up to +2.4% mAP improvement over state-of-the-art baselines. Our results highlight the critical role of representation learning in BEV perception and offer a complementary avenue to conventional task-specific optimizations.", "date": "2025-08-06", "categories": []} {"id": "2406.00830", "title": "Collaborative Novel Object Discovery and Box-Guided Cross-Modal Alignment for Open-Vocabulary 3D Object Detection", "abstract": "Open-vocabulary 3D Object Detection (OV-3DDet) addresses the detection of objects from an arbitrary list of novel categories in 3D scenes, which remains a very challenging problem. In this work, we propose CoDAv2, a unified framework designed to innovatively tackle both the localization and classification of novel 3D objects, under the condition of limited base categories. For localization, the proposed 3D Novel Object Discovery (3D-NOD) strategy utilizes 3D geometries and 2D open-vocabulary semantic priors to discover pseudo labels for novel objects during training. 3D-NOD is further extended with an Enrichment strategy that significantly enriches the novel object distribution in the training scenes, and then enhances the model's ability to localize more novel objects. The 3D-NOD with Enrichment is termed 3D-NODE. For classification, the Discovery-driven Cross-modal Alignment (DCMA) module aligns features from 3D point clouds and 2D/textual modalities, employing both class-agnostic and class-specific alignments that are iteratively refined to handle the expanding vocabulary of objects. Besides, 2D box guidance boosts the classification accuracy against complex background noises, which is coined as Box-DCMA. Extensive evaluation demonstrates the superiority of CoDAv2. CoDAv2 outperforms the best-performing method by a large margin (AP_Novel of 9.17 vs. 3.61 on SUN-RGBD and 9.12 vs. 3.74 on ScanNetv2). Source code and pre-trained models are available at the GitHub project page.", "date": "2024-06-02", "categories": []} {"id": "2503.04154", "title": "CA-W3D: Leveraging Context-Aware Knowledge for Weakly Supervised Monocular 3D Detection", "abstract": "Weakly supervised monocular 3D detection, while less annotation-intensive, often struggles to capture the global context required for reliable 3D reasoning. Conventional label-efficient methods focus on object-centric features, neglecting contextual semantic relationships that are critical in complex scenes. In this work, we propose a Context-Aware Weak Supervision for Monocular 3D object detection, namely CA-W3D, to address this limitation in a two-stage training paradigm. Specifically, we first introduce a pre-training stage employing Region-wise Object Contrastive Matching (ROCM), which aligns regional object embeddings derived from a trainable monocular 3D encoder and a frozen open-vocabulary 2D visual grounding model. This alignment encourages the monocular encoder to discriminate scene-specific attributes and acquire richer contextual knowledge. In the second stage, we incorporate a pseudo-label training process with a Dual-to-One Distillation (D2OD) mechanism, which effectively transfers contextual priors into the monocular encoder while preserving spatial fidelity and maintaining computational efficiency during inference. Extensive experiments conducted on the public KITTI benchmark demonstrate the effectiveness of our approach, surpassing the SoTA method over all metrics, highlighting the importance of contextual-aware knowledge in weakly-supervised monocular 3D detection.", "date": "2025-03-06", "categories": []} {"id": "2507.21358", "title": "Collaborative Perceiver: Elevating Vision-based 3D Object Detection via Local Density-Aware Spatial Occupancy", "abstract": "Vision-based bird's-eye-view (BEV) 3D object detection has advanced significantly in autonomous driving by offering cost-effectiveness and rich contextual information. However, existing methods often construct BEV representations by collapsing extracted object features, neglecting intrinsic environmental contexts, such as roads and pavements. This hinders detectors from comprehensively perceiving the characteristics of the physical world. To alleviate this, we introduce a multi-task learning framework, Collaborative Perceiver (CoP), that leverages spatial occupancy as auxiliary information to mine consistent structural and conceptual similarities shared between 3D object detection and occupancy prediction tasks, bridging gaps in spatial representations and feature refinement. To this end, we first propose a pipeline to generate dense occupancy ground truths incorporating local density information (LDO) for reconstructing detailed environmental information. Next, we employ a voxel-height-guided sampling (VHS) strategy to distill fine-grained local features according to distinct object properties. Furthermore, we develop a global-local collaborative feature fusion (CFF) module that seamlessly integrates complementary knowledge between both tasks, thus composing more robust BEV representations. Extensive experiments on the nuScenes benchmark demonstrate that CoP outperforms existing vision-based frameworks, achieving 49.5\\% mAP and 59.2\\% NDS on the test set. Code and supplementary materials are available at this link https://github.com/jichengyuan/Collaborative-Perceiver.", "date": "2025-07-28", "categories": []} {"id": "2508.02288", "title": "Unleashing the Temporal Potential of Stereo Event Cameras for Continuous-Time 3D Object Detection", "abstract": "3D object detection is essential for autonomous systems, enabling precise localization and dimension estimation. While LiDAR and RGB cameras are widely used, their fixed frame rates create perception gaps in high-speed scenarios. Event cameras, with their asynchronous nature and high temporal resolution, offer a solution by capturing motion continuously. The recent approach, which integrates event cameras with conventional sensors for continuous-time detection, struggles in fast-motion scenarios due to its dependency on synchronized sensors. We propose a novel stereo 3D object detection framework that relies solely on event cameras, eliminating the need for conventional 3D sensors. To compensate for the lack of semantic and geometric information in event data, we introduce a dual filter mechanism that extracts both. Additionally, we enhance regression by aligning bounding boxes with object-centric information. Experiments show that our method outperforms prior approaches in dynamic environments, demonstrating the potential of event cameras for robust, continuous-time 3D perception. The code is available at https://github.com/mickeykang16/Ev-Stereo3D.", "date": "2025-08-04", "categories": []} {"id": "2508.00169", "title": "Robust 3D Object Detection using Probabilistic Point Clouds from Single-Photon LiDARs", "abstract": "LiDAR-based 3D sensors provide point clouds, a canonical 3D representation used in various scene understanding tasks. Modern LiDARs face key challenges in several real-world scenarios, such as long-distance or low-albedo objects, producing sparse or erroneous point clouds. These errors, which are rooted in the noisy raw LiDAR measurements, get propagated to downstream perception models, resulting in potentially severe loss of accuracy. This is because conventional 3D processing pipelines do not retain any uncertainty information from the raw measurements when constructing point clouds.\n We propose Probabilistic Point Clouds (PPC), a novel 3D scene representation where each point is augmented with a probability attribute that encapsulates the measurement uncertainty (or confidence) in the raw data. We further introduce inference approaches that leverage PPC for robust 3D object detection; these methods are versatile and can be used as computationally lightweight drop-in modules in 3D inference pipelines. We demonstrate, via both simulations and real captures, that PPC-based 3D inference methods outperform several baselines using LiDAR as well as camera-LiDAR fusion models, across challenging indoor and outdoor scenarios involving small, distant, and low-albedo objects, as well as strong ambient light.\n Our project webpage is at https://bhavyagoyal.github.io/ppc .", "date": "2025-07-31", "categories": []} {"id": "2412.18884", "title": "HV-BEV: Decoupling Horizontal and Vertical Feature Sampling for Multi-View 3D Object Detection", "abstract": "The application of vision-based multi-view environmental perception system has been increasingly recognized in autonomous driving technology, especially the BEV-based models. Current state-of-the-art solutions primarily encode image features from each camera view into the BEV space through explicit or implicit depth prediction. However, these methods often overlook the structured correlations among different parts of objects in 3D space and the fact that different categories of objects often occupy distinct local height ranges. For example, trucks appear at higher elevations, whereas traffic cones are near the ground. In this work, we propose a novel approach that decouples feature sampling in the \\textbf{BEV} grid queries paradigm into \\textbf{H}orizontal feature aggregation and \\textbf{V}ertical adaptive height-aware reference point sampling (HV-BEV), aiming to improve both the aggregation of objects' complete information and awareness of diverse objects' height distribution. Specifically, a set of relevant neighboring points is dynamically constructed for each 3D reference point on the ground-aligned horizontal plane, enhancing the association of the same instance across different BEV grids, especially when the instance spans multiple image views around the vehicle. Additionally, instead of relying on uniform sampling within a fixed height range, we introduce a height-aware module that incorporates historical information, enabling the reference points to adaptively focus on the varying heights at which objects appear in different scenes. Extensive experiments validate the effectiveness of our proposed method, demonstrating its superior performance over the baseline across the nuScenes dataset. Moreover, our best-performing model achieves a remarkable 50.5\\% mAP and 59.8\\% NDS on the nuScenes testing set. The code is available at https://github.com/Uddd821/HV-BEV.", "date": "2024-12-25", "categories": []} {"id": "2507.19304", "title": "Multistream Network for LiDAR and Camera-based 3D Object Detection in Outdoor Scenes", "abstract": "Fusion of LiDAR and RGB data has the potential to enhance outdoor 3D object detection accuracy. To address real-world challenges in outdoor 3D object detection, fusion of LiDAR and RGB input has started gaining traction. However, effective integration of these modalities for precise object detection task still remains a largely open problem. To address that, we propose a MultiStream Detection (MuStD) network, that meticulously extracts task-relevant information from both data modalities. The network follows a three-stream structure. Its LiDAR-PillarNet stream extracts sparse 2D pillar features from the LiDAR input while the LiDAR-Height Compression stream computes Bird's-Eye View features. An additional 3D Multimodal stream combines RGB and LiDAR features using UV mapping and polar coordinate indexing. Eventually, the features containing comprehensive spatial, textural and geometric information are carefully fused and fed to a detection head for 3D object detection. Our extensive evaluation on the challenging KITTI Object Detection Benchmark using public testing server at https://www.cvlibs.net/datasets/kitti/eval_object_detail.php?&result=d162ec699d6992040e34314d19ab7f5c217075e0 establishes the efficacy of our method by achieving new state-of-the-art or highly competitive results in different categories while remaining among the most efficient methods. Our code will be released through MuStD GitHub repository at https://github.com/IbrahimUWA/MuStD.git", "date": "2025-07-25", "categories": []} {"id": "2507.18331", "title": "Boosting Multi-View Indoor 3D Object Detection via Adaptive 3D Volume Construction", "abstract": "This work presents SGCDet, a novel multi-view indoor 3D object detection framework based on adaptive 3D volume construction. Unlike previous approaches that restrict the receptive field of voxels to fixed locations on images, we introduce a geometry and context aware aggregation module to integrate geometric and contextual information within adaptive regions in each image and dynamically adjust the contributions from different views, enhancing the representation capability of voxel features. Furthermore, we propose a sparse volume construction strategy that adaptively identifies and selects voxels with high occupancy probabilities for feature refinement, minimizing redundant computation in free space. Benefiting from the above designs, our framework achieves effective and efficient volume construction in an adaptive way. Better still, our network can be supervised using only 3D bounding boxes, eliminating the dependence on ground-truth scene geometry. Experimental results demonstrate that SGCDet achieves state-of-the-art performance on the ScanNet, ScanNet200 and ARKitScenes datasets. The source code is available at https://github.com/RM-Zhang/SGCDet.", "date": "2025-07-24", "categories": []} {"id": "2507.17079", "title": "Few-Shot Learning in Video and 3D Object Detection: A Survey", "abstract": "Few-shot learning (FSL) enables object detection models to recognize novel classes given only a few annotated examples, thereby reducing expensive manual data labeling. This survey examines recent FSL advances for video and 3D object detection. For video, FSL is especially valuable since annotating objects across frames is more laborious than for static images. By propagating information across frames, techniques like tube proposals and temporal matching networks can detect new classes from a couple examples, efficiently leveraging spatiotemporal structure. FSL for 3D detection from LiDAR or depth data faces challenges like sparsity and lack of texture. Solutions integrate FSL with specialized point cloud networks and losses tailored for class imbalance. Few-shot 3D detection enables practical autonomous driving deployment by minimizing costly 3D annotation needs. Core issues in both domains include balancing generalization and overfitting, integrating prototype matching, and handling data modality properties. In summary, FSL shows promise for reducing annotation requirements and enabling real-world video, 3D, and other applications by efficiently leveraging information across feature, temporal, and data modalities. By comprehensively surveying recent advancements, this paper illuminates FSL's potential to minimize supervision needs and enable deployment across video, 3D, and other real-world applications.", "date": "2025-07-22", "categories": []} {"id": "2507.17445", "title": "IndoorBEV: Joint Detection and Footprint Completion of Objects via Mask-based Prediction in Indoor Scenarios for Bird's-Eye View Perception", "abstract": "Detecting diverse objects within complex indoor 3D point clouds presents significant challenges for robotic perception, particularly with varied object shapes, clutter, and the co-existence of static and dynamic elements where traditional bounding box methods falter. To address these limitations, we propose IndoorBEV, a novel mask-based Bird's-Eye View (BEV) method for indoor mobile robots.\n In a BEV method, a 3D scene is projected into a 2D BEV grid which handles naturally occlusions and provides a consistent top-down view aiding to distinguish static obstacles from dynamic agents. The obtained 2D BEV results is directly usable to downstream robotic tasks like navigation, motion prediction, and planning. Our architecture utilizes an axis compact encoder and a window-based backbone to extract rich spatial features from this BEV map. A query-based decoder head then employs learned object queries to concurrently predict object classes and instance masks in the BEV space. This mask-centric formulation effectively captures the footprint of both static and dynamic objects regardless of their shape, offering a robust alternative to bounding box regression. We demonstrate the effectiveness of IndoorBEV on a custom indoor dataset featuring diverse object classes including static objects\n and dynamic elements like robots and miscellaneous items, showcasing its potential for robust indoor scene understanding.", "date": "2025-07-23", "categories": []} {"id": "2507.04123", "title": "Towards Accurate and Efficient 3D Object Detection for Autonomous Driving: A Mixture of Experts Computing System on Edge", "abstract": "This paper presents Edge-based Mixture of Experts (MoE) Collaborative Computing (EMC2), an optimal computing system designed for autonomous vehicles (AVs) that simultaneously achieves low-latency and high-accuracy 3D object detection. Unlike conventional approaches, EMC2 incorporates a scenario-aware MoE architecture specifically optimized for edge platforms. By effectively fusing LiDAR and camera data, the system leverages the complementary strengths of sparse 3D point clouds and dense 2D images to generate robust multimodal representations. To enable this, EMC2 employs an adaptive multimodal data bridge that performs multi-scale preprocessing on sensor inputs, followed by a scenario-aware routing mechanism that dynamically dispatches features to dedicated expert models based on object visibility and distance. In addition, EMC2 integrates joint hardware-software optimizations, including hardware resource utilization optimization and computational graph simplification, to ensure efficient and real-time inference on resource-constrained edge devices. Experiments on open-source benchmarks clearly show the EMC2 advancements as an end-to-end system. On the KITTI dataset, it achieves an average accuracy improvement of 3.58% and a 159.06% inference speedup compared to 15 baseline methods on Jetson platforms, with similar performance gains on the nuScenes dataset, highlighting its capability to advance reliable, real-time 3D object detection tasks for AVs. The official implementation is available at https://github.com/LinshenLiu622/EMC2.", "date": "2025-07-05", "categories": []} {"id": "2507.16413", "title": "Towards Railway Domain Adaptation for LiDAR-based 3D Detection: Road-to-Rail and Sim-to-Real via SynDRA-BBox", "abstract": "In recent years, interest in automatic train operations has significantly increased. To enable advanced functionalities, robust vision-based algorithms are essential for perceiving and understanding the surrounding environment. However, the railway sector suffers from a lack of publicly available real-world annotated datasets, making it challenging to test and validate new perception solutions in this domain. To address this gap, we introduce SynDRA-BBox, a synthetic dataset designed to support object detection and other vision-based tasks in realistic railway scenarios. To the best of our knowledge, is the first synthetic dataset specifically tailored for 2D and 3D object detection in the railway domain, the dataset is publicly available at https://syndra.retis.santannapisa.it. In the presented evaluation, a state-of-the-art semi-supervised domain adaptation method, originally developed for automotive perception, is adapted to the railway context, enabling the transferability of synthetic data to 3D object detection. Experimental results demonstrate promising performance, highlighting the effectiveness of synthetic datasets and domain adaptation techniques in advancing perception capabilities for railway environments.", "date": "2025-07-22", "categories": []} {"id": "2507.13899", "title": "Enhancing LiDAR Point Features with Foundation Model Priors for 3D Object Detection", "abstract": "Recent advances in foundation models have opened up new possibilities for enhancing 3D perception. In particular, DepthAnything offers dense and reliable geometric priors from monocular RGB images, which can complement sparse LiDAR data in autonomous driving scenarios. However, such priors remain underutilized in LiDAR-based 3D object detection. In this paper, we address the limited expressiveness of raw LiDAR point features, especially the weak discriminative capability of the reflectance attribute, by introducing depth priors predicted by DepthAnything. These priors are fused with the original LiDAR attributes to enrich each point's representation. To leverage the enhanced point features, we propose a point-wise feature extraction module. Then, a Dual-Path RoI feature extraction framework is employed, comprising a voxel-based branch for global semantic context and a point-based branch for fine-grained structural details. To effectively integrate the complementary RoI features, we introduce a bidirectional gated RoI feature fusion module that balances global and local cues. Extensive experiments on the KITTI benchmark show that our method consistently improves detection accuracy, demonstrating the value of incorporating visual foundation model priors into LiDAR-based 3D object detection.", "date": "2025-07-18", "categories": []} {"id": "2409.00839", "title": "Entropy Loss: An Interpretability Amplifier of 3D Object Detection Network for Intelligent Driving", "abstract": "With the increasing complexity of the traffic environment, the significance of safety perception in intelligent driving is intensifying. Traditional methods in the field of intelligent driving perception rely on deep learning, which suffers from limited interpretability, often described as a \"black box.\" This paper introduces a novel type of loss function, termed \"Entropy Loss,\" along with an innovative training strategy. Entropy Loss is formulated based on the functionality of feature compression networks within the perception model. Drawing inspiration from communication systems, the information transmission process in a feature compression network is expected to demonstrate steady changes in information volume and a continuous decrease in information entropy. By modeling network layer outputs as continuous random variables, we construct a probabilistic model that quantifies changes in information volume. Entropy Loss is then derived based on these expectations, guiding the update of network parameters to enhance network interpretability. Our experiments indicate that the Entropy Loss training strategy accelerates the training process. Utilizing the same 60 training epochs, the accuracy of 3D object detection models using Entropy Loss on the KITTI test set improved by up to 4.47\\% compared to models without Entropy Loss, underscoring the method's efficacy. The implementation code is available at https://github.com/yhbcode000/Eloss-Interpretability.", "date": "2024-09-01", "categories": []} {"id": "2507.13363", "title": "Just Add Geometry: Gradient-Free Open-Vocabulary 3D Detection Without Human-in-the-Loop", "abstract": "Modern 3D object detection datasets are constrained by narrow class taxonomies and costly manual annotations, limiting their ability to scale to open-world settings. In contrast, 2D vision-language models trained on web-scale image-text pairs exhibit rich semantic understanding and support open-vocabulary detection via natural language prompts. In this work, we leverage the maturity and category diversity of 2D foundation models to perform open-vocabulary 3D object detection without any human-annotated 3D labels.\n Our pipeline uses a 2D vision-language detector to generate text-conditioned proposals, which are segmented with SAM and back-projected into 3D using camera geometry and either LiDAR or monocular pseudo-depth. We introduce a geometric inflation strategy based on DBSCAN clustering and Rotating Calipers to infer 3D bounding boxes without training. To simulate adverse real-world conditions, we construct Pseudo-nuScenes, a fog-augmented, RGB-only variant of the nuScenes dataset.\n Experiments demonstrate that our method achieves competitive localization performance across multiple settings, including LiDAR-based and purely RGB-D inputs, all while remaining training-free and open-vocabulary. Our results highlight the untapped potential of 2D foundation models for scalable 3D perception. We open-source our code and resources at https://github.com/atharv0goel/open-world-3D-det.", "date": "2025-07-06", "categories": []} {"id": "2411.10715", "title": "EVT: Efficient View Transformation for Multi-Modal 3D Object Detection", "abstract": "Multi-modal sensor fusion in Bird's Eye View (BEV) representation has become the leading approach for 3D object detection. However, existing methods often rely on depth estimators or transformer encoders to transform image features into BEV space, which reduces robustness or introduces significant computational overhead. Moreover, the insufficient geometric guidance in view transformation results in ray-directional misalignments, limiting the effectiveness of BEV representations. To address these challenges, we propose Efficient View Transformation (EVT), a novel 3D object detection framework that constructs a well-structured BEV representation, improving both accuracy and efficiency. Our approach focuses on two key aspects. First, Adaptive Sampling and Adaptive Projection (ASAP), which utilizes LiDAR guidance to generate 3D sampling points and adaptive kernels, enables more effective transformation of image features into BEV space and a refined BEV representation. Second, an improved query-based detection framework, incorporating group-wise mixed query selection and geometry-aware cross-attention, effectively captures both the common properties and the geometric structure of objects in the transformer decoder. On the nuScenes test set, EVT achieves state-of-the-art performance of 75.3% NDS with real-time inference speed.", "date": "2024-11-16", "categories": []} {"id": "2507.08644", "title": "OnlineBEV: Recurrent Temporal Fusion in Bird's Eye View Representations for Multi-Camera 3D Perception", "abstract": "Multi-view camera-based 3D perception can be conducted using bird's eye view (BEV) features obtained through perspective view-to-BEV transformations. Several studies have shown that the performance of these 3D perception methods can be further enhanced by combining sequential BEV features obtained from multiple camera frames. However, even after compensating for the ego-motion of an autonomous agent, the performance gain from temporal aggregation is limited when combining a large number of image frames. This limitation arises due to dynamic changes in BEV features over time caused by object motion. In this paper, we introduce a novel temporal 3D perception method called OnlineBEV, which combines BEV features over time using a recurrent structure. This structure increases the effective number of combined features with minimal memory usage. However, it is critical to spatially align the features over time to maintain strong performance. OnlineBEV employs the Motion-guided BEV Fusion Network (MBFNet) to achieve temporal feature alignment. MBFNet extracts motion features from consecutive BEV frames and dynamically aligns historical BEV features with current ones using these motion features. To enforce temporal feature alignment explicitly, we use Temporal Consistency Learning Loss, which captures discrepancies between historical and target BEV features. Experiments conducted on the nuScenes benchmark demonstrate that OnlineBEV achieves significant performance gains over the current best method, SOLOFusion. OnlineBEV achieves 63.9% NDS on the nuScenes test set, recording state-of-the-art performance in the camera-only 3D object detection task.", "date": "2025-07-11", "categories": []} {"id": "2507.06976", "title": "DenoiseCP-Net: Efficient Collective Perception in Adverse Weather via Joint LiDAR-Based 3D Object Detection and Denoising", "abstract": "While automated vehicles hold the potential to significantly reduce traffic accidents, their perception systems remain vulnerable to sensor degradation caused by adverse weather and environmental occlusions. Collective perception, which enables vehicles to share information, offers a promising approach to overcoming these limitations. However, to this date collective perception in adverse weather is mostly unstudied. Therefore, we conduct the first study of LiDAR-based collective perception under diverse weather conditions and present a novel multi-task architecture for LiDAR-based collective perception under adverse weather. Adverse weather conditions can not only degrade perception capabilities, but also negatively affect bandwidth requirements and latency due to the introduced noise that is also transmitted and processed. Denoising prior to communication can effectively mitigate these issues. Therefore, we propose DenoiseCP-Net, a novel multi-task architecture for LiDAR-based collective perception under adverse weather conditions. DenoiseCP-Net integrates voxel-level noise filtering and object detection into a unified sparse convolution backbone, eliminating redundant computations associated with two-stage pipelines. This design not only reduces inference latency and computational cost but also minimizes communication overhead by removing non-informative noise. We extended the well-known OPV2V dataset by simulating rain, snow, and fog using our realistic weather simulation models. We demonstrate that DenoiseCP-Net achieves near-perfect denoising accuracy in adverse weather, reduces the bandwidth requirements by up to 23.6% while maintaining the same detection accuracy and reducing the inference latency for cooperative vehicles.", "date": "2025-07-09", "categories": []} {"id": "2504.09506", "title": "Pillar-Voxel Fusion Network for 3D Object Detection in Airborne Hyperspectral Point Clouds", "abstract": "Hyperspectral point clouds (HPCs) can simultaneously characterize 3D spatial and spectral information of ground objects, offering excellent 3D perception and target recognition capabilities. Current approaches for generating HPCs often involve fusion techniques with hyperspectral images and LiDAR point clouds, which inevitably lead to geometric-spectral distortions due to fusion errors and obstacle occlusions. These adverse effects limit their performance in downstream fine-grained tasks across multiple scenarios, particularly in airborne applications. To address these issues, we propose PiV-AHPC, a 3D object detection network for airborne HPCs. To the best of our knowledge, this is the first attempt at this HPCs task. Specifically, we first develop a pillar-voxel dual-branch encoder, where the former captures spectral and vertical structural features from HPCs to overcome spectral distortion, while the latter emphasizes extracting accurate 3D spatial features from point clouds. A multi-level feature fusion mechanism is devised to enhance information interaction between the two branches, achieving neighborhood feature alignment and channel-adaptive selection, thereby organically integrating heterogeneous features and mitigating geometric distortion. Extensive experiments on two airborne HPCs datasets demonstrate that PiV-AHPC possesses state-of-the-art detection performance and high generalization capability.", "date": "2025-04-13", "categories": []} {"id": "2507.02393", "title": "PLOT: Pseudo-Labeling via Video Object Tracking for Scalable Monocular 3D Object Detection", "abstract": "Monocular 3D object detection (M3OD) has long faced challenges due to data scarcity caused by high annotation costs and inherent 2D-to-3D ambiguity. Although various weakly supervised methods and pseudo-labeling methods have been proposed to address these issues, they are mostly limited by domain-specific learning or rely solely on shape information from a single observation. In this paper, we propose a novel pseudo-labeling framework that uses only video data and is more robust to occlusion, without requiring a multi-view setup, additional sensors, camera poses, or domain-specific training. Specifically, we explore a technique for aggregating the pseudo-LiDARs of both static and dynamic objects across temporally adjacent frames using object point tracking, enabling 3D attribute extraction in scenarios where 3D data acquisition is infeasible. Extensive experiments demonstrate that our method ensures reliable accuracy and strong scalability, making it a practical and effective solution for M3OD.", "date": "2025-07-03", "categories": []} {"id": "2507.04369", "title": "MambaFusion: Height-Fidelity Dense Global Fusion for Multi-modal 3D Object Detection", "abstract": "We present the first work demonstrating that a pure Mamba block can achieve efficient Dense Global Fusion, meanwhile guaranteeing top performance for camera-LiDAR multi-modal 3D object detection. Our motivation stems from the observation that existing fusion strategies are constrained by their inability to simultaneously achieve efficiency, long-range modeling, and retaining complete scene information. Inspired by recent advances in state-space models (SSMs) and linear attention, we leverage their linear complexity and long-range modeling capabilities to address these challenges. However, this is non-trivial since our experiments reveal that simply adopting efficient linear-complexity methods does not necessarily yield improvements and may even degrade performance. We attribute this degradation to the loss of height information during multi-modal alignment, leading to deviations in sequence order. To resolve this, we propose height-fidelity LiDAR encoding that preserves precise height information through voxel compression in continuous space, thereby enhancing camera-LiDAR alignment. Subsequently, we introduce the Hybrid Mamba Block, which leverages the enriched height-informed features to conduct local and global contextual learning. By integrating these components, our method achieves state-of-the-art performance with the top-tire NDS score of 75.0 on the nuScenes validation benchmark, even surpassing methods that utilize high-resolution inputs. Meanwhile, our method maintains efficiency, achieving faster inference speed than most recent state-of-the-art methods.", "date": "2025-07-06", "categories": []} {"id": "2507.04587", "title": "CVFusion: Cross-View Fusion of 4D Radar and Camera for 3D Object Detection", "abstract": "4D radar has received significant attention in autonomous driving thanks to its robustness under adverse weathers. Due to the sparse points and noisy measurements of the 4D radar, most of the research finish the 3D object detection task by integrating images from camera and perform modality fusion in BEV space. However, the potential of the radar and the fusion mechanism is still largely unexplored, hindering the performance improvement. In this study, we propose a cross-view two-stage fusion network called CVFusion. In the first stage, we design a radar guided iterative (RGIter) BEV fusion module to generate high-recall 3D proposal boxes. In the second stage, we aggregate features from multiple heterogeneous views including points, image, and BEV for each proposal. These comprehensive instance level features greatly help refine the proposals and generate high-quality predictions. Extensive experiments on public datasets show that our method outperforms the previous state-of-the-art methods by a large margin, with 9.10% and 3.68% mAP improvements on View-of-Delft (VoD) and TJ4DRadSet, respectively. Our code will be made publicly available.", "date": "2025-07-07", "categories": []} {"id": "2408.05945", "title": "MV2DFusion: Leveraging Modality-Specific Object Semantics for Multi-Modal 3D Detection", "abstract": "The rise of autonomous vehicles has significantly increased the demand for robust 3D object detection systems. While cameras and LiDAR sensors each offer unique advantages--cameras provide rich texture information and LiDAR offers precise 3D spatial data--relying on a single modality often leads to performance limitations. This paper introduces MV2DFusion, a multi-modal detection framework that integrates the strengths of both worlds through an advanced query-based fusion mechanism. By introducing an image query generator to align with image-specific attributes and a point cloud query generator, MV2DFusion effectively combines modality-specific object semantics without biasing toward one single modality. Then the sparse fusion process can be accomplished based on the valuable object semantics, ensuring efficient and accurate object detection across various scenarios. Our framework's flexibility allows it to integrate with any image and point cloud-based detectors, showcasing its adaptability and potential for future advancements. Extensive evaluations on the nuScenes and Argoverse2 datasets demonstrate that MV2DFusion achieves state-of-the-art performance, particularly excelling in long-range detection scenarios.", "date": "2024-08-12", "categories": []} {"id": "2310.05346", "title": "Anyview: Generalizable Indoor 3D Object Detection with Variable Frames", "abstract": "In this paper, we propose a novel network framework for indoor 3D object detection to handle variable input frame numbers in practical scenarios. Existing methods only consider fixed frames of input data for a single detector, such as monocular RGB-D images or point clouds reconstructed from dense multi-view RGB-D images. While in practical application scenes such as robot navigation and manipulation, the raw input to the 3D detectors is the RGB-D images with variable frame numbers instead of the reconstructed scene point cloud. However, the previous approaches can only handle fixed frame input data and have poor performance with variable frame input. In order to facilitate 3D object detection methods suitable for practical tasks, we present a novel 3D detection framework named AnyView for our practical applications, which generalizes well across different numbers of input frames with a single model. To be specific, we propose a geometric learner to mine the local geometric features of each input RGB-D image frame and implement local-global feature interaction through a designed spatial mixture module. Meanwhile, we further utilize a dynamic token strategy to adaptively adjust the number of extracted features for each frame, which ensures consistent global feature density and further enhances the generalization after fusion. Extensive experiments on the ScanNet dataset show our method achieves both great generalizability and high detection accuracy with a simple and clean architecture containing a similar amount of parameters with the baselines.", "date": "2023-10-09", "categories": []} {"id": "2405.18734", "title": "Information Entropy Guided Height-aware Histogram for Quantization-friendly Pillar Feature Encoder", "abstract": "Real-time and high-performance 3D object detection plays a critical role in autonomous driving and robotics. Recent pillar-based 3D object detectors have gained significant attention due to their compact representation and low computational overhead, making them suitable for onboard deployment and quantization. However, existing pillar-based detectors still suffer from information loss along height dimension and large numerical distribution difference during pillar feature encoding (PFE), which severely limits their performance and quantization potential. To address above issue, we first unveil the importance of different input information during PFE and identify the height dimension as a key factor in enhancing 3D detection performance. Motivated by this observation, we propose a height-aware pillar feature encoder, called PillarHist. Specifically, PillarHist statistics the discrete distribution of points at different heights within one pillar with the information entropy guidance. This simple yet effective design greatly preserves the information along the height dimension while significantly reducing the computation overhead of the PFE. Meanwhile, PillarHist also constrains the arithmetic distribution of PFE input to a stable range, making it quantization-friendly. Notably, PillarHist operates exclusively within the PFE stage to enhance performance, enabling seamless integration into existing pillar-based methods without introducing complex operations. Extensive experiments show the effectiveness of PillarHist in terms of both efficiency and performance.", "date": "2024-05-29", "categories": []} {"id": "2408.06832", "title": "FlatFusion: Delving into Details of Sparse Transformer-based Camera-LiDAR Fusion for Autonomous Driving", "abstract": "The integration of data from diverse sensor modalities (e.g., camera and LiDAR) constitutes a prevalent methodology within the ambit of autonomous driving scenarios. Recent advancements in efficient point cloud transformers have underscored the efficacy of integrating information in sparse formats. When it comes to fusion, since image patches are dense in pixel space with ambiguous depth, it necessitates additional design considerations for effective fusion. In this paper, we conduct a comprehensive exploration of design choices for Transformer-based sparse cameraLiDAR fusion. This investigation encompasses strategies for image-to-3D and LiDAR-to-2D mapping, attention neighbor grouping, single modal tokenizer, and micro-structure of Transformer. By amalgamating the most effective principles uncovered through our investigation, we introduce FlatFusion, a carefully designed framework for sparse camera-LiDAR fusion. Notably, FlatFusion significantly outperforms state-of-the-art sparse Transformer-based methods, including UniTR, CMT, and SparseFusion, achieving 73.7 NDS on the nuScenes validation set with 10.1 FPS with PyTorch.", "date": "2024-08-13", "categories": []} {"id": "2411.09062", "title": "Multimodal Object Detection using Depth and Image Data for Manufacturing Parts", "abstract": "Manufacturing requires reliable object detection methods for precise picking and handling of diverse types of manufacturing parts and components. Traditional object detection methods utilize either only 2D images from cameras or 3D data from lidars or similar 3D sensors. However, each of these sensors have weaknesses and limitations. Cameras do not have depth perception and 3D sensors typically do not carry color information. These weaknesses can undermine the reliability and robustness of industrial manufacturing systems. To address these challenges, this work proposes a multi-sensor system combining an red-green-blue (RGB) camera and a 3D point cloud sensor. The two sensors are calibrated for precise alignment of the multimodal data captured from the two hardware devices. A novel multimodal object detection method is developed to process both RGB and depth data. This object detector is based on the Faster R-CNN baseline that was originally designed to process only camera images. The results show that the multimodal model significantly outperforms the depth-only and RGB-only baselines on established object detection metrics. More specifically, the multimodal model improves mAP by 13% and raises Mean Precision by 11.8% in comparison to the RGB-only baseline. Compared to the depth-only baseline, it improves mAP by 78% and raises Mean Precision by 57%. Hence, this method facilitates more reliable and robust object detection in service to smart manufacturing applications.", "date": "2024-11-13", "categories": []} {"id": "2503.08101", "title": "Accelerate 3D Object Detection Models via Zero-Shot Attention Key Pruning", "abstract": "Query-based methods with dense features have demonstrated remarkable success in 3D object detection tasks. However, the computational demands of these models, particularly with large image sizes and multiple transformer layers, pose significant challenges for efficient running on edge devices. Existing pruning and distillation methods either need retraining or are designed for ViT models, which are hard to migrate to 3D detectors. To address this issue, we propose a zero-shot runtime pruning method for transformer decoders in 3D object detection models. The method, termed tgGBC (trim keys gradually Guided By Classification scores), systematically trims keys in transformer modules based on their importance. We expand the classification score to multiply it with the attention map to get the importance score of each key and then prune certain keys after each transformer layer according to their importance scores. Our method achieves a 1.99x speedup in the transformer decoder of the latest ToC3D model, with only a minimal performance loss of less than 1%. Interestingly, for certain models, our method even enhances their performance. Moreover, we deploy 3D detectors with tgGBC on an edge device, further validating the effectiveness of our method. The code can be found at https://github.com/iseri27/tg_gbc.", "date": "2025-03-11", "categories": []} {"id": "2506.23565", "title": "OcRFDet: Object-Centric Radiance Fields for Multi-View 3D Object Detection in Autonomous Driving", "abstract": "Current multi-view 3D object detection methods typically transfer 2D features into 3D space using depth estimation or 3D position encoder, but in a fully data-driven and implicit manner, which limits the detection performance. Inspired by the success of radiance fields on 3D reconstruction, we assume they can be used to enhance the detector's ability of 3D geometry estimation. However, we observe a decline in detection performance, when we directly use them for 3D rendering as an auxiliary task. From our analysis, we find the performance drop is caused by the strong responses on the background when rendering the whole scene. To address this problem, we propose object-centric radiance fields, focusing on modeling foreground objects while discarding background noises. Specifically, we employ Object-centric Radiance Fields (OcRF) to enhance 3D voxel features via an auxiliary task of rendering foreground objects. We further use opacity - the side-product of rendering- to enhance the 2D foreground BEV features via Height-aware Opacity-based Attention (HOA), where attention maps at different height levels are generated separately via multiple networks in parallel. Extensive experiments on the nuScenes validation and test datasets demonstrate that our OcRFDet achieves superior performance, outperforming previous state-of-the-art methods with 57.2$\\%$ mAP and 64.8$\\%$ NDS on the nuScenes test benchmark. Code will be available at https://github.com/Mingqj/OcRFDet.", "date": "2025-06-30", "categories": []} {"id": "2410.10298", "title": "ROA-BEV: 2D Region-Oriented Attention for BEV-based 3D Object Detection", "abstract": "Vision-based Bird's-Eye-View (BEV) 3D object detection has recently become popular in autonomous driving. However, objects with a high similarity to the background from a camera perspective cannot be detected well by existing methods. In this paper, we propose a BEV-based 3D Object Detection Network with 2D Region-Oriented Attention (ROA-BEV), which enables the backbone to focus more on feature learning of the regions where objects exist. Moreover, our method further enhances the information feature learning ability of ROA through multi-scale structures. Each block of ROA utilizes a large kernel to ensure that the receptive field is large enough to catch information about large objects. Experiments on nuScenes show that ROA-BEV improves the performance based on BEVDepth. The source codes of this work will be available at https://github.com/DFLyan/ROA-BEV.", "date": "2024-10-14", "categories": []} {"id": "2506.21358", "title": "ToosiCubix: Monocular 3D Cuboid Labeling via Vehicle Part Annotations", "abstract": "Many existing methods for 3D cuboid annotation of vehicles rely on expensive and carefully calibrated camera-LiDAR or stereo setups, limiting their accessibility for large-scale data collection. We introduce ToosiCubix, a simple yet powerful approach for annotating ground-truth cuboids using only monocular images and intrinsic camera parameters. Our method requires only about 10 user clicks per vehicle, making it highly practical for adding 3D annotations to existing datasets originally collected without specialized equipment. By annotating specific features (e.g., wheels, car badge, symmetries) across different vehicle parts, we accurately estimate each vehicle's position, orientation, and dimensions up to a scale ambiguity (8 DoF). The geometric constraints are formulated as an optimization problem, which we solve using a coordinate descent strategy, alternating between Perspective-n-Points (PnP) and least-squares subproblems. To handle common ambiguities such as scale and unobserved dimensions, we incorporate probabilistic size priors, enabling 9 DoF cuboid placements. We validate our annotations against the KITTI and Cityscapes3D datasets, demonstrating that our method offers a cost-effective and scalable solution for high-quality 3D cuboid annotation.", "date": "2025-06-26", "categories": []} {"id": "2506.17958", "title": "ELMAR: Enhancing LiDAR Detection with 4D Radar Motion Awareness and Cross-modal Uncertainty", "abstract": "LiDAR and 4D radar are widely used in autonomous driving and robotics. While LiDAR provides rich spatial information, 4D radar offers velocity measurement and remains robust under adverse conditions. As a result, increasing studies have focused on the 4D radar-LiDAR fusion method to enhance the perception. However, the misalignment between different modalities is often overlooked. To address this challenge and leverage the strengths of both modalities, we propose a LiDAR detection framework enhanced by 4D radar motion status and cross-modal uncertainty. The object movement information from 4D radar is first captured using a Dynamic Motion-Aware Encoding module during feature extraction to enhance 4D radar predictions. Subsequently, the instance-wise uncertainties of bounding boxes are estimated to mitigate the cross-modal misalignment and refine the final LiDAR predictions. Extensive experiments on the View-of-Delft (VoD) dataset highlight the effectiveness of our method, achieving state-of-the-art performance with the mAP of 74.89% in the entire area and 88.70% within the driving corridor while maintaining a real-time inference speed of 30.02 FPS.", "date": "2025-06-22", "categories": []} {"id": "2501.09481", "title": "MonoSOWA: Scalable monocular 3D Object detector Without human Annotations", "abstract": "Inferring object 3D position and orientation from a single RGB camera is a foundational task in computer vision with many important applications. Traditionally, 3D object detection methods are trained in a fully-supervised setup, requiring LiDAR and vast amounts of human annotations, which are laborious, costly, and do not scale well with the ever-increasing amounts of data being captured.\n We present a novel method to train a 3D object detector from a single RGB camera without domain-specific human annotations, making orders of magnitude more data available for training. The method uses newly proposed Local Object Motion Model to disentangle object movement source between subsequent frames, is approximately 700 times faster than previous work and compensates camera focal length differences to aggregate multiple datasets.\n The method is evaluated on three public datasets, where despite using no human labels, it outperforms prior work by a significant margin. It also shows its versatility as a pre-training tool for fully-supervised training and shows that combining pseudo-labels from multiple datasets can achieve comparable accuracy to using human labels from a single dataset. The source code and model are available at https://github.com/jskvrna/MonoSOWA.", "date": "2025-01-16", "categories": []} {"id": "2506.16531", "title": "How Hard Is Snow? A Paired Domain Adaptation Dataset for Clear and Snowy Weather: CADC+", "abstract": "The impact of snowfall on 3D object detection performance remains underexplored. Conducting such an evaluation requires a dataset with sufficient labelled data from both weather conditions, ideally captured in the same driving environment. Current driving datasets with LiDAR point clouds either do not provide enough labelled data in both snowy and clear weather conditions, or rely on de-snowing methods to generate synthetic clear weather. Synthetic data often lacks realism and introduces an additional domain shift that confounds accurate evaluations. To address these challenges, we present CADC+, the first paired weather domain adaptation dataset for autonomous driving in winter conditions. CADC+ extends the Canadian Adverse Driving Conditions dataset (CADC) using clear weather data that was recorded on the same roads and in the same period as CADC. To create CADC+, we pair each CADC sequence with a clear weather sequence that matches the snowy sequence as closely as possible. CADC+ thus minimizes the domain shift resulting from factors unrelated to the presence of snow. We also present some preliminary results using CADC+ to evaluate the effect of snow on 3D object detection performance. We observe that snow introduces a combination of aleatoric and epistemic uncertainties, acting as both noise and a distinct data domain.", "date": "2025-06-19", "categories": []} {"id": "2506.14835", "title": "MonoVQD: Monocular 3D Object Detection with Variational Query Denoising and Self-Distillation", "abstract": "Precisely localizing 3D objects from a single image constitutes a central challenge in monocular 3D detection. While DETR-like architectures offer a powerful paradigm, their direct application in this domain encounters inherent limitations, preventing optimal performance. Our work addresses these challenges by introducing MonoVQD, a novel framework designed to fundamentally advance DETR-based monocular 3D detection. We propose three main contributions. First, we propose the Mask Separated Self-Attention mechanism that enables the integration of the denoising process into a DETR architecture. This improves the stability of Hungarian matching to achieve a consistent optimization objective. Second, we present the Variational Query Denoising technique to address the gradient vanishing problem of conventional denoising methods, which severely restricts the efficiency of the denoising process. This explicitly introduces stochastic properties to mitigate this fundamental limitation and unlock substantial performance gains. Finally, we introduce a sophisticated self-distillation strategy, leveraging insights from later decoder layers to synergistically improve query quality in earlier layers, thereby amplifying the iterative refinement process. Rigorous experimentation demonstrates that MonoVQD achieves superior performance on the challenging KITTI monocular benchmark. Highlighting its broad applicability, MonoVQD's core components seamlessly integrate into other architectures, delivering significant performance gains even in multi-view 3D detection scenarios on the nuScenes dataset and underscoring its robust generalization capabilities.", "date": "2025-06-14", "categories": []} {"id": "2503.21099", "title": "Learning Class Prototypes for Unified Sparse Supervised 3D Object Detection", "abstract": "Both indoor and outdoor scene perceptions are essential for embodied intelligence. However, current sparse supervised 3D object detection methods focus solely on outdoor scenes without considering indoor settings. To this end, we propose a unified sparse supervised 3D object detection method for both indoor and outdoor scenes through learning class prototypes to effectively utilize unlabeled objects. Specifically, we first propose a prototype-based object mining module that converts the unlabeled object mining into a matching problem between class prototypes and unlabeled features. By using optimal transport matching results, we assign prototype labels to high-confidence features, thereby achieving the mining of unlabeled objects. We then present a multi-label cooperative refinement module to effectively recover missed detections through pseudo label quality control and prototype label cooperation. Experiments show that our method achieves state-of-the-art performance under the one object per scene sparse supervised setting across indoor and outdoor datasets. With only one labeled object per scene, our method achieves about 78%, 90%, and 96% performance compared to the fully supervised detector on ScanNet V2, SUN RGB-D, and KITTI, respectively, highlighting the scalability of our method. Code is available at https://github.com/zyrant/CPDet3D.", "date": "2025-03-27", "categories": []} {"id": "2506.11804", "title": "Teleoperated Driving: a New Challenge for 3D Object Detection in Compressed Point Clouds", "abstract": "In recent years, the development of interconnected devices has expanded in many fields, from infotainment to education and industrial applications. This trend has been accelerated by the increased number of sensors and accessibility to powerful hardware and software. One area that significantly benefits from these advancements is Teleoperated Driving (TD). In this scenario, a controller drives safely a vehicle from remote leveraging sensors data generated onboard the vehicle, and exchanged via Vehicle-to-Everything (V2X) communications. In this work, we tackle the problem of detecting the presence of cars and pedestrians from point cloud data to enable safe TD operations. More specifically, we exploit the SELMA dataset, a multimodal, open-source, synthetic dataset for autonomous driving, that we expanded by including the ground-truth bounding boxes of 3D objects to support object detection. We analyze the performance of state-of-the-art compression algorithms and object detectors under several metrics, including compression efficiency, (de)compression and inference time, and detection accuracy. Moreover, we measure the impact of compression and detection on the V2X network in terms of data rate and latency with respect to 3GPP requirements for TD applications.", "date": "2025-06-13", "categories": []} {"id": "2506.11839", "title": "Vision-based Lifting of 2D Object Detections for Automated Driving", "abstract": "Image-based 3D object detection is an inevitable part of autonomous driving because cheap onboard cameras are already available in most modern cars. Because of the accurate depth information, currently, most state-of-the-art 3D object detectors heavily rely on LiDAR data. In this paper, we propose a pipeline which lifts the results of existing vision-based 2D algorithms to 3D detections using only cameras as a cost-effective alternative to LiDAR. In contrast to existing approaches, we focus not only on cars but on all types of road users. To the best of our knowledge, we are the first using a 2D CNN to process the point cloud for each 2D detection to keep the computational effort as low as possible. Our evaluation on the challenging KITTI 3D object detection benchmark shows results comparable to state-of-the-art image-based approaches while having a runtime of only a third.", "date": "2025-06-13", "categories": []} {"id": "2506.10242", "title": "DySS: Dynamic Queries and State-Space Learning for Efficient 3D Object Detection from Multi-Camera Videos", "abstract": "Camera-based 3D object detection in Bird's Eye View (BEV) is one of the most important perception tasks in autonomous driving. Earlier methods rely on dense BEV features, which are costly to construct. More recent works explore sparse query-based detection. However, they still require a large number of queries and can become expensive to run when more video frames are used. In this paper, we propose DySS, a novel method that employs state-space learning and dynamic queries. More specifically, DySS leverages a state-space model (SSM) to sequentially process the sampled features over time steps. In order to encourage the model to better capture the underlying motion and correspondence information, we introduce auxiliary tasks of future prediction and masked reconstruction to better train the SSM. The state of the SSM then provides an informative yet efficient summarization of the scene. Based on the state-space learned features, we dynamically update the queries via merge, remove, and split operations, which help maintain a useful, lean set of detection queries throughout the network. Our proposed DySS achieves both superior detection performance and efficient inference. Specifically, on the nuScenes test split, DySS achieves 65.31 NDS and 57.4 mAP, outperforming the latest state of the art. On the val split, DySS achieves 56.2 NDS and 46.2 mAP, as well as a real-time inference speed of 33 FPS.", "date": "2025-06-11", "categories": []} {"id": "2406.19048", "title": "BiCo-Fusion: Bidirectional Complementary LiDAR-Camera Fusion for Semantic- and Spatial-Aware 3D Object Detection", "abstract": "3D object detection is an important task that has been widely applied in autonomous driving. To perform this task, a new trend is to fuse multi-modal inputs, i.e., LiDAR and camera. Under such a trend, recent methods fuse these two modalities by unifying them in the same 3D space. However, during direct fusion in a unified space, the drawbacks of both modalities (LiDAR features struggle with detailed semantic information and the camera lacks accurate 3D spatial information) are also preserved, diluting semantic and spatial awareness of the final unified representation. To address the issue, this letter proposes a novel bidirectional complementary LiDAR-camera fusion framework, called BiCo-Fusion that can achieve robust semantic- and spatial-aware 3D object detection. The key insight is to fuse LiDAR and camera features in a bidirectional complementary way to enhance the semantic awareness of the LiDAR and the 3D spatial awareness of the camera. The enhanced features from both modalities are then adaptively fused to build a semantic- and spatial-aware unified representation. Specifically, we introduce Pre-Fusion consisting of a Voxel Enhancement Module (VEM) to enhance the semantic awareness of voxel features from 2D camera features and Image Enhancement Module (IEM) to enhance the 3D spatial awareness of camera features from 3D voxel features. We then introduce Unified Fusion (U-Fusion) to adaptively fuse the enhanced features from the last stage to build a unified representation. Extensive experiments demonstrate the superiority of our BiCo-Fusion against the prior arts. Project page: https://t-ys.github.io/BiCo-Fusion/.", "date": "2024-06-27", "categories": []} {"id": "2506.09541", "title": "3DGeoDet: General-purpose Geometry-aware Image-based 3D Object Detection", "abstract": "This paper proposes 3DGeoDet, a novel geometry-aware 3D object detection approach that effectively handles single- and multi-view RGB images in indoor and outdoor environments, showcasing its general-purpose applicability. The key challenge for image-based 3D object detection tasks is the lack of 3D geometric cues, which leads to ambiguity in establishing correspondences between images and 3D representations. To tackle this problem, 3DGeoDet generates efficient 3D geometric representations in both explicit and implicit manners based on predicted depth information. Specifically, we utilize the predicted depth to learn voxel occupancy and optimize the voxelized 3D feature volume explicitly through the proposed voxel occupancy attention. To further enhance 3D awareness, the feature volume is integrated with an implicit 3D representation, the truncated signed distance function (TSDF). Without requiring supervision from 3D signals, we significantly improve the model's comprehension of 3D geometry by leveraging intermediate 3D representations and achieve end-to-end training. Our approach surpasses the performance of state-of-the-art image-based methods on both single- and multi-view benchmark datasets across diverse environments, achieving a 9.3 mAP@0.5 improvement on the SUN RGB-D dataset, a 3.3 mAP@0.5 improvement on the ScanNetV2 dataset, and a 0.19 AP3D@0.7 improvement on the KITTI dataset. The project page is available at: https://cindy0725.github.io/3DGeoDet/.", "date": "2025-06-11", "categories": []} {"id": "2407.05769", "title": "Boosting 3D Object Detection with Semantic-Aware Multi-Branch Framework", "abstract": "In autonomous driving, LiDAR sensors are vital for acquiring 3D point clouds, providing reliable geometric information. However, traditional sampling methods of preprocessing often ignore semantic features, leading to detail loss and ground point interference in 3D object detection. To address this, we propose a multi-branch two-stage 3D object detection framework using a Semantic-aware Multi-branch Sampling (SMS) module and multi-view consistency constraints. The SMS module includes random sampling, Density Equalization Sampling (DES) for enhancing distant objects, and Ground Abandonment Sampling (GAS) to focus on non-ground points. The sampled multi-view points are processed through a Consistent KeyPoint Selection (CKPS) module to generate consistent keypoint masks for efficient proposal sampling. The first-stage detector uses multi-branch parallel learning with multi-view consistency loss for feature aggregation, while the second-stage detector fuses multi-view data through a Multi-View Fusion Pooling (MVFP) module to precisely predict 3D objects. The experimental results on the KITTI dataset and Waymo Open Dataset show that our method achieves excellent detection performance improvement for a variety of backbones, especially for low-performance backbones with the simple network structures.", "date": "2024-07-08", "categories": []} {"id": "2506.07375", "title": "DINO-CoDT: Multi-class Collaborative Detection and Tracking with Vision Foundation Models", "abstract": "Collaborative perception plays a crucial role in enhancing environmental understanding by expanding the perceptual range and improving robustness against sensor failures, which primarily involves collaborative 3D detection and tracking tasks. The former focuses on object recognition in individual frames, while the latter captures continuous instance tracklets over time. However, existing works in both areas predominantly focus on the vehicle superclass, lacking effective solutions for both multi-class collaborative detection and tracking. This limitation hinders their applicability in real-world scenarios, which involve diverse object classes with varying appearances and motion patterns. To overcome these limitations, we propose a multi-class collaborative detection and tracking framework tailored for diverse road users. We first present a detector with a global spatial attention fusion (GSAF) module, enhancing multi-scale feature learning for objects of varying sizes. Next, we introduce a tracklet RE-IDentification (REID) module that leverages visual semantics with a vision foundation model to effectively reduce ID SWitch (IDSW) errors, in cases of erroneous mismatches involving small objects like pedestrians. We further design a velocity-based adaptive tracklet management (VATM) module that adjusts the tracking interval dynamically based on object motion. Extensive experiments on the V2X-Real and OPV2V datasets show that our approach significantly outperforms existing state-of-the-art methods in both detection and tracking accuracy.", "date": "2025-06-09", "categories": []} {"id": "2504.12643", "title": "RoPETR: Improving Temporal Camera-Only 3D Detection by Integrating Enhanced Rotary Position Embedding", "abstract": "This technical report introduces a targeted improvement to the StreamPETR framework, specifically aimed at enhancing velocity estimation, a critical factor influencing the overall NuScenes Detection Score. While StreamPETR exhibits strong 3D bounding box detection performance as reflected by its high mean Average Precision our analysis identified velocity estimation as a substantial bottleneck when evaluated on the NuScenes dataset. To overcome this limitation, we propose a customized positional embedding strategy tailored to enhance temporal modeling capabilities. Experimental evaluations conducted on the NuScenes test set demonstrate that our improved approach achieves a state-of-the-art NDS of 70.86% using the ViT-L backbone, setting a new benchmark for camera-only 3D object detection.", "date": "2025-04-17", "categories": []} {"id": "2506.05780", "title": "Robust sensor fusion against on-vehicle sensor staleness", "abstract": "Sensor fusion is crucial for a performant and robust Perception system in autonomous vehicles, but sensor staleness, where data from different sensors arrives with varying delays, poses significant challenges. Temporal misalignment between sensor modalities leads to inconsistent object state estimates, severely degrading the quality of trajectory predictions that are critical for safety. We present a novel and model-agnostic approach to address this problem via (1) a per-point timestamp offset feature (for LiDAR and radar both relative to camera) that enables fine-grained temporal awareness in sensor fusion, and (2) a data augmentation strategy that simulates realistic sensor staleness patterns observed in deployed vehicles. Our method is integrated into a perspective-view detection model that consumes sensor data from multiple LiDARs, radars and cameras. We demonstrate that while a conventional model shows significant regressions when one sensor modality is stale, our approach reaches consistently good performance across both synchronized and stale conditions.", "date": "2025-06-06", "categories": []} {"id": "2506.03714", "title": "FSHNet: Fully Sparse Hybrid Network for 3D Object Detection", "abstract": "Fully sparse 3D detectors have recently gained significant attention due to their efficiency in long-range detection. However, sparse 3D detectors extract features only from non-empty voxels, which impairs long-range interactions and causes the center feature missing. The former weakens the feature extraction capability, while the latter hinders network optimization. To address these challenges, we introduce the Fully Sparse Hybrid Network (FSHNet). FSHNet incorporates a proposed SlotFormer block to enhance the long-range feature extraction capability of existing sparse encoders. The SlotFormer divides sparse voxels using a slot partition approach, which, compared to traditional window partition, provides a larger receptive field. Additionally, we propose a dynamic sparse label assignment strategy to deeply optimize the network by providing more high-quality positive samples. To further enhance performance, we introduce a sparse upsampling module to refine downsampled voxels, preserving fine-grained details crucial for detecting small objects. Extensive experiments on the Waymo, nuScenes, and Argoverse2 benchmarks demonstrate the effectiveness of FSHNet. The code is available at https://github.com/Say2L/FSHNet.", "date": "2025-06-04", "categories": []} {"id": "2505.23756", "title": "Rooms from Motion: Un-posed Indoor 3D Object Detection as Localization and Mapping", "abstract": "We revisit scene-level 3D object detection as the output of an object-centric framework capable of both localization and mapping using 3D oriented boxes as the underlying geometric primitive. While existing 3D object detection approaches operate globally and implicitly rely on the a priori existence of metric camera poses, our method, Rooms from Motion (RfM) operates on a collection of un-posed images. By replacing the standard 2D keypoint-based matcher of structure-from-motion with an object-centric matcher based on image-derived 3D boxes, we estimate metric camera poses, object tracks, and finally produce a global, semantic 3D object map. When a priori pose is available, we can significantly improve map quality through optimization of global 3D boxes against individual observations. RfM shows strong localization performance and subsequently produces maps of higher quality than leading point-based and multi-view 3D object detection methods on CA-1M and ScanNet++, despite these global methods relying on overparameterization through point clouds or dense volumes. Rooms from Motion achieves a general, object-centric representation which not only extends the work of Cubify Anything to full scenes but also allows for inherently sparse localization and parametric mapping proportional to the number of objects in a scene.", "date": "2025-05-29", "categories": []} {"id": "2401.03846", "title": "UOD: Unseen Object Detection in 3D Point Cloud", "abstract": "Existing 3D object detectors encounter extreme challenges in localizing unseen 3D objects and recognizing them as unseen, which is a crucial technology in autonomous driving in the wild. To address these challenges, we propose practical methods to enhance the performance of 3D detection and Out-Of-Distribution (OOD) classification for unseen objects. The proposed methods include anomaly sample augmentation, learning of universal objectness, learning of detecting unseen objects, and learning of distinguishing unseen objects. To demonstrate the effectiveness of our approach, we propose the KITTI Misc benchmark and two additional synthetic OOD benchmarks: the Nuscenes OOD benchmark and the SUN-RGBD OOD benchmark. The proposed methods consistently enhance performance by a large margin across all existing methods, giving insight for future work on unseen 3D object detection in the wild.", "date": "2024-01-08", "categories": []} {"id": "2503.17168", "title": "Hi-ALPS -- An Experimental Robustness Quantification of Six LiDAR-based Object Detection Systems for Autonomous Driving", "abstract": "Light Detection and Ranging (LiDAR) is an essential sensor technology for autonomous driving as it can capture high-resolution 3D data. As 3D object detection systems (OD) can interpret such point cloud data, they play a key role in the driving decisions of autonomous vehicles. Consequently, such 3D OD must be robust against all types of perturbations and must therefore be extensively tested. One approach is the use of adversarial examples, which are small, sometimes sophisticated perturbations in the input data that change, i.e., falsify, the prediction of the OD. These perturbations are carefully designed based on the weaknesses of the OD. The robustness of the OD cannot be quantified with adversarial examples in general, because if the OD is vulnerable to a given attack, it is unclear whether this is due to the robustness of the OD or whether the attack algorithm produces particularly strong adversarial examples. The contribution of this work is Hi-ALPS -- Hierarchical Adversarial-example-based LiDAR Perturbation Level System, where higher robustness of the OD is required to withstand the perturbations as the perturbation levels increase. In doing so, the Hi-ALPS levels successively implement a heuristic followed by established adversarial example approaches. In a series of comprehensive experiments using Hi-ALPS, we quantify the robustness of six state-of-the-art 3D OD under different types of perturbations. The results of the experiments show that none of the OD is robust against all Hi-ALPS levels; an important factor for the ranking is that human observers can still correctly recognize the perturbed objects, as the respective perturbations are small. To increase the robustness of the OD, we discuss the applicability of state-of-the-art countermeasures. In addition, we derive further suggestions for countermeasures based on our experimental results.", "date": "2025-03-21", "categories": []} {"id": "2502.06114", "title": "Enhanced 3D Object Detection via Diverse Feature Representations of 4D Radar Tensor", "abstract": "Recent advances in automotive four-dimensional (4D) Radar have enabled access to raw 4D Radar Tensor (4DRT), offering richer spatial and Doppler information than conventional point clouds. While most existing methods rely on heavily pre-processed, sparse Radar data, recent attempts to leverage raw 4DRT face high computational costs and limited scalability. To address these limitations, we propose a novel three-dimensional (3D) object detection framework that maximizes the utility of 4DRT while preserving efficiency. Our method introduces a multi-teacher knowledge distillation (KD), where multiple teacher models are trained on point clouds derived from diverse 4DRT pre-processing techniques, each capturing complementary signal characteristics. These teacher representations are fused via a dedicated aggregation module and distilled into a lightweight student model that operates solely on a sparse Radar input. Experimental results on the K-Radar dataset demonstrate that our framework achieves improvements of 7.3% in AP_3D and 9.5% in AP_BEV over the baseline RTNH model when using extremely sparse inputs. Furthermore, it attains comparable performance to denser-input baselines while significantly reducing the input data size by about 90 times, confirming the scalability and efficiency of our approach.", "date": "2025-02-10", "categories": []} {"id": "2505.16029", "title": "Learning better representations for crowded pedestrians in offboard LiDAR-camera 3D tracking-by-detection", "abstract": "Perceiving pedestrians in highly crowded urban environments is a difficult long-tail problem for learning-based autonomous perception. Speeding up 3D ground truth generation for such challenging scenes is performance-critical yet very challenging. The difficulties include the sparsity of the captured pedestrian point cloud and a lack of suitable benchmarks for a specific system design study. To tackle the challenges, we first collect a new multi-view LiDAR-camera 3D multiple-object-tracking benchmark of highly crowded pedestrians for in-depth analysis. We then build an offboard auto-labeling system that reconstructs pedestrian trajectories from LiDAR point cloud and multi-view images. To improve the generalization power for crowded scenes and the performance for small objects, we propose to learn high-resolution representations that are density-aware and relationship-aware. Extensive experiments validate that our approach significantly improves the 3D pedestrian tracking performance towards higher auto-labeling efficiency. The code will be publicly available at this HTTP URL.", "date": "2025-05-21", "categories": []} {"id": "2505.13266", "title": "DB3D-L: Depth-aware BEV Feature Transformation for Accurate 3D Lane Detection", "abstract": "3D Lane detection plays an important role in autonomous driving. Recent advances primarily build Birds-Eye-View (BEV) feature from front-view (FV) images to perceive 3D information of Lane more effectively. However, constructing accurate BEV information from FV image is limited due to the lacking of depth information, causing previous works often rely heavily on the assumption of a flat ground plane. Leveraging monocular depth estimation to assist in constructing BEV features is less constrained, but existing methods struggle to effectively integrate the two tasks. To address the above issue, in this paper, an accurate 3D lane detection method based on depth-aware BEV feature transtormation is proposed. In detail, an effective feature extraction module is designed, in which a Depth Net is integrated to obtain the vital depth information for 3D perception, thereby simplifying the complexity of view transformation. Subquently a feature reduce module is proposed to reduce height dimension of FV features and depth features, thereby enables effective fusion of crucial FV features and depth features. Then a fusion module is designed to build BEV feature from prime FV feature and depth information. The proposed method performs comparably with state-of-the-art methods on both synthetic Apollo, realistic OpenLane datasets.", "date": "2025-05-19", "categories": []} {"id": "2505.09422", "title": "MoRAL: Motion-aware Multi-Frame 4D Radar and LiDAR Fusion for Robust 3D Object Detection", "abstract": "Reliable autonomous driving systems require accurate detection of traffic participants. To this end, multi-modal fusion has emerged as an effective strategy. In particular, 4D radar and LiDAR fusion methods based on multi-frame radar point clouds have demonstrated the effectiveness in bridging the point density gap. However, they often neglect radar point clouds' inter-frame misalignment caused by object movement during accumulation and do not fully exploit the object dynamic information from 4D radar. In this paper, we propose MoRAL, a motion-aware multi-frame 4D radar and LiDAR fusion framework for robust 3D object detection. First, a Motion-aware Radar Encoder (MRE) is designed to compensate for inter-frame radar misalignment from moving objects. Later, a Motion Attention Gated Fusion (MAGF) module integrate radar motion features to guide LiDAR features to focus on dynamic foreground objects. Extensive evaluations on the View-of-Delft (VoD) dataset demonstrate that MoRAL outperforms existing methods, achieving the highest mAP of 73.30% in the entire area and 88.68% in the driving corridor. Notably, our method also achieves the best AP of 69.67% for pedestrians in the entire area and 96.25% for cyclists in the driving corridor.", "date": "2025-05-14", "categories": []} {"id": "2505.07398", "title": "DepthFusion: Depth-Aware Hybrid Feature Fusion for LiDAR-Camera 3D Object Detection", "abstract": "State-of-the-art LiDAR-camera 3D object detectors usually focus on feature fusion. However, they neglect the factor of depth while designing the fusion strategy. In this work, we are the first to observe that different modalities play different roles as depth varies via statistical analysis and visualization. Based on this finding, we propose a Depth-Aware Hybrid Feature Fusion (DepthFusion) strategy that guides the weights of point cloud and RGB image modalities by introducing depth encoding at both global and local levels. Specifically, the Depth-GFusion module adaptively adjusts the weights of image Bird's-Eye-View (BEV) features in multi-modal global features via depth encoding. Furthermore, to compensate for the information lost when transferring raw features to the BEV space, we propose a Depth-LFusion module, which adaptively adjusts the weights of original voxel features and multi-view image features in multi-modal local features via depth encoding. Extensive experiments on the nuScenes and KITTI datasets demonstrate that our DepthFusion method surpasses previous state-of-the-art methods. Moreover, our DepthFusion is more robust to various kinds of corruptions, outperforming previous methods on the nuScenes-C dataset.", "date": "2025-05-12", "categories": []} {"id": "2505.05397", "title": "PillarMamba: Learning Local-Global Context for Roadside Point Cloud via\n Hybrid State Space Model", "abstract": " Serving the Intelligent Transport System (ITS) and Vehicle-to-Everything\n(V2X) tasks, roadside perception has received increasing attention in recent\nyears, as it can extend the perception range of connected vehicles and improve\ntraffic safety. However, roadside point cloud oriented 3D object detection has\nnot been effectively explored. To some extent, the key to the performance of a\npoint cloud detector lies in the receptive field of the network and the ability\nto effectively utilize the scene context. The recent emergence of Mamba, based\non State Space Model (SSM), has shaken up the traditional convolution and\ntransformers that have long been the foundational building blocks, due to its\nefficient global receptive field. In this work, we introduce Mamba to\npillar-based roadside point cloud perception and propose a framework based on\nCross-stage State-space Group (CSG), called PillarMamba. It enhances the\nexpressiveness of the network and achieves efficient computation through\ncross-stage feature fusion. However, due to the limitations of scan directions,\nstate space model faces local connection disrupted and historical relationship\nforgotten. To address this, we propose the Hybrid State-space Block (HSB) to\nobtain the local-global context of roadside point cloud. Specifically, it\nenhances neighborhood connections through local convolution and preserves\nhistorical memory through residual attention. The proposed method outperforms\nthe state-of-the-art methods on the popular large scale roadside benchmark:\nDAIR-V2X-I. The code will be released soon.\n", "date": "2025-05-08", "categories": []} {"id": "2505.00757", "title": "Efficient On-Chip Implementation of 4D Radar-Based 3D Object Detection\n on Hailo-8L", "abstract": " 4D radar has attracted attention in autonomous driving due to its ability to\nenable robust 3D object detection even under adverse weather conditions. To\npractically deploy such technologies, it is essential to achieve real-time\nprocessing within low-power embedded environments. Addressing this, we present\nthe first on-chip implementation of a 4D radar-based 3D object detection model\non the Hailo-8L AI accelerator. Although conventional 3D convolutional neural\nnetwork (CNN) architectures require 5D inputs, the Hailo-8L only supports 4D\ntensors, posing a significant challenge. To overcome this limitation, we\nintroduce a tensor transformation method that reshapes 5D inputs into 4D\nformats during the compilation process, enabling direct deployment without\naltering the model structure. The proposed system achieves 46.47% AP_3D and\n52.75% AP_BEV, maintaining comparable accuracy to GPU-based models while\nachieving an inference speed of 13.76 Hz. These results demonstrate the\napplicability of 4D radar-based perception technologies to autonomous driving\nsystems.\n", "date": "2025-05-01", "categories": []} {"id": "2505.00511", "title": "Inconsistency-based Active Learning for LiDAR Object Detection", "abstract": " Deep learning models for object detection in autonomous driving have recently\nachieved impressive performance gains and are already being deployed in\nvehicles worldwide. However, current models require increasingly large datasets\nfor training. Acquiring and labeling such data is costly, necessitating the\ndevelopment of new strategies to optimize this process. Active learning is a\npromising approach that has been extensively researched in the image domain. In\nour work, we extend this concept to the LiDAR domain by developing several\ninconsistency-based sample selection strategies and evaluate their\neffectiveness in various settings. Our results show that using a naive\ninconsistency approach based on the number of detected boxes, we achieve the\nsame mAP as the random sampling strategy with 50% of the labeled data.\n", "date": "2025-05-01", "categories": []} {"id": "2504.11014", "title": "GATE3D: Generalized Attention-based Task-synergized Estimation in 3D*", "abstract": " The emerging trend in computer vision emphasizes developing universal models\ncapable of simultaneously addressing multiple diverse tasks. Such universality\ntypically requires joint training across multi-domain datasets to ensure\neffective generalization. However, monocular 3D object detection presents\nunique challenges in multi-domain training due to the scarcity of datasets\nannotated with accurate 3D ground-truth labels, especially beyond typical\nroad-based autonomous driving contexts. To address this challenge, we introduce\na novel weakly supervised framework leveraging pseudo-labels. Current\npretrained models often struggle to accurately detect pedestrians in non-road\nenvironments due to inherent dataset biases. Unlike generalized image-based 2D\nobject detection models, achieving similar generalization in monocular 3D\ndetection remains largely unexplored. In this paper, we propose GATE3D, a novel\nframework designed specifically for generalized monocular 3D object detection\nvia weak supervision. GATE3D effectively bridges domain gaps by employing\nconsistency losses between 2D and 3D predictions. Remarkably, our model\nachieves competitive performance on the KITTI benchmark as well as on an\nindoor-office dataset collected by us to evaluate the generalization\ncapabilities of our framework. Our results demonstrate that GATE3D\nsignificantly accelerates learning from limited annotated data through\neffective pre-training strategies, highlighting substantial potential for\nbroader impacts in robotics, augmented reality, and virtual reality\napplications. Project page: https://ies0411.github.io/GATE3D/\n", "date": "2025-04-15", "categories": []} {"id": "2504.18738", "title": "A Review of 3D Object Detection with Vision-Language Models", "abstract": " This review provides a systematic analysis of comprehensive survey of 3D\nobject detection with vision-language models(VLMs) , a rapidly advancing area\nat the intersection of 3D vision and multimodal AI. By examining over 100\nresearch papers, we provide the first systematic analysis dedicated to 3D\nobject detection with vision-language models. We begin by outlining the unique\nchallenges of 3D object detection with vision-language models, emphasizing\ndifferences from 2D detection in spatial reasoning and data complexity.\nTraditional approaches using point clouds and voxel grids are compared to\nmodern vision-language frameworks like CLIP and 3D LLMs, which enable\nopen-vocabulary detection and zero-shot generalization. We review key\narchitectures, pretraining strategies, and prompt engineering methods that\nalign textual and 3D features for effective 3D object detection with\nvision-language models. Visualization examples and evaluation benchmarks are\ndiscussed to illustrate performance and behavior. Finally, we highlight current\nchallenges, such as limited 3D-language datasets and computational demands, and\npropose future research directions to advance 3D object detection with\nvision-language models. >Object Detection, Vision-Language Models, Agents,\nVLMs, LLMs, AI\n", "date": "2025-04-25", "categories": []} {"id": "2504.19115", "title": "Towards Latency-Aware 3D Streaming Perception for Autonomous Driving", "abstract": " Although existing 3D perception algorithms have demonstrated significant\nimprovements in performance, their deployment on edge devices continues to\nencounter critical challenges due to substantial runtime latency. We propose a\nnew benchmark tailored for online evaluation by considering runtime latency.\nBased on the benchmark, we build a Latency-Aware 3D Streaming Perception (LASP)\nframework that addresses the latency issue through two primary components: 1)\nlatency-aware history integration, which extends query propagation into a\ncontinuous process, ensuring the integration of historical feature regardless\nof varying latency; 2) latency-aware predictive detection, a module that\ncompensates the detection results with the predicted trajectory and the\nposterior accessed latency. By incorporating the latency-aware mechanism, our\nmethod shows generalization across various latency levels, achieving an online\nperformance that closely aligns with 80\\% of its offline evaluation on the\nJetson AGX Orin without any acceleration techniques.\n", "date": "2025-04-27", "categories": []} {"id": "2504.18203", "title": "LiDAR-Guided Monocular 3D Object Detection for Long-Range Railway\n Monitoring", "abstract": " Railway systems, particularly in Germany, require high levels of automation\nto address legacy infrastructure challenges and increase train traffic safely.\nA key component of automation is robust long-range perception, essential for\nearly hazard detection, such as obstacles at level crossings or pedestrians on\ntracks. Unlike automotive systems with braking distances of ~70 meters, trains\nrequire perception ranges exceeding 1 km. This paper presents an\ndeep-learning-based approach for long-range 3D object detection tailored for\nautonomous trains. The method relies solely on monocular images, inspired by\nthe Faraway-Frustum approach, and incorporates LiDAR data during training to\nimprove depth estimation. The proposed pipeline consists of four key modules:\n(1) a modified YOLOv9 for 2.5D object detection, (2) a depth estimation\nnetwork, and (3-4) dedicated short- and long-range 3D detection heads.\nEvaluations on the OSDaR23 dataset demonstrate the effectiveness of the\napproach in detecting objects up to 250 meters. Results highlight its potential\nfor railway automation and outline areas for future improvement.\n", "date": "2025-04-25", "categories": []} {"id": "2504.16368", "title": "Revisiting Radar Camera Alignment by Contrastive Learning for 3D Object\n Detection", "abstract": " Recently, 3D object detection algorithms based on radar and camera fusion\nhave shown excellent performance, setting the stage for their application in\nautonomous driving perception tasks. Existing methods have focused on dealing\nwith feature misalignment caused by the domain gap between radar and camera.\nHowever, existing methods either neglect inter-modal features interaction\nduring alignment or fail to effectively align features at the same spatial\nlocation across modalities. To alleviate the above problems, we propose a new\nalignment model called Radar Camera Alignment (RCAlign). Specifically, we\ndesign a Dual-Route Alignment (DRA) module based on contrastive learning to\nalign and fuse the features between radar and camera. Moreover, considering the\nsparsity of radar BEV features, a Radar Feature Enhancement (RFE) module is\nproposed to improve the densification of radar BEV features with the knowledge\ndistillation loss. Experiments show RCAlign achieves a new state-of-the-art on\nthe public nuScenes benchmark in radar camera fusion for 3D Object Detection.\nFurthermore, the RCAlign achieves a significant performance gain (4.3\\% NDS and\n8.4\\% mAP) in real-time 3D detection compared to the latest state-of-the-art\nmethod (RCBEVDet).\n", "date": "2025-04-23", "categories": []} {"id": "2504.16740", "title": "Gaussian Splatting is an Effective Data Generator for 3D Object\n Detection", "abstract": " We investigate data augmentation for 3D object detection in autonomous\ndriving. We utilize recent advancements in 3D reconstruction based on Gaussian\nSplatting for 3D object placement in driving scenes. Unlike existing\ndiffusion-based methods that synthesize images conditioned on BEV layouts, our\napproach places 3D objects directly in the reconstructed 3D space with\nexplicitly imposed geometric transformations. This ensures both the physical\nplausibility of object placement and highly accurate 3D pose and position\nannotations.\n Our experiments demonstrate that even by integrating a limited number of\nexternal 3D objects into real scenes, the augmented data significantly enhances\n3D object detection performance and outperforms existing diffusion-based 3D\naugmentation for object detection. Extensive testing on the nuScenes dataset\nreveals that imposing high geometric diversity in object placement has a\ngreater impact compared to the appearance diversity of objects. Additionally,\nwe show that generating hard examples, either by maximizing detection loss or\nimposing high visual occlusion in camera images, does not lead to more\nefficient 3D data augmentation for camera-based 3D object detection in\nautonomous driving.\n", "date": "2025-04-23", "categories": []} {"id": "2504.13297", "title": "Weak Cube R-CNN: Weakly Supervised 3D Detection using only 2D Bounding\n Boxes", "abstract": " Monocular 3D object detection is an essential task in computer vision, and it\nhas several applications in robotics and virtual reality. However, 3D object\ndetectors are typically trained in a fully supervised way, relying extensively\non 3D labeled data, which is labor-intensive and costly to annotate. This work\nfocuses on weakly-supervised 3D detection to reduce data needs using a\nmonocular method that leverages a singlecamera system over expensive LiDAR\nsensors or multi-camera setups. We propose a general model Weak Cube R-CNN,\nwhich can predict objects in 3D at inference time, requiring only 2D box\nannotations for training by exploiting the relationship between 2D projections\nof 3D cubes. Our proposed method utilizes pre-trained frozen foundation 2D\nmodels to estimate depth and orientation information on a training set. We use\nthese estimated values as pseudo-ground truths during training. We design loss\nfunctions that avoid 3D labels by incorporating information from the external\nmodels into the loss. In this way, we aim to implicitly transfer knowledge from\nthese large foundation 2D models without having access to 3D bounding box\nannotations. Experimental results on the SUN RGB-D dataset show increased\nperformance in accuracy compared to an annotation time equalized Cube R-CNN\nbaseline. While not precise for centimetre-level measurements, this method\nprovides a strong foundation for further research.\n", "date": "2025-04-17", "categories": []} {"id": "2308.06985", "title": "PatchContrast: Self-Supervised Pre-training for 3D Object Detection", "abstract": " Accurately detecting objects in the environment is a key challenge for\nautonomous vehicles. However, obtaining annotated data for detection is\nexpensive and time-consuming. We introduce PatchContrast, a novel\nself-supervised point cloud pre-training framework for 3D object detection. We\npropose to utilize two levels of abstraction to learn discriminative\nrepresentation from unlabeled data: proposal-level and patch-level. The\nproposal-level aims at localizing objects in relation to their surroundings,\nwhereas the patch-level adds information about the internal connections between\nthe object's components, hence distinguishing between different objects based\non their individual components. We demonstrate how these levels can be\nintegrated into self-supervised pre-training for various backbones to enhance\nthe downstream 3D detection task. We show that our method outperforms existing\nstate-of-the-art models on three commonly-used 3D detection datasets.\n", "date": "2023-08-14", "categories": []} {"id": "2504.09086", "title": "RICCARDO: Radar Hit Prediction and Convolution for Camera-Radar 3D\n Object Detection", "abstract": " Radar hits reflect from points on both the boundary and internal to object\noutlines. This results in a complex distribution of radar hits that depends on\nfactors including object category, size, and orientation. Current radar-camera\nfusion methods implicitly account for this with a black-box neural network. In\nthis paper, we explicitly utilize a radar hit distribution model to assist\nfusion. First, we build a model to predict radar hit distributions conditioned\non object properties obtained from a monocular detector. Second, we use the\npredicted distribution as a kernel to match actual measured radar points in the\nneighborhood of the monocular detections, generating matching scores at nearby\npositions. Finally, a fusion stage combines context with the kernel detector to\nrefine the matching scores. Our method achieves the state-of-the-art\nradar-camera detection performance on nuScenes. Our source code is available at\nhttps://github.com/longyunf/riccardo.\n", "date": "2025-04-12", "categories": []} {"id": "2504.06801", "title": "MonoPlace3D: Learning 3D-Aware Object Placement for 3D Monocular\n Detection", "abstract": " Current monocular 3D detectors are held back by the limited diversity and\nscale of real-world datasets. While data augmentation certainly helps, it's\nparticularly difficult to generate realistic scene-aware augmented data for\noutdoor settings. Most current approaches to synthetic data generation focus on\nrealistic object appearance through improved rendering techniques. However, we\nshow that where and how objects are positioned is just as crucial for training\neffective 3D monocular detectors. The key obstacle lies in automatically\ndetermining realistic object placement parameters - including position,\ndimensions, and directional alignment when introducing synthetic objects into\nactual scenes. To address this, we introduce MonoPlace3D, a novel system that\nconsiders the 3D scene content to create realistic augmentations. Specifically,\ngiven a background scene, MonoPlace3D learns a distribution over plausible 3D\nbounding boxes. Subsequently, we render realistic objects and place them\naccording to the locations sampled from the learned distribution. Our\ncomprehensive evaluation on two standard datasets KITTI and NuScenes,\ndemonstrates that MonoPlace3D significantly improves the accuracy of multiple\nexisting monocular 3D detectors while being highly data efficient.\n", "date": "2025-04-09", "categories": []} {"id": "2504.07441", "title": "WS-DETR: Robust Water Surface Object Detection through Vision-Radar\n Fusion with Detection Transformer", "abstract": " Robust object detection for Unmanned Surface Vehicles (USVs) in complex water\nenvironments is essential for reliable navigation and operation. Specifically,\nwater surface object detection faces challenges from blurred edges and diverse\nobject scales. Although vision-radar fusion offers a feasible solution,\nexisting approaches suffer from cross-modal feature conflicts, which negatively\naffect model robustness. To address this problem, we propose a robust\nvision-radar fusion model WS-DETR. In particular, we first introduce a\nMulti-Scale Edge Information Integration (MSEII) module to enhance edge\nperception and a Hierarchical Feature Aggregator (HiFA) to boost multi-scale\nobject detection in the encoder. Then, we adopt self-moving point\nrepresentations for continuous convolution and residual connection to\nefficiently extract irregular features under the scenarios of irregular point\ncloud data. To further mitigate cross-modal conflicts, an Adaptive Feature\nInteractive Fusion (AFIF) module is introduced to integrate visual and radar\nfeatures through geometric alignment and semantic fusion. Extensive experiments\non the WaterScenes dataset demonstrate that WS-DETR achieves state-of-the-art\n(SOTA) performance, maintaining its superiority even under adverse weather and\nlighting conditions.\n", "date": "2025-04-10", "categories": []} {"id": "2405.10577", "title": "DuoSpaceNet: Leveraging Both Bird's-Eye-View and Perspective View\n Representations for 3D Object Detection", "abstract": " Multi-view camera-only 3D object detection largely follows two primary\nparadigms: exploiting bird's-eye-view (BEV) representations or focusing on\nperspective-view (PV) features, each with distinct advantages. Although several\nrecent approaches explore combining BEV and PV, many rely on partial fusion or\nmaintain separate detection heads. In this paper, we propose DuoSpaceNet, a\nnovel framework that fully unifies BEV and PV feature spaces within a single\ndetection pipeline for comprehensive 3D perception. Our design includes a\ndecoder to integrate BEV and PV features into unified detection queries, as\nwell as a feature enhancement strategy that enriches different feature\nrepresentations. In addition, DuoSpaceNet can be extended to handle multi-frame\ninputs, enabling more robust temporal analysis. Extensive experiments on\nnuScenes dataset show that DuoSpaceNet surpasses both BEV-based baselines\n(e.g., BEVFormer) and PV-based baselines (e.g., Sparse4D) in 3D object\ndetection and BEV map segmentation, verifying the effectiveness of our proposed\ndesign.\n", "date": "2024-05-17", "categories": []} {"id": "2504.05649", "title": "POD: Predictive Object Detection with Single-Frame FMCW LiDAR Point\n Cloud", "abstract": " LiDAR-based 3D object detection is a fundamental task in the field of\nautonomous driving. This paper explores the unique advantage of Frequency\nModulated Continuous Wave (FMCW) LiDAR in autonomous perception. Given a single\nframe FMCW point cloud with radial velocity measurements, we expect that our\nobject detector can detect the short-term future locations of objects using\nonly the current frame sensor data and demonstrate a fast ability to respond to\nintermediate danger. To achieve this, we extend the standard object detection\ntask to a novel task named predictive object detection (POD), which aims to\npredict the short-term future location and dimensions of objects based solely\non current observations. Typically, a motion prediction task requires\nhistorical sensor information to process the temporal contexts of each object,\nwhile our detector's avoidance of multi-frame historical information enables a\nmuch faster response time to potential dangers. The core advantage of FMCW\nLiDAR lies in the radial velocity associated with every reflected point. We\npropose a novel POD framework, the core idea of which is to generate a virtual\nfuture point using a ray casting mechanism, create virtual two-frame point\nclouds with the current and virtual future frames, and encode these two-frame\nvoxel features with a sparse 4D encoder. Subsequently, the 4D voxel features\nare separated by temporal indices and remapped into two Bird's Eye View (BEV)\nfeatures: one decoded for standard current frame object detection and the other\nfor future predictive object detection. Extensive experiments on our in-house\ndataset demonstrate the state-of-the-art standard and predictive detection\nperformance of the proposed POD framework.\n", "date": "2025-04-08", "categories": []} {"id": "2504.03438", "title": "ZFusion: An Effective Fuser of Camera and 4D Radar for 3D Object\n Perception in Autonomous Driving", "abstract": " Reliable 3D object perception is essential in autonomous driving. Owing to\nits sensing capabilities in all weather conditions, 4D radar has recently\nreceived much attention. However, compared to LiDAR, 4D radar provides much\nsparser point cloud. In this paper, we propose a 3D object detection method,\ntermed ZFusion, which fuses 4D radar and vision modality. As the core of\nZFusion, our proposed FP-DDCA (Feature Pyramid-Double Deformable Cross\nAttention) fuser complements the (sparse) radar information and (dense) vision\ninformation, effectively. Specifically, with a feature-pyramid structure, the\nFP-DDCA fuser packs Transformer blocks to interactively fuse multi-modal\nfeatures at different scales, thus enhancing perception accuracy. In addition,\nwe utilize the Depth-Context-Split view transformation module due to the\nphysical properties of 4D radar. Considering that 4D radar has a much lower\ncost than LiDAR, ZFusion is an attractive alternative to LiDAR-based methods.\nIn typical traffic scenarios like the VoD (View-of-Delft) dataset, experiments\nshow that with reasonable inference speed, ZFusion achieved the\nstate-of-the-art mAP (mean average precision) in the region of interest, while\nhaving competitive mAP in the entire area compared to the baseline methods,\nwhich demonstrates performance close to LiDAR and greatly outperforms those\ncamera-only methods.\n", "date": "2025-04-04", "categories": []} {"id": "2504.03654", "title": "PointSplit: Towards On-device 3D Object Detection with Heterogeneous\n Low-power Accelerators", "abstract": " Running deep learning models on resource-constrained edge devices has drawn\nsignificant attention due to its fast response, privacy preservation, and\nrobust operation regardless of Internet connectivity. While these devices\nalready cope with various intelligent tasks, the latest edge devices that are\nequipped with multiple types of low-power accelerators (i.e., both mobile GPU\nand NPU) can bring another opportunity; a task that used to be too heavy for an\nedge device in the single-accelerator world might become viable in the upcoming\nheterogeneous-accelerator world.To realize the potential in the context of 3D\nobject detection, we identify several technical challenges and propose\nPointSplit, a novel 3D object detection framework for multi-accelerator edge\ndevices that addresses the problems. Specifically, our PointSplit design\nincludes (1) 2D semantics-aware biased point sampling, (2) parallelized 3D\nfeature extraction, and (3) role-based group-wise quantization. We implement\nPointSplit on TensorFlow Lite and evaluate it on a customized hardware platform\ncomprising both mobile GPU and EdgeTPU. Experimental results on representative\nRGB-D datasets, SUN RGB-D and Scannet V2, demonstrate that PointSplit on a\nmulti-accelerator device is 24.7 times faster with similar accuracy compared to\nthe full-precision, 2D-3D fusion-based 3D detector on a GPU-only device.\n", "date": "2025-03-04", "categories": []} {"id": "2504.05170", "title": "SSLFusion: Scale & Space Aligned Latent Fusion Model for Multimodal 3D\n Object Detection", "abstract": " Multimodal 3D object detection based on deep neural networks has indeed made\nsignificant progress. However, it still faces challenges due to the\nmisalignment of scale and spatial information between features extracted from\n2D images and those derived from 3D point clouds. Existing methods usually\naggregate multimodal features at a single stage. However, leveraging\nmulti-stage cross-modal features is crucial for detecting objects of various\nscales. Therefore, these methods often struggle to integrate features across\ndifferent scales and modalities effectively, thereby restricting the accuracy\nof detection. Additionally, the time-consuming Query-Key-Value-based\n(QKV-based) cross-attention operations often utilized in existing methods aid\nin reasoning the location and existence of objects by capturing non-local\ncontexts. However, this approach tends to increase computational complexity. To\naddress these challenges, we present SSLFusion, a novel Scale & Space Aligned\nLatent Fusion Model, consisting of a scale-aligned fusion strategy (SAF), a\n3D-to-2D space alignment module (SAM), and a latent cross-modal fusion module\n(LFM). SAF mitigates scale misalignment between modalities by aggregating\nfeatures from both images and point clouds across multiple levels. SAM is\ndesigned to reduce the inter-modal gap between features from images and point\nclouds by incorporating 3D coordinate information into 2D image features.\nAdditionally, LFM captures cross-modal non-local contexts in the latent space\nwithout utilizing the QKV-based attention operations, thus mitigating\ncomputational complexity. Experiments on the KITTI and DENSE datasets\ndemonstrate that our SSLFusion outperforms state-of-the-art methods. Our\napproach obtains an absolute gain of 2.15% in 3D AP, compared with the\nstate-of-art method GraphAlign on the moderate level of the KITTI test set.\n", "date": "2025-04-07", "categories": []} {"id": "2504.02920", "title": "LiDAR-based Object Detection with Real-time Voice Specifications", "abstract": " This paper presents a LiDAR-based object detection system with real-time\nvoice specifications, integrating KITTI's 3D point clouds and RGB images\nthrough a multi-modal PointNet framework. It achieves 87.0% validation accuracy\non a 3000-sample subset, surpassing a 200-sample baseline of 67.5% by combining\nspatial and visual data, addressing class imbalance with weighted loss, and\nrefining training via adaptive techniques. A Tkinter prototype provides natural\nIndian male voice output using Edge TTS (en-IN-PrabhatNeural), alongside 3D\nvisualizations and real-time feedback, enhancing accessibility and safety in\nautonomous navigation, assistive technology, and beyond. The study offers a\ndetailed methodology, comprehensive experimental analysis, and a broad review\nof applications and challenges, establishing this work as a scalable\nadvancement in human-computer interaction and environmental perception, aligned\nwith current research trends.\n", "date": "2025-04-03", "categories": []} {"id": "2504.03563", "title": "PF3Det: A Prompted Foundation Feature Assisted Visual LiDAR 3D Detector", "abstract": " 3D object detection is crucial for autonomous driving, leveraging both LiDAR\npoint clouds for precise depth information and camera images for rich semantic\ninformation. Therefore, the multi-modal methods that combine both modalities\noffer more robust detection results. However, efficiently fusing LiDAR points\nand images remains challenging due to the domain gaps. In addition, the\nperformance of many models is limited by the amount of high quality labeled\ndata, which is expensive to create. The recent advances in foundation models,\nwhich use large-scale pre-training on different modalities, enable better\nmulti-modal fusion. Combining the prompt engineering techniques for efficient\ntraining, we propose the Prompted Foundational 3D Detector (PF3Det), which\nintegrates foundation model encoders and soft prompts to enhance LiDAR-camera\nfeature fusion. PF3Det achieves the state-of-the-art results under limited\ntraining data, improving NDS by 1.19% and mAP by 2.42% on the nuScenes dataset,\ndemonstrating its efficiency in 3D detection.\n", "date": "2025-04-04", "categories": []} {"id": "2504.02464", "title": "CornerPoint3D: Look at the Nearest Corner Instead of the Center", "abstract": " 3D object detection aims to predict object centers, dimensions, and rotations\nfrom LiDAR point clouds. Despite its simplicity, LiDAR captures only the near\nside of objects, making center-based detectors prone to poor localization\naccuracy in cross-domain tasks with varying point distributions. Meanwhile,\nexisting evaluation metrics designed for single-domain assessment also suffer\nfrom overfitting due to dataset-specific size variations. A key question\narises: Do we really need models to maintain excellent performance in the\nentire 3D bounding boxes after being applied across domains? Actually, one of\nour main focuses is on preventing collisions between vehicles and other\nobstacles, especially in cross-domain scenarios where correctly predicting the\nsizes is much more difficult. To address these issues, we rethink cross-domain\n3D object detection from a practical perspective. We propose two new metrics\nthat evaluate a model's ability to detect objects' closer-surfaces to the LiDAR\nsensor. Additionally, we introduce EdgeHead, a refinement head that guides\nmodels to focus more on learnable closer surfaces, significantly improving\ncross-domain performance under both our new and traditional BEV/3D metrics.\nFurthermore, we argue that predicting the nearest corner rather than the object\ncenter enhances robustness. We propose a novel 3D object detector, coined as\nCornerPoint3D, which is built upon CenterPoint and uses heatmaps to supervise\nthe learning and detection of the nearest corner of each object. Our proposed\nmethods realize a balanced trade-off between the detection quality of entire\nbounding boxes and the locating accuracy of closer surfaces to the LiDAR\nsensor, outperforming the traditional center-based detector CenterPoint in\nmultiple cross-domain tasks and providing a more practically reasonable and\nrobust cross-domain 3D object detection solution.\n", "date": "2025-04-03", "categories": []} {"id": "2401.06946", "title": "3D Object Detection and High-Resolution Traffic Parameters Extraction\n Using Low-Resolution LiDAR Data", "abstract": " Traffic volume data collection is a crucial aspect of transportation\nengineering and urban planning, as it provides vital insights into traffic\npatterns, congestion, and infrastructure efficiency. Traditional manual methods\nof traffic data collection are both time-consuming and costly. However, the\nemergence of modern technologies, particularly Light Detection and Ranging\n(LiDAR), has revolutionized the process by enabling efficient and accurate data\ncollection. Despite the benefits of using LiDAR for traffic data collection,\nprevious studies have identified two major limitations that have impeded its\nwidespread adoption. These are the need for multiple LiDAR systems to obtain\ncomplete point cloud information of objects of interest, as well as the\nlabor-intensive process of annotating 3D bounding boxes for object detection\ntasks. In response to these challenges, the current study proposes an\ninnovative framework that alleviates the need for multiple LiDAR systems and\nsimplifies the laborious 3D annotation process. To achieve this goal, the study\nemployed a single LiDAR system, that aims at reducing the data acquisition cost\nand addressed its accompanying limitation of missing point cloud information by\ndeveloping a Point Cloud Completion (PCC) framework to fill in missing point\ncloud information using point density. Furthermore, we also used zero-shot\nlearning techniques to detect vehicles and pedestrians, as well as proposed a\nunique framework for extracting low to high features from the object of\ninterest, such as height, acceleration, and speed. Using the 2D bounding box\ndetection and extracted height information, this study is able to generate 3D\nbounding boxes automatically without human intervention.\n", "date": "2024-01-13", "categories": []} {"id": "2504.00382", "title": "Intrinsic-feature-guided 3D Object Detection", "abstract": " LiDAR-based 3D object detection is essential for autonomous driving systems.\nHowever, LiDAR point clouds may appear to have sparsity, uneven distribution,\nand incomplete structures, significantly limiting the detection performance. In\nroad driving environments, target objects referring to vehicles, pedestrians\nand cyclists are well-suited for enhancing representation through the complete\ntemplate guidance, considering their grid and topological structures.\nTherefore, this paper presents an intrinsic-feature-guided 3D object detection\nmethod based on a template-assisted feature enhancement module, which extracts\nintrinsic features from relatively generalized templates and provides rich\nstructural information for foreground objects. Furthermore, a proposal-level\ncontrastive learning mechanism is designed to enhance the feature differences\nbetween foreground and background objects. The proposed modules can act as\nplug-and-play components and improve the performance of multiple existing\nmethods. Extensive experiments illustrate that the proposed method achieves the\nhighly competitive detection results. Code will be available at\nhttps://github.com/zhangwanjingjj/IfgNet.git.\n", "date": "2025-04-01", "categories": []} {"id": "2410.02646", "title": "Learning 3D Perception from Others' Predictions", "abstract": " Accurate 3D object detection in real-world environments requires a huge\namount of annotated data with high quality. Acquiring such data is tedious and\nexpensive, and often needs repeated effort when a new sensor is adopted or when\nthe detector is deployed in a new environment. We investigate a new scenario to\nconstruct 3D object detectors: learning from the predictions of a nearby unit\nthat is equipped with an accurate detector. For example, when a self-driving\ncar enters a new area, it may learn from other traffic participants whose\ndetectors have been optimized for that area. This setting is label-efficient,\nsensor-agnostic, and communication-efficient: nearby units only need to share\nthe predictions with the ego agent (e.g., car). Naively using the received\npredictions as ground-truths to train the detector for the ego car, however,\nleads to inferior performance. We systematically study the problem and identify\nviewpoint mismatches and mislocalization (due to synchronization and GPS\nerrors) as the main causes, which unavoidably result in false positives, false\nnegatives, and inaccurate pseudo labels. We propose a distance-based\ncurriculum, first learning from closer units with similar viewpoints and\nsubsequently improving the quality of other units' predictions via\nself-training. We further demonstrate that an effective pseudo label refinement\nmodule can be trained with a handful of annotated data, largely reducing the\ndata quantity necessary to train an object detector. We validate our approach\non the recently released real-world collaborative driving dataset, using\nreference cars' predictions as pseudo labels for the ego car. Extensive\nexperiments including several scenarios (e.g., different sensors, detectors,\nand domains) demonstrate the effectiveness of our approach toward\nlabel-efficient learning of 3D perception from other units' predictions.\n", "date": "2024-10-03", "categories": []} {"id": "2503.19776", "title": "Resilient Sensor Fusion under Adverse Sensor Failures via Multi-Modal\n Expert Fusion", "abstract": " Modern autonomous driving perception systems utilize complementary\nmulti-modal sensors, such as LiDAR and cameras. Although sensor fusion\narchitectures enhance performance in challenging environments, they still\nsuffer significant performance drops under severe sensor failures, such as\nLiDAR beam reduction, LiDAR drop, limited field of view, camera drop, and\nocclusion. This limitation stems from inter-modality dependencies in current\nsensor fusion frameworks. In this study, we introduce an efficient and robust\nLiDAR-camera 3D object detector, referred to as MoME, which can achieve robust\nperformance through a mixture of experts approach. Our MoME fully decouples\nmodality dependencies using three parallel expert decoders, which use camera\nfeatures, LiDAR features, or a combination of both to decode object queries,\nrespectively. We propose Multi-Expert Decoding (MED) framework, where each\nquery is decoded selectively using one of three expert decoders. MoME utilizes\nan Adaptive Query Router (AQR) to select the most appropriate expert decoder\nfor each query based on the quality of camera and LiDAR features. This ensures\nthat each query is processed by the best-suited expert, resulting in robust\nperformance across diverse sensor failure scenarios. We evaluated the\nperformance of MoME on the nuScenes-R benchmark. Our MoME achieved\nstate-of-the-art performance in extreme weather and sensor failure conditions,\nsignificantly outperforming the existing models across various sensor failure\nscenarios.\n", "date": "2025-03-25", "categories": []} {"id": "2412.00692", "title": "MCBLT: Multi-Camera Multi-Object 3D Tracking in Long Videos", "abstract": " Object perception from multi-view cameras is crucial for intelligent systems,\nparticularly in indoor environments, e.g., warehouses, retail stores, and\nhospitals. Most traditional multi-target multi-camera (MTMC) detection and\ntracking methods rely on 2D object detection, single-view multi-object tracking\n(MOT), and cross-view re-identification (ReID) techniques, without properly\nhandling important 3D information by multi-view image aggregation. In this\npaper, we propose a 3D object detection and tracking framework, named MCBLT,\nwhich first aggregates multi-view images with necessary camera calibration\nparameters to obtain 3D object detections in bird's-eye view (BEV). Then, we\nintroduce hierarchical graph neural networks (GNNs) to track these 3D\ndetections in BEV for MTMC tracking results. Unlike existing methods, MCBLT has\nimpressive generalizability across different scenes and diverse camera\nsettings, with exceptional capability for long-term association handling. As a\nresult, our proposed MCBLT establishes a new state-of-the-art on the AICity'24\ndataset with $81.22$ HOTA, and on the WildTrack dataset with $95.6$ IDF1.\n", "date": "2024-12-01", "categories": []} {"id": "2403.07746", "title": "Unleashing HyDRa: Hybrid Fusion, Depth Consistency and Radar for Unified\n 3D Perception", "abstract": " Low-cost, vision-centric 3D perception systems for autonomous driving have\nmade significant progress in recent years, narrowing the gap to expensive\nLiDAR-based methods. The primary challenge in becoming a fully reliable\nalternative lies in robust depth prediction capabilities, as camera-based\nsystems struggle with long detection ranges and adverse lighting and weather\nconditions. In this work, we introduce HyDRa, a novel camera-radar fusion\narchitecture for diverse 3D perception tasks. Building upon the principles of\ndense BEV (Bird's Eye View)-based architectures, HyDRa introduces a hybrid\nfusion approach to combine the strengths of complementary camera and radar\nfeatures in two distinct representation spaces. Our Height Association\nTransformer module leverages radar features already in the perspective view to\nproduce more robust and accurate depth predictions. In the BEV, we refine the\ninitial sparse representation by a Radar-weighted Depth Consistency. HyDRa\nachieves a new state-of-the-art for camera-radar fusion of 64.2 NDS (+1.8) and\n58.4 AMOTA (+1.5) on the public nuScenes dataset. Moreover, our new\nsemantically rich and spatially accurate BEV features can be directly converted\ninto a powerful occupancy representation, beating all previous camera-based\nmethods on the Occ3D benchmark by an impressive 3.7 mIoU. Code and models are\navailable at https://github.com/phi-wol/hydra.\n", "date": "2024-03-12", "categories": []} {"id": "2404.04910", "title": "MonoTAKD: Teaching Assistant Knowledge Distillation for Monocular 3D\n Object Detection", "abstract": " Monocular 3D object detection (Mono3D) holds noteworthy promise for\nautonomous driving applications owing to the cost-effectiveness and rich visual\ncontext of monocular camera sensors. However, depth ambiguity poses a\nsignificant challenge, as it requires extracting precise 3D scene geometry from\na single image, resulting in suboptimal performance when transferring knowledge\nfrom a LiDAR-based teacher model to a camera-based student model. To facilitate\neffective distillation, we introduce Monocular Teaching Assistant Knowledge\nDistillation (MonoTAKD), which proposes a camera-based teaching assistant (TA)\nmodel to transfer robust 3D visual knowledge to the student model, leveraging\nthe smaller feature representation gap. Additionally, we define 3D spatial cues\nas residual features that capture the differences between the teacher and the\nTA models. We then leverage these cues to improve the student model's 3D\nperception capabilities. Experimental results show that our MonoTAKD achieves\nstate-of-the-art performance on the KITTI3D dataset. Furthermore, we evaluate\nthe performance on nuScenes and KITTI raw datasets to demonstrate the\ngeneralization of our model to multi-view 3D and unsupervised data settings.\nOur code is available at https://github.com/hoiliu-0801/MonoTAKD.\n", "date": "2024-04-07", "categories": []} {"id": "2503.20682", "title": "GLRD: Global-Local Collaborative Reason and Debate with PSL for 3D\n Open-Vocabulary Detection", "abstract": " The task of LiDAR-based 3D Open-Vocabulary Detection (3D OVD) requires the\ndetector to learn to detect novel objects from point clouds without\noff-the-shelf training labels. Previous methods focus on the learning of\nobject-level representations and ignore the scene-level information, thus it is\nhard to distinguish objects with similar classes. In this work, we propose a\nGlobal-Local Collaborative Reason and Debate with PSL (GLRD) framework for the\n3D OVD task, considering both local object-level information and global\nscene-level information. Specifically, LLM is utilized to perform common sense\nreasoning based on object-level and scene-level information, where the\ndetection result is refined accordingly. To further boost the LLM's ability of\nprecise decisions, we also design a probabilistic soft logic solver (OV-PSL) to\nsearch for the optimal solution, and a debate scheme to confirm the class of\nconfusable objects. In addition, to alleviate the uneven distribution of\nclasses, a static balance scheme (SBC) and a dynamic balance scheme (DBC) are\ndesigned. In addition, to reduce the influence of noise in data and training,\nwe further propose Reflected Pseudo Labels Generation (RPLG) and\nBackground-Aware Object Localization (BAOL). Extensive experiments conducted on\nScanNet and SUN RGB-D demonstrate the superiority of GLRD, where absolute\nimprovements in mean average precision are $+2.82\\%$ on SUN RGB-D and $+3.72\\%$\non ScanNet in the partial open-vocabulary setting. In the full open-vocabulary\nsetting, the absolute improvements in mean average precision are $+4.03\\%$ on\nScanNet and $+14.11\\%$ on SUN RGB-D.\n", "date": "2025-03-26", "categories": []} {"id": "2206.10965", "title": "Polar Parametrization for Vision-based Surround-View 3D Detection", "abstract": " 3D detection based on surround-view camera system is a critical technique in\nautopilot. In this work, we present Polar Parametrization for 3D detection,\nwhich reformulates position parametrization, velocity decomposition, perception\nrange, label assignment and loss function in polar coordinate system. Polar\nParametrization establishes explicit associations between image patterns and\nprediction targets, exploiting the view symmetry of surround-view cameras as\ninductive bias to ease optimization and boost performance. Based on Polar\nParametrization, we propose a surround-view 3D DEtection TRansformer, named\nPolarDETR. PolarDETR achieves promising performance-speed trade-off on\ndifferent backbone configurations. Besides, PolarDETR ranks 1st on the\nleaderboard of nuScenes benchmark in terms of both 3D detection and 3D tracking\nat the submission time (Mar. 4th, 2022). Code will be released at\n\\url{https://github.com/hustvl/PolarDETR}.\n", "date": "2022-06-22", "categories": []} {"id": "2412.12725", "title": "RaCFormer: Towards High-Quality 3D Object Detection via Query-based\n Radar-Camera Fusion", "abstract": " We propose Radar-Camera fusion transformer (RaCFormer) to boost the accuracy\nof 3D object detection by the following insight. The Radar-Camera fusion in\noutdoor 3D scene perception is capped by the image-to-BEV transformation--if\nthe depth of pixels is not accurately estimated, the naive combination of BEV\nfeatures actually integrates unaligned visual content. To avoid this problem,\nwe propose a query-based framework that enables adaptive sampling of\ninstance-relevant features from both the bird's-eye view (BEV) and the original\nimage view. Furthermore, we enhance system performance by two key designs:\noptimizing query initialization and strengthening the representational capacity\nof BEV. For the former, we introduce an adaptive circular distribution in polar\ncoordinates to refine the initialization of object queries, allowing for a\ndistance-based adjustment of query density. For the latter, we initially\nincorporate a radar-guided depth head to refine the transformation from image\nview to BEV. Subsequently, we focus on leveraging the Doppler effect of radar\nand introduce an implicit dynamic catcher to capture the temporal elements\nwithin the BEV. Extensive experiments on nuScenes and View-of-Delft (VoD)\ndatasets validate the merits of our design. Remarkably, our method achieves\nsuperior results of 64.9% mAP and 70.2% NDS on nuScenes. RaCFormer also secures\nthe state-of-the-art performance on the VoD dataset. Code is available at\nhttps://github.com/cxmomo/RaCFormer.\n", "date": "2024-12-17", "categories": []} {"id": "2412.19165", "title": "Revisiting Monocular 3D Object Detection with Depth Thickness Field", "abstract": " Monocular 3D object detection is challenging due to the lack of accurate\ndepth. However, existing depth-assisted solutions still exhibit inferior\nperformance, whose reason is universally acknowledged as the unsatisfactory\naccuracy of monocular depth estimation models. In this paper, we revisit\nmonocular 3D object detection from the depth perspective and formulate an\nadditional issue as the limited 3D structure-aware capability of existing depth\nrepresentations (e.g., depth one-hot encoding or depth distribution). To\naddress this issue, we introduce a novel Depth Thickness Field approach to\nembed clear 3D structures of the scenes. Specifically, we present MonoDTF, a\nscene-to-instance depth-adapted network for monocular 3D object detection. The\nframework mainly comprises a Scene-Level Depth Retargeting (SDR) module and an\nInstance-Level Spatial Refinement (ISR) module. The former retargets\ntraditional depth representations to the proposed depth thickness field,\nincorporating the scene-level perception of 3D structures. The latter refines\nthe voxel space with the guidance of instances, enhancing the 3D instance-aware\ncapability of the depth thickness field and thus improving detection accuracy.\nExtensive experiments on the KITTI and Waymo datasets demonstrate our\nsuperiority to existing state-of-the-art (SoTA) methods and the universality\nwhen equipped with different depth estimation models. The code will be\navailable.\n", "date": "2024-12-26", "categories": []} {"id": "2501.15449", "title": "Breaking the SSL-AL Barrier: A Synergistic Semi-Supervised Active\n Learning Framework for 3D Object Detection", "abstract": " To address the annotation burden in LiDAR-based 3D object detection, active\nlearning (AL) methods offer a promising solution. However, traditional active\nlearning approaches solely rely on a small amount of labeled data to train an\ninitial model for data selection, overlooking the potential of leveraging the\nabundance of unlabeled data. Recently, attempts to integrate semi-supervised\nlearning (SSL) into AL with the goal of leveraging unlabeled data have faced\nchallenges in effectively resolving the conflict between the two paradigms,\nresulting in less satisfactory performance. To tackle this conflict, we propose\na Synergistic Semi-Supervised Active Learning framework, dubbed as S-SSAL.\nSpecifically, from the perspective of SSL, we propose a Collaborative\nPseudoScene Pre-training (CPSP) method that effectively learns from unlabeled\ndata without introducing adverse effects. From the perspective of AL, we design\na Collaborative Active Learning (CAL) method, which complements the uncertainty\nand diversity methods by model cascading. This allows us to fully exploit the\npotential of the CPSP pre-trained model. Extensive experiments conducted on\nKITTI and Waymo demonstrate the effectiveness of our S-SSAL framework. Notably,\non the KITTI dataset, utilizing only 2% labeled data, S-SSAL can achieve\nperformance comparable to models trained on the full dataset. The code has been\nreleased at https://github.com/LandDreamer/S_SSAL.\n", "date": "2025-01-26", "categories": []} {"id": "2503.16811", "title": "Seg2Box: 3D Object Detection by Point-Wise Semantics Supervision", "abstract": " LiDAR-based 3D object detection and semantic segmentation are critical tasks\nin 3D scene understanding. Traditional detection and segmentation methods\nsupervise their models through bounding box labels and semantic mask labels.\nHowever, these two independent labels inherently contain significant\nredundancy. This paper aims to eliminate the redundancy by supervising 3D\nobject detection using only semantic labels. However, the challenge arises due\nto the incomplete geometry structure and boundary ambiguity of point-cloud\ninstances, leading to inaccurate pseudo labels and poor detection results. To\naddress these challenges, we propose a novel method, named Seg2Box. We first\nintroduce a Multi-Frame Multi-Scale Clustering (MFMS-C) module, which leverages\nthe spatio-temporal consistency of point clouds to generate accurate box-level\npseudo-labels. Additionally, the Semantic?Guiding Iterative-Mining\nSelf-Training (SGIM-ST) module is proposed to enhance the performance by\nprogressively refining the pseudo-labels and mining the instances without\ngenerating pseudo-labels. Experiments on the Waymo Open Dataset and nuScenes\nDataset show that our method significantly outperforms other competitive\nmethods by 23.7\\% and 10.3\\% in mAP, respectively. The results demonstrate the\ngreat label-efficient potential and advancement of our method.\n", "date": "2025-03-21", "categories": []} {"id": "2403.11371", "title": "V2X-DGW: Domain Generalization for Multi-agent Perception under Adverse\n Weather Conditions", "abstract": " Current LiDAR-based Vehicle-to-Everything (V2X) multi-agent perception\nsystems have shown the significant success on 3D object detection. While these\nmodels perform well in the trained clean weather, they struggle in unseen\nadverse weather conditions with the domain gap. In this paper, we propose a\nDomain Generalization based approach, named \\textit{V2X-DGW}, for LiDAR-based\n3D object detection on multi-agent perception system under adverse weather\nconditions. Our research aims to not only maintain favorable multi-agent\nperformance in the clean weather but also promote the performance in the unseen\nadverse weather conditions by learning only on the clean weather data. To\nrealize the Domain Generalization, we first introduce the Adaptive Weather\nAugmentation (AWA) to mimic the unseen adverse weather conditions, and then\npropose two alignments for generalizable representation learning: Trust-region\nWeather-invariant Alignment (TWA) and Agent-aware Contrastive Alignment (ACA).\nTo evaluate this research, we add Fog, Rain, Snow conditions on two publicized\nmulti-agent datasets based on physics-based models, resulting in two new\ndatasets: OPV2V-w and V2XSet-w. Extensive experiments demonstrate that our\nV2X-DGW achieved significant improvements in the unseen adverse weathers. The\ncode is available at https://github.com/Baolu1998/V2X-DGW.\n", "date": "2024-03-17", "categories": []} {"id": "2503.15997", "title": "Automating 3D Dataset Generation with Neural Radiance Fields", "abstract": " 3D detection is a critical task to understand spatial characteristics of the\nenvironment and is used in a variety of applications including robotics,\naugmented reality, and image retrieval. Training performant detection models\nrequire diverse, precisely annotated, and large scale datasets that involve\ncomplex and expensive creation processes. Hence, there are only few public 3D\ndatasets that are additionally limited in their range of classes. In this work,\nwe propose a pipeline for automatic generation of 3D datasets for arbitrary\nobjects. By utilizing the universal 3D representation and rendering\ncapabilities of Radiance Fields, our pipeline generates high quality 3D models\nfor arbitrary objects. These 3D models serve as input for a synthetic dataset\ngenerator. Our pipeline is fast, easy to use and has a high degree of\nautomation. Our experiments demonstrate, that 3D pose estimation networks,\ntrained with our generated datasets, archive strong performance in typical\napplication scenarios.\n", "date": "2025-03-20", "categories": []} {"id": "2503.14493", "title": "State Space Model Meets Transformer: A New Paradigm for 3D Object\n Detection", "abstract": " DETR-based methods, which use multi-layer transformer decoders to refine\nobject queries iteratively, have shown promising performance in 3D indoor\nobject detection. However, the scene point features in the transformer decoder\nremain fixed, leading to minimal contributions from later decoder layers,\nthereby limiting performance improvement. Recently, State Space Models (SSM)\nhave shown efficient context modeling ability with linear complexity through\niterative interactions between system states and inputs. Inspired by SSMs, we\npropose a new 3D object DEtection paradigm with an interactive STate space\nmodel (DEST). In the interactive SSM, we design a novel state-dependent SSM\nparameterization method that enables system states to effectively serve as\nqueries in 3D indoor detection tasks. In addition, we introduce four key\ndesigns tailored to the characteristics of point cloud and SSM: The\nserialization and bidirectional scanning strategies enable bidirectional\nfeature interaction among scene points within the SSM. The inter-state\nattention mechanism models the relationships between state points, while the\ngated feed-forward network enhances inter-channel correlations. To the best of\nour knowledge, this is the first method to model queries as system states and\nscene points as system inputs, which can simultaneously update scene point\nfeatures and query features with linear complexity. Extensive experiments on\ntwo challenging datasets demonstrate the effectiveness of our DEST-based\nmethod. Our method improves the GroupFree baseline in terms of AP50 on ScanNet\nV2 (+5.3) and SUN RGB-D (+3.2) datasets. Based on the VDETR baseline, Our\nmethod sets a new SOTA on the ScanNetV2 and SUN RGB-D datasets.\n", "date": "2025-03-18", "categories": []} {"id": "2503.15022", "title": "xMOD: Cross-Modal Distillation for 2D/3D Multi-Object Discovery from 2D\n motion", "abstract": " Object discovery, which refers to the task of localizing objects without\nhuman annotations, has gained significant attention in 2D image analysis.\nHowever, despite this growing interest, it remains under-explored in 3D data,\nwhere approaches rely exclusively on 3D motion, despite its several challenges.\nIn this paper, we present a novel framework that leverages advances in 2D\nobject discovery which are based on 2D motion to exploit the advantages of such\nmotion cues being more flexible and generalizable and to bridge the gap between\n2D and 3D modalities. Our primary contributions are twofold: (i) we introduce\nDIOD-3D, the first baseline for multi-object discovery in 3D data using 2D\nmotion, incorporating scene completion as an auxiliary task to enable dense\nobject localization from sparse input data; (ii) we develop xMOD, a cross-modal\ntraining framework that integrates 2D and 3D data while always using 2D motion\ncues. xMOD employs a teacher-student training paradigm across the two\nmodalities to mitigate confirmation bias by leveraging the domain gap. During\ninference, the model supports both RGB-only and point cloud-only inputs.\nAdditionally, we propose a late-fusion technique tailored to our pipeline that\nfurther enhances performance when both modalities are available at inference.\nWe evaluate our approach extensively on synthetic (TRIP-PD) and challenging\nreal-world datasets (KITTI and Waymo). Notably, our approach yields a\nsubstantial performance improvement compared with the 2D object discovery\nstate-of-the-art on all datasets with gains ranging from +8.7 to +15.1 in F1@50\nscore. The code is available at https://github.com/CEA-LIST/xMOD\n", "date": "2025-03-19", "categories": []} {"id": "2503.15087", "title": "An Investigation of Beam Density on LiDAR Object Detection Performance", "abstract": " Accurate 3D object detection is a critical component of autonomous driving,\nenabling vehicles to perceive their surroundings with precision and make\ninformed decisions. LiDAR sensors, widely used for their ability to provide\ndetailed 3D measurements, are key to achieving this capability. However,\nvariations between training and inference data can cause significant\nperformance drops when object detection models are employed in different sensor\nsettings. One critical factor is beam density, as inference on sparse,\ncost-effective LiDAR sensors is often preferred in real-world applications.\nDespite previous work addressing the beam-density-induced domain gap,\nsubstantial knowledge gaps remain, particularly concerning dense 128-beam\nsensors in cross-domain scenarios. To gain better understanding of the impact\nof beam density on domain gaps, we conduct a comprehensive investigation that\nincludes an evaluation of different object detection architectures. Our\narchitecture evaluation reveals that combining voxel- and point-based\napproaches yields superior cross-domain performance by leveraging the strengths\nof both representations. Building on these findings, we analyze\nbeam-density-induced domain gaps and argue that these domain gaps must be\nevaluated in conjunction with other domain shifts. Contrary to conventional\nbeliefs, our experiments reveal that detectors benefit from training on denser\ndata and exhibit robustness to beam density variations during inference.\n", "date": "2025-03-19", "categories": []} {"id": "2503.15211", "title": "GO-N3RDet: Geometry Optimized NeRF-enhanced 3D Object Detector", "abstract": " We propose GO-N3RDet, a scene-geometry optimized multi-view 3D object\ndetector enhanced by neural radiance fields. The key to accurate 3D object\ndetection is in effective voxel representation. However, due to occlusion and\nlack of 3D information, constructing 3D features from multi-view 2D images is\nchallenging. Addressing that, we introduce a unique 3D positional information\nembedded voxel optimization mechanism to fuse multi-view features. To\nprioritize neural field reconstruction in object regions, we also devise a\ndouble importance sampling scheme for the NeRF branch of our detector. We\nadditionally propose an opacity optimization module for precise voxel opacity\nprediction by enforcing multi-view consistency constraints. Moreover, to\nfurther improve voxel density consistency across multiple perspectives, we\nincorporate ray distance as a weighting factor to minimize cumulative ray\nerrors. Our unique modules synergetically form an end-to-end neural model that\nestablishes new state-of-the-art in NeRF-based multi-view 3D detection,\nverified with extensive experiments on ScanNet and ARKITScenes. Code will be\navailable at https://github.com/ZechuanLi/GO-N3RDet.\n", "date": "2025-03-19", "categories": []} {"id": "2503.15435", "title": "V2X-DG: Domain Generalization for Vehicle-to-Everything Cooperative\n Perception", "abstract": " LiDAR-based Vehicle-to-Everything (V2X) cooperative perception has\ndemonstrated its impact on the safety and effectiveness of autonomous driving.\nSince current cooperative perception algorithms are trained and tested on the\nsame dataset, the generalization ability of cooperative perception systems\nremains underexplored. This paper is the first work to study the Domain\nGeneralization problem of LiDAR-based V2X cooperative perception (V2X-DG) for\n3D detection based on four widely-used open source datasets: OPV2V, V2XSet,\nV2V4Real and DAIR-V2X. Our research seeks to sustain high performance not only\nwithin the source domain but also across other unseen domains, achieved solely\nthrough training on source domain. To this end, we propose Cooperative Mixup\nAugmentation based Generalization (CMAG) to improve the model generalization\ncapability by simulating the unseen cooperation, which is designed compactly\nfor the domain gaps in cooperative perception. Furthermore, we propose a\nconstraint for the regularization of the robust generalized feature\nrepresentation learning: Cooperation Feature Consistency (CFC), which aligns\nthe intermediately fused features of the generalized cooperation by CMAG and\nthe early fused features of the original cooperation in source domain.\nExtensive experiments demonstrate that our approach achieves significant\nperformance gains when generalizing to other unseen datasets while it also\nmaintains strong performance on the source dataset.\n", "date": "2025-03-19", "categories": []} {"id": "2406.18414", "title": "BiTrack: Bidirectional Offline 3D Multi-Object Tracking Using\n Camera-LiDAR Data", "abstract": " Compared with real-time multi-object tracking (MOT), offline multi-object\ntracking (OMOT) has the advantages to perform 2D-3D detection fusion, erroneous\nlink correction, and full track optimization but has to deal with the\nchallenges from bounding box misalignment and track evaluation, editing, and\nrefinement. This paper proposes \"BiTrack\", a 3D OMOT framework that includes\nmodules of 2D-3D detection fusion, initial trajectory generation, and\nbidirectional trajectory re-optimization to achieve optimal tracking results\nfrom camera-LiDAR data. The novelty of this paper includes threefold: (1)\ndevelopment of a point-level object registration technique that employs a\ndensity-based similarity metric to achieve accurate fusion of 2D-3D detection\nresults; (2) development of a set of data association and track management\nskills that utilizes a vertex-based similarity metric as well as false alarm\nrejection and track recovery mechanisms to generate reliable bidirectional\nobject trajectories; (3) development of a trajectory re-optimization scheme\nthat re-organizes track fragments of different fidelities in a greedy fashion,\nas well as refines each trajectory with completion and smoothing techniques.\nThe experiment results on the KITTI dataset demonstrate that BiTrack achieves\nthe state-of-the-art performance for 3D OMOT tasks in terms of accuracy and\nefficiency.\n", "date": "2024-06-26", "categories": []} {"id": "2501.10266", "title": "MutualForce: Mutual-Aware Enhancement for 4D Radar-LiDAR 3D Object\n Detection", "abstract": " Radar and LiDAR have been widely used in autonomous driving as LiDAR provides\nrich structure information, and radar demonstrates high robustness under\nadverse weather. Recent studies highlight the effectiveness of fusing radar and\nLiDAR point clouds. However, challenges remain due to the modality misalignment\nand information loss during feature extractions. To address these issues, we\npropose a 4D radar-LiDAR framework to mutually enhance their representations.\nInitially, the indicative features from radar are utilized to guide both radar\nand LiDAR geometric feature learning. Subsequently, to mitigate their sparsity\ngap, the shape information from LiDAR is used to enrich radar BEV features.\nExtensive experiments on the View-of-Delft (VoD) dataset demonstrate our\napproach's superiority over existing methods, achieving the highest mAP of\n71.76% across the entire area and 86.36\\% within the driving corridor.\nEspecially for cars, we improve the AP by 4.17% and 4.20% due to the strong\nindicative features and symmetric shapes.\n", "date": "2025-01-17", "categories": []} {"id": "2503.13743", "title": "MonoCT: Overcoming Monocular 3D Detection Domain Shift with Consistent\n Teacher Models", "abstract": " We tackle the problem of monocular 3D object detection across different\nsensors, environments, and camera setups. In this paper, we introduce a novel\nunsupervised domain adaptation approach, MonoCT, that generates highly accurate\npseudo labels for self-supervision. Inspired by our observation that accurate\ndepth estimation is critical to mitigating domain shifts, MonoCT introduces a\nnovel Generalized Depth Enhancement (GDE) module with an ensemble concept to\nimprove depth estimation accuracy. Moreover, we introduce a novel Pseudo Label\nScoring (PLS) module by exploring inner-model consistency measurement and a\nDiversity Maximization (DM) strategy to further generate high-quality pseudo\nlabels for self-training. Extensive experiments on six benchmarks show that\nMonoCT outperforms existing SOTA domain adaptation methods by large margins\n(~21% minimum for AP Mod.) and generalizes well to car, traffic camera and\ndrone views.\n", "date": "2025-03-17", "categories": []} {"id": "2503.13946", "title": "Is Discretization Fusion All You Need for Collaborative Perception?", "abstract": " Collaborative perception in multi-agent system enhances overall perceptual\ncapabilities by facilitating the exchange of complementary information among\nagents. Current mainstream collaborative perception methods rely on discretized\nfeature maps to conduct fusion, which however, lacks flexibility in extracting\nand transmitting the informative features and can hardly focus on the\ninformative features during fusion. To address these problems, this paper\nproposes a novel Anchor-Centric paradigm for Collaborative Object detection\n(ACCO). It avoids grid precision issues and allows more flexible and efficient\nanchor-centric communication and fusion. ACCO is composed by three main\ncomponents: (1) Anchor featuring block (AFB) that targets to generate anchor\nproposals and projects prepared anchor queries to image features. (2) Anchor\nconfidence generator (ACG) is designed to minimize communication by selecting\nonly the features in the confident anchors to transmit. (3) A local-global\nfusion module, in which local fusion is anchor alignment-based fusion (LAAF)\nand global fusion is conducted by spatial-aware cross-attention (SACA). LAAF\nand SACA run in multi-layers, so agents conduct anchor-centric fusion\niteratively to adjust the anchor proposals. Comprehensive experiments are\nconducted to evaluate ACCO on OPV2V and Dair-V2X datasets, which demonstrate\nACCO's superiority in reducing the communication volume, and in improving the\nperception range and detection performances. Code can be found at:\n\\href{https://github.com/sidiangongyuan/ACCO}{https://github.com/sidiangongyuan/ACCO}.\n", "date": "2025-03-18", "categories": []} {"id": "2503.13951", "title": "FrustumFusionNets: A Three-Dimensional Object Detection Network Based on\n Tractor Road Scene", "abstract": " To address the issues of the existing frustum-based methods' underutilization\nof image information in road three-dimensional object detection as well as the\nlack of research on agricultural scenes, we constructed an object detection\ndataset using an 80-line Light Detection And Ranging (LiDAR) and a camera in a\ncomplex tractor road scene and proposed a new network called FrustumFusionNets\n(FFNets). Initially, we utilize the results of image-based two-dimensional\nobject detection to narrow down the search region in the three-dimensional\nspace of the point cloud. Next, we introduce a Gaussian mask to enhance the\npoint cloud information. Then, we extract the features from the frustum point\ncloud and the crop image using the point cloud feature extraction pipeline and\nthe image feature extraction pipeline, respectively. Finally, we concatenate\nand fuse the data features from both modalities to achieve three-dimensional\nobject detection. Experiments demonstrate that on the constructed test set of\ntractor road data, the FrustumFusionNetv2 achieves 82.28% and 95.68% accuracy\nin the three-dimensional object detection of the two main road objects, cars\nand people, respectively. This performance is 1.83% and 2.33% better than the\noriginal model. It offers a hybrid fusion-based multi-object, high-precision,\nreal-time three-dimensional object detection technique for unmanned\nagricultural machines in tractor road scenarios. On the Karlsruhe Institute of\nTechnology and Toyota Technological Institute (KITTI) Benchmark Suite\nvalidation set, the FrustumFusionNetv2 also demonstrates significant\nsuperiority in detecting road pedestrian objects compared with other\nfrustum-based three-dimensional object detection methods.\n", "date": "2025-03-18", "categories": []} {"id": "2402.06423", "title": "CurveFormer++: 3D Lane Detection by Curve Propagation with Temporal\n Curve Queries and Attention", "abstract": " In autonomous driving, accurate 3D lane detection using monocular cameras is\nimportant for downstream tasks. Recent CNN and Transformer approaches usually\napply a two-stage model design. The first stage transforms the image feature\nfrom a front image into a bird's-eye-view (BEV) representation. Subsequently, a\nsub-network processes the BEV feature to generate the 3D detection results.\nHowever, these approaches heavily rely on a challenging image feature\ntransformation module from a perspective view to a BEV representation. In our\nwork, we present CurveFormer++, a single-stage Transformer-based method that\ndoes not require the view transform module and directly infers 3D lane results\nfrom the perspective image features. Specifically, our approach models the 3D\nlane detection task as a curve propagation problem, where each lane is\nrepresented by a curve query with a dynamic and ordered anchor point set. By\nemploying a Transformer decoder, the model can iteratively refine the 3D lane\nresults. A curve cross-attention module is introduced to calculate similarities\nbetween image features and curve queries. To handle varying lane lengths, we\nemploy context sampling and anchor point restriction techniques to compute more\nrelevant image features. Furthermore, we apply a temporal fusion module that\nincorporates selected informative sparse curve queries and their corresponding\nanchor point sets to leverage historical information. In the experiments, we\nevaluate our approach on two publicly real-world datasets. The results\ndemonstrate that our method provides outstanding performance compared with both\nCNN and Transformer based methods. We also conduct ablation studies to analyze\nthe impact of each component.\n", "date": "2024-02-09", "categories": []} {"id": "2503.08992", "title": "Dual-Domain Homogeneous Fusion with Cross-Modal Mamba and Progressive\n Decoder for 3D Object Detection", "abstract": " Fusing LiDAR and image features in a homogeneous BEV domain has become\npopular for 3D object detection in autonomous driving. However, this paradigm\nis constrained by the excessive feature compression. While some works explore\ndense voxel fusion to enable better feature interaction, they face high\ncomputational costs and challenges in query generation. Additionally, feature\nmisalignment in both domains results in suboptimal detection accuracy. To\naddress these limitations, we propose a Dual-Domain Homogeneous Fusion network\n(DDHFusion), which leverages the complementarily of both BEV and voxel domains\nwhile mitigating their drawbacks. Specifically, we first transform image\nfeatures into BEV and sparse voxel representations using lift-splat-shot and\nour proposed Semantic-Aware Feature Sampling (SAFS) module. The latter\nsignificantly reduces computational overhead by discarding unimportant voxels.\nNext, we introduce Homogeneous Voxel and BEV Fusion (HVF and HBF) networks for\nmulti-modal fusion within respective domains. They are equipped with novel\ncross-modal Mamba blocks to resolve feature misalignment and enable\ncomprehensive scene perception. The output voxel features are injected into the\nBEV space to compensate for the information loss brought by direct height\ncompression. During query selection, the Progressive Query Generation (PQG)\nmechanism is implemented in the BEV domain to reduce false negatives caused by\nfeature compression. Furthermore, we propose a Progressive Decoder (QD) that\nsequentially aggregates not only context-rich BEV features but also\ngeometry-aware voxel features with deformable attention and the Multi-Modal\nVoxel Feature Mixing (MMVFM) block for precise classification and box\nregression.\n", "date": "2025-03-12", "categories": []} {"id": "2503.10579", "title": "Semantic-Supervised Spatial-Temporal Fusion for LiDAR-based 3D Object\n Detection", "abstract": " LiDAR-based 3D object detection presents significant challenges due to the\ninherent sparsity of LiDAR points. A common solution involves long-term\ntemporal LiDAR data to densify the inputs. However, efficiently leveraging\nspatial-temporal information remains an open problem. In this paper, we propose\na novel Semantic-Supervised Spatial-Temporal Fusion (ST-Fusion) method, which\nintroduces a novel fusion module to relieve the spatial misalignment caused by\nthe object motion over time and a feature-level semantic supervision to\nsufficiently unlock the capacity of the proposed fusion module. Specifically,\nthe ST-Fusion consists of a Spatial Aggregation (SA) module and a Temporal\nMerging (TM) module. The SA module employs a convolutional layer with\nprogressively expanding receptive fields to aggregate the object features from\nthe local regions to alleviate the spatial misalignment, the TM module\ndynamically extracts object features from the preceding frames based on the\nattention mechanism for a comprehensive sequential presentation. Besides, in\nthe semantic supervision, we propose a Semantic Injection method to enrich the\nsparse LiDAR data via injecting the point-wise semantic labels, using it for\ntraining a teacher model and providing a reconstruction target at the feature\nlevel supervised by the proposed object-aware loss. Extensive experiments on\nvarious LiDAR-based detectors demonstrate the effectiveness and universality of\nour proposal, yielding an improvement of approximately +2.8% in NDS based on\nthe nuScenes benchmark.\n", "date": "2025-03-13", "categories": []} {"id": "2503.12914", "title": "Efficient Multimodal 3D Object Detector via Instance-Level Contrastive\n Distillation", "abstract": " Multimodal 3D object detectors leverage the strengths of both geometry-aware\nLiDAR point clouds and semantically rich RGB images to enhance detection\nperformance. However, the inherent heterogeneity between these modalities,\nincluding unbalanced convergence and modal misalignment, poses significant\nchallenges. Meanwhile, the large size of the detection-oriented feature also\nconstrains existing fusion strategies to capture long-range dependencies for\nthe 3D detection tasks. In this work, we introduce a fast yet effective\nmultimodal 3D object detector, incorporating our proposed Instance-level\nContrastive Distillation (ICD) framework and Cross Linear Attention Fusion\nModule (CLFM). ICD aligns instance-level image features with LiDAR\nrepresentations through object-aware contrastive distillation, ensuring\nfine-grained cross-modal consistency. Meanwhile, CLFM presents an efficient and\nscalable fusion strategy that enhances cross-modal global interactions within\nsizable multimodal BEV features. Extensive experiments on the KITTI and\nnuScenes 3D object detection benchmarks demonstrate the effectiveness of our\nmethods. Notably, our 3D object detector outperforms state-of-the-art (SOTA)\nmethods while achieving superior efficiency. The implementation of our method\nhas been released as open-source at: https://github.com/nubot-nudt/ICD-Fusion.\n", "date": "2025-03-17", "categories": []} {"id": "2503.12982", "title": "SparseAlign: A Fully Sparse Framework for Cooperative Object Detection", "abstract": " Cooperative perception can increase the view field and decrease the occlusion\nof an ego vehicle, hence improving the perception performance and safety of\nautonomous driving. Despite the success of previous works on cooperative object\ndetection, they mostly operate on dense Bird's Eye View (BEV) feature maps,\nwhich are computationally demanding and can hardly be extended to long-range\ndetection problems. More efficient fully sparse frameworks are rarely explored.\nIn this work, we design a fully sparse framework, SparseAlign, with three key\nfeatures: an enhanced sparse 3D backbone, a query-based temporal context\nlearning module, and a robust detection head specially tailored for sparse\nfeatures. Extensive experimental results on both OPV2V and DairV2X datasets\nshow that our framework, despite its sparsity, outperforms the state of the art\nwith less communication bandwidth requirements. In addition, experiments on the\nOPV2Vt and DairV2Xt datasets for time-aligned cooperative object detection also\nshow a significant performance gain compared to the baseline works.\n", "date": "2025-03-17", "categories": []} {"id": "2503.10777", "title": "HeightFormer: Learning Height Prediction in Voxel Features for Roadside\n Vision Centric 3D Object Detection via Transformer", "abstract": " Roadside vision centric 3D object detection has received increasing attention\nin recent years. It expands the perception range of autonomous vehicles,\nenhances the road safety. Previous methods focused on predicting per-pixel\nheight rather than depth, making significant gains in roadside visual\nperception. While it is limited by the perspective property of near-large and\nfar-small on image features, making it difficult for network to understand real\ndimension of objects in the 3D world. BEV features and voxel features present\nthe real distribution of objects in 3D world compared to the image features.\nHowever, BEV features tend to lose details due to the lack of explicit height\ninformation, and voxel features are computationally expensive. Inspired by this\ninsight, an efficient framework learning height prediction in voxel features\nvia transformer is proposed, dubbed HeightFormer. It groups the voxel features\ninto local height sequences, and utilize attention mechanism to obtain height\ndistribution prediction. Subsequently, the local height sequences are\nreassembled to generate accurate 3D features. The proposed method is applied to\ntwo large-scale roadside benchmarks, DAIR-V2X-I and Rope3D. Extensive\nexperiments are performed and the HeightFormer outperforms the state-of-the-art\nmethods in roadside vision centric 3D object detection task.\n", "date": "2025-03-13", "categories": []} {"id": "2403.11848", "title": "GraphBEV: Towards Robust BEV Feature Alignment for Multi-Modal 3D Object\n Detection", "abstract": " Integrating LiDAR and camera information into Bird's-Eye-View (BEV)\nrepresentation has emerged as a crucial aspect of 3D object detection in\nautonomous driving. However, existing methods are susceptible to the inaccurate\ncalibration relationship between LiDAR and the camera sensor. Such inaccuracies\nresult in errors in depth estimation for the camera branch, ultimately causing\nmisalignment between LiDAR and camera BEV features. In this work, we propose a\nrobust fusion framework called Graph BEV. Addressing errors caused by\ninaccurate point cloud projection, we introduce a Local Align module that\nemploys neighbor-aware depth features via Graph matching. Additionally, we\npropose a Global Align module to rectify the misalignment between LiDAR and\ncamera BEV features. Our Graph BEV framework achieves state-of-the-art\nperformance, with an mAP of 70.1\\%, surpassing BEV Fusion by 1.6\\% on the\nnuscenes validation set. Importantly, our Graph BEV outperforms BEV Fusion by\n8.3\\% under conditions with misalignment noise.\n", "date": "2024-03-18", "categories": []} {"id": "2503.08421", "title": "Learning to Detect Objects from Multi-Agent LiDAR Scans without Manual\n Labels", "abstract": " Unsupervised 3D object detection serves as an important solution for offline\n3D object annotation. However, due to the data sparsity and limited views, the\nclustering-based label fitting in unsupervised object detection often generates\nlow-quality pseudo-labels. Multi-agent collaborative dataset, which involves\nthe sharing of complementary observations among agents, holds the potential to\nbreak through this bottleneck. In this paper, we introduce a novel unsupervised\nmethod that learns to Detect Objects from Multi-Agent LiDAR scans, termed DOtA,\nwithout using labels from external. DOtA first uses the internally shared\nego-pose and ego-shape of collaborative agents to initialize the detector,\nleveraging the generalization performance of neural networks to infer\npreliminary labels. Subsequently,DOtA uses the complementary observations\nbetween agents to perform multi-scale encoding on preliminary labels, then\ndecodes high-quality and low-quality labels. These labels are further used as\nprompts to guide a correct feature learning process, thereby enhancing the\nperformance of the unsupervised object detection task. Extensive experiments on\nthe V2V4Real and OPV2V datasets show that our DOtA outperforms state-of-the-art\nunsupervised 3D object detection methods. Additionally, we also validate the\neffectiveness of the DOtA labels under various collaborative perception\nframeworks.The code is available at https://github.com/xmuqimingxia/DOtA.\n", "date": "2025-03-11", "categories": []} {"id": "2410.19590", "title": "MonoDGP: Monocular 3D Object Detection with Decoupled-Query and\n Geometry-Error Priors", "abstract": " Perspective projection has been extensively utilized in monocular 3D object\ndetection methods. It introduces geometric priors from 2D bounding boxes and 3D\nobject dimensions to reduce the uncertainty of depth estimation. However, due\nto depth errors originating from the object's visual surface, the height of the\nbounding box often fails to represent the actual projected central height,\nwhich undermines the effectiveness of geometric depth. Direct prediction for\nthe projected height unavoidably results in a loss of 2D priors, while\nmulti-depth prediction with complex branches does not fully leverage geometric\ndepth. This paper presents a Transformer-based monocular 3D object detection\nmethod called MonoDGP, which adopts perspective-invariant geometry errors to\nmodify the projection formula. We also try to systematically discuss and\nexplain the mechanisms and efficacy behind geometry errors, which serve as a\nsimple but effective alternative to multi-depth prediction. Additionally,\nMonoDGP decouples the depth-guided decoder and constructs a 2D decoder only\ndependent on visual features, providing 2D priors and initializing object\nqueries without the disturbance of 3D detection. To further optimize and\nfine-tune input tokens of the transformer decoder, we also introduce a Region\nSegment Head (RSH) that generates enhanced features and segment embeddings. Our\nmonocular method demonstrates state-of-the-art performance on the KITTI\nbenchmark without extra data. Code is available at\nhttps://github.com/PuFanqi23/MonoDGP.\n", "date": "2024-10-25", "categories": []} {"id": "2503.08092", "title": "SparseVoxFormer: Sparse Voxel-based Transformer for Multi-modal 3D\n Object Detection", "abstract": " Most previous 3D object detection methods that leverage the multi-modality of\nLiDAR and cameras utilize the Bird's Eye View (BEV) space for intermediate\nfeature representation. However, this space uses a low x, y-resolution and\nsacrifices z-axis information to reduce the overall feature resolution, which\nmay result in declined accuracy. To tackle the problem of using low-resolution\nfeatures, this paper focuses on the sparse nature of LiDAR point cloud data.\nFrom our observation, the number of occupied cells in the 3D voxels constructed\nfrom a LiDAR data can be even fewer than the number of total cells in the BEV\nmap, despite the voxels' significantly higher resolution. Based on this, we\nintroduce a novel sparse voxel-based transformer network for 3D object\ndetection, dubbed as SparseVoxFormer. Instead of performing BEV feature\nextraction, we directly leverage sparse voxel features as the input for a\ntransformer-based detector. Moreover, with regard to the camera modality, we\nintroduce an explicit modality fusion approach that involves projecting 3D\nvoxel coordinates onto 2D images and collecting the corresponding image\nfeatures. Thanks to these components, our approach can leverage geometrically\nricher multi-modal features while even reducing the computational cost. Beyond\nthe proof-of-concept level, we further focus on facilitating better multi-modal\nfusion and flexible control over the number of sparse features. Finally,\nthorough experimental results demonstrate that utilizing a significantly\nsmaller number of sparse features drastically reduces computational costs in a\n3D object detector while enhancing both overall and long-range performance.\n", "date": "2025-03-11", "categories": []} {"id": "2503.08639", "title": "GBlobs: Explicit Local Structure via Gaussian Blobs for Improved\n Cross-Domain LiDAR-based 3D Object Detection", "abstract": " LiDAR-based 3D detectors need large datasets for training, yet they struggle\nto generalize to novel domains. Domain Generalization (DG) aims to mitigate\nthis by training detectors that are invariant to such domain shifts. Current DG\napproaches exclusively rely on global geometric features (point cloud Cartesian\ncoordinates) as input features. Over-reliance on these global geometric\nfeatures can, however, cause 3D detectors to prioritize object location and\nabsolute position, resulting in poor cross-domain performance. To mitigate\nthis, we propose to exploit explicit local point cloud structure for DG, in\nparticular by encoding point cloud neighborhoods with Gaussian blobs, GBlobs.\nOur proposed formulation is highly efficient and requires no additional\nparameters. Without any bells and whistles, simply by integrating GBlobs in\nexisting detectors, we beat the current state-of-the-art in challenging\nsingle-source DG benchmarks by over 21 mAP (Waymo->KITTI), 13 mAP\n(KITTI->Waymo), and 12 mAP (nuScenes->KITTI), without sacrificing in-domain\nperformance. Additionally, GBlobs demonstrate exceptional performance in\nmulti-source DG, surpassing the current state-of-the-art by 17, 12, and 5 mAP\non Waymo, KITTI, and ONCE, respectively.\n", "date": "2025-03-11", "categories": []} {"id": "2503.06435", "title": "OV-SCAN: Semantically Consistent Alignment for Novel Object Discovery in\n Open-Vocabulary 3D Object Detection", "abstract": " Open-vocabulary 3D object detection for autonomous driving aims to detect\nnovel objects beyond the predefined training label sets in point cloud scenes.\nExisting approaches achieve this by connecting traditional 3D object detectors\nwith vision-language models (VLMs) to regress 3D bounding boxes for novel\nobjects and perform open-vocabulary classification through cross-modal\nalignment between 3D and 2D features. However, achieving robust cross-modal\nalignment remains a challenge due to semantic inconsistencies when generating\ncorresponding 3D and 2D feature pairs. To overcome this challenge, we present\nOV-SCAN, an Open-Vocabulary 3D framework that enforces Semantically Consistent\nAlignment for Novel object discovery. OV-SCAN employs two core strategies:\ndiscovering precise 3D annotations and filtering out low-quality or corrupted\nalignment pairs (arising from 3D annotation, occlusion-induced, or\nresolution-induced noise). Extensive experiments on the nuScenes dataset\ndemonstrate that OV-SCAN achieves state-of-the-art performance.\n", "date": "2025-03-09", "categories": []} {"id": "2503.06467", "title": "SP3D: Boosting Sparsely-Supervised 3D Object Detection via Accurate\n Cross-Modal Semantic Prompts", "abstract": " Recently, sparsely-supervised 3D object detection has gained great attention,\nachieving performance close to fully-supervised 3D objectors while requiring\nonly a few annotated instances. Nevertheless, these methods suffer challenges\nwhen accurate labels are extremely absent. In this paper, we propose a boosting\nstrategy, termed SP3D, explicitly utilizing the cross-modal semantic prompts\ngenerated from Large Multimodal Models (LMMs) to boost the 3D detector with\nrobust feature discrimination capability under sparse annotation settings.\nSpecifically, we first develop a Confident Points Semantic Transfer (CPST)\nmodule that generates accurate cross-modal semantic prompts through\nboundary-constrained center cluster selection. Based on these accurate semantic\nprompts, which we treat as seed points, we introduce a Dynamic Cluster\nPseudo-label Generation (DCPG) module to yield pseudo-supervision signals from\nthe geometry shape of multi-scale neighbor points. Additionally, we design a\nDistribution Shape score (DS score) that chooses high-quality supervision\nsignals for the initial training of the 3D detector. Experiments on the KITTI\ndataset and Waymo Open Dataset (WOD) have validated that SP3D can enhance the\nperformance of sparsely supervised detectors by a large margin under meager\nlabeling conditions. Moreover, we verified SP3D in the zero-shot setting, where\nits performance exceeded that of the state-of-the-art methods. The code is\navailable at https://github.com/xmuqimingxia/SP3D.\n", "date": "2025-03-09", "categories": []} {"id": "2503.07133", "title": "A Light Perspective for 3D Object Detection", "abstract": " Comprehending the environment and accurately detecting objects in 3D space\nare essential for advancing autonomous vehicle technologies. Integrating Camera\nand LIDAR data has emerged as an effective approach for achieving high accuracy\nin 3D Object Detection models. However, existing methodologies often rely on\nheavy, traditional backbones that are computationally demanding. This paper\nintroduces a novel approach that incorporates cutting-edge Deep Learning\ntechniques into the feature extraction process, aiming to create more efficient\nmodels without compromising performance. Our model, NextBEV, surpasses\nestablished feature extractors like ResNet50 and MobileNetV2. On the KITTI 3D\nMonocular detection benchmark, NextBEV achieves an accuracy improvement of\n2.39%, having less than 10% of the MobileNetV3 parameters. Moreover, we propose\nchanges in LIDAR backbones that decreased the original inference time to 10 ms.\nAdditionally, by fusing these lightweight proposals, we have enhanced the\naccuracy of the VoxelNet-based model by 2.93% and improved the F1-score of the\nPointPillar-based model by approximately 20%. Therefore, this work contributes\nto establishing lightweight and powerful models for individual or fusion\ntechniques, making them more suitable for onboard implementations.\n", "date": "2025-03-10", "categories": []} {"id": "2503.07593", "title": "Hierarchical Cross-Modal Alignment for Open-Vocabulary 3D Object\n Detection", "abstract": " Open-vocabulary 3D object detection (OV-3DOD) aims at localizing and\nclassifying novel objects beyond closed sets. The recent success of\nvision-language models (VLMs) has demonstrated their remarkable capabilities to\nunderstand open vocabularies. Existing works that leverage VLMs for 3D object\ndetection (3DOD) generally resort to representations that lose the rich scene\ncontext required for 3D perception. To address this problem, we propose in this\npaper a hierarchical framework, named HCMA, to simultaneously learn local\nobject and global scene information for OV-3DOD. Specifically, we first design\na Hierarchical Data Integration (HDI) approach to obtain coarse-to-fine\n3D-image-text data, which is fed into a VLM to extract object-centric\nknowledge. To facilitate the association of feature hierarchies, we then\npropose an Interactive Cross-Modal Alignment (ICMA) strategy to establish\neffective intra-level and inter-level feature connections. To better align\nfeatures across different levels, we further propose an Object-Focusing Context\nAdjustment (OFCA) module to refine multi-level features by emphasizing\nobject-related features. Extensive experiments demonstrate that the proposed\nmethod outperforms SOTA methods on the existing OV-3DOD benchmarks. It also\nachieves promising OV-3DOD results even without any 3D annotations.\n", "date": "2025-03-10", "categories": []} {"id": "2503.03548", "title": "Simulation-Based Performance Evaluation of 3D Object Detection Methods\n with Deep Learning for a LiDAR Point Cloud Dataset in a SOTIF-related Use\n Case", "abstract": " Safety of the Intended Functionality (SOTIF) addresses sensor performance\nlimitations and deep learning-based object detection insufficiencies to ensure\nthe intended functionality of Automated Driving Systems (ADS). This paper\npresents a methodology examining the adaptability and performance evaluation of\nthe 3D object detection methods on a LiDAR point cloud dataset generated by\nsimulating a SOTIF-related Use Case. The major contributions of this paper\ninclude defining and modelling a SOTIF-related Use Case with 21 diverse weather\nconditions and generating a LiDAR point cloud dataset suitable for application\nof 3D object detection methods. The dataset consists of 547 frames,\nencompassing clear, cloudy, rainy weather conditions, corresponding to\ndifferent times of the day, including noon, sunset, and night. Employing\nMMDetection3D and OpenPCDET toolkits, the performance of State-of-the-Art\n(SOTA) 3D object detection methods is evaluated and compared by testing the\npre-trained Deep Learning (DL) models on the generated dataset using Average\nPrecision (AP) and Recall metrics.\n", "date": "2025-03-05", "categories": []} {"id": "2401.02702", "title": "VoxelNextFusion: A Simple, Unified and Effective Voxel Fusion Framework\n for Multi-Modal 3D Object Detection", "abstract": " LiDAR-camera fusion can enhance the performance of 3D object detection by\nutilizing complementary information between depth-aware LiDAR points and\nsemantically rich images. Existing voxel-based methods face significant\nchallenges when fusing sparse voxel features with dense image features in a\none-to-one manner, resulting in the loss of the advantages of images, including\nsemantic and continuity information, leading to sub-optimal detection\nperformance, especially at long distances. In this paper, we present\nVoxelNextFusion, a multi-modal 3D object detection framework specifically\ndesigned for voxel-based methods, which effectively bridges the gap between\nsparse point clouds and dense images. In particular, we propose a voxel-based\nimage pipeline that involves projecting point clouds onto images to obtain both\npixel- and patch-level features. These features are then fused using a\nself-attention to obtain a combined representation. Moreover, to address the\nissue of background features present in patches, we propose a feature\nimportance module that effectively distinguishes between foreground and\nbackground features, thus minimizing the impact of the background features.\nExtensive experiments were conducted on the widely used KITTI and nuScenes 3D\nobject detection benchmarks. Notably, our VoxelNextFusion achieved around\n+3.20% in AP@0.7 improvement for car detection in hard level compared to the\nVoxel R-CNN baseline on the KITTI test dataset\n", "date": "2024-01-05", "categories": []} {"id": "2503.01899", "title": "FASTer: Focal Token Acquiring-and-Scaling Transformer for Long-term 3D\n Object Detection", "abstract": " Recent top-performing temporal 3D detectors based on Lidars have increasingly\nadopted region-based paradigms. They first generate coarse proposals, followed\nby encoding and fusing regional features. However, indiscriminate sampling and\nfusion often overlook the varying contributions of individual points and lead\nto exponentially increased complexity as the number of input frames grows.\nMoreover, arbitrary result-level concatenation limits the global information\nextraction. In this paper, we propose a Focal Token Acquring-and-Scaling\nTransformer (FASTer), which dynamically selects focal tokens and condenses\ntoken sequences in an adaptive and lightweight manner. Emphasizing the\ncontribution of individual tokens, we propose a simple but effective Adaptive\nScaling mechanism to capture geometric contexts while sifting out focal points.\nAdaptively storing and processing only focal points in historical frames\ndramatically reduces the overall complexity. Furthermore, a novel Grouped\nHierarchical Fusion strategy is proposed, progressively performing sequence\nscaling and Intra-Group Fusion operations to facilitate the exchange of global\nspatial and temporal information. Experiments on the Waymo Open Dataset\ndemonstrate that our FASTer significantly outperforms other state-of-the-art\ndetectors in both performance and efficiency while also exhibiting improved\nflexibility and robustness. The code is available at\nhttps://github.com/MSunDYY/FASTer.git.\n", "date": "2025-02-28", "categories": []} {"id": "2503.02201", "title": "MonoLite3D: Lightweight 3D Object Properties Estimation", "abstract": " Reliable perception of the environment plays a crucial role in enabling\nefficient self-driving vehicles. Therefore, the perception system necessitates\nthe acquisition of comprehensive 3D data regarding the surrounding objects\nwithin a specific time constrain, including their dimensions, spatial location\nand orientation. Deep learning has gained significant popularity in perception\nsystems, enabling the conversion of image features captured by a camera into\nmeaningful semantic information. This research paper introduces MonoLite3D\nnetwork, an embedded-device friendly lightweight deep learning methodology\ndesigned for hardware environments with limited resources. MonoLite3D network\nis a cutting-edge technique that focuses on estimating multiple properties of\n3D objects, encompassing their dimensions and spatial orientation, solely from\nmonocular images. This approach is specifically designed to meet the\nrequirements of resource-constrained environments, making it highly suitable\nfor deployment on devices with limited computational capabilities. The\nexperimental results validate the accuracy and efficiency of the proposed\napproach on the orientation benchmark of the KITTI dataset. It achieves an\nimpressive score of 82.27% on the moderate class and 69.81% on the hard class,\nwhile still meeting the real-time requirements.\n", "date": "2025-03-04", "categories": []} {"id": "2412.10739", "title": "DSRC: Learning Density-insensitive and Semantic-aware Collaborative\n Representation against Corruptions", "abstract": " As a potential application of Vehicle-to-Everything (V2X) communication,\nmulti-agent collaborative perception has achieved significant success in 3D\nobject detection. While these methods have demonstrated impressive results on\nstandard benchmarks, the robustness of such approaches in the face of complex\nreal-world environments requires additional verification. To bridge this gap,\nwe introduce the first comprehensive benchmark designed to evaluate the\nrobustness of collaborative perception methods in the presence of natural\ncorruptions typical of real-world environments. Furthermore, we propose DSRC, a\nrobustness-enhanced collaborative perception method aiming to learn\nDensity-insensitive and Semantic-aware collaborative Representation against\nCorruptions. DSRC consists of two key designs: i) a semantic-guided\nsparse-to-dense distillation framework, which constructs multi-view dense\nobjects painted by ground truth bounding boxes to effectively learn\ndensity-insensitive and semantic-aware collaborative representation; ii) a\nfeature-to-point cloud reconstruction approach to better fuse critical\ncollaborative representation across agents. To thoroughly evaluate DSRC, we\nconduct extensive experiments on real-world and simulated datasets. The results\ndemonstrate that our method outperforms SOTA collaborative perception methods\nin both clean and corrupted conditions. Code is available at\nhttps://github.com/Terry9a/DSRC.\n", "date": "2024-12-14", "categories": []} {"id": "2502.19630", "title": "Ev-3DOD: Pushing the Temporal Boundaries of 3D Object Detection with\n Event Cameras", "abstract": " Detecting 3D objects in point clouds plays a crucial role in autonomous\ndriving systems. Recently, advanced multi-modal methods incorporating camera\ninformation have achieved notable performance. For a safe and effective\nautonomous driving system, algorithms that excel not only in accuracy but also\nin speed and low latency are essential. However, existing algorithms fail to\nmeet these requirements due to the latency and bandwidth limitations of fixed\nframe rate sensors, e.g., LiDAR and camera. To address this limitation, we\nintroduce asynchronous event cameras into 3D object detection for the first\ntime. We leverage their high temporal resolution and low bandwidth to enable\nhigh-speed 3D object detection. Our method enables detection even during\ninter-frame intervals when synchronized data is unavailable, by retrieving\nprevious 3D information through the event camera. Furthermore, we introduce the\nfirst event-based 3D object detection dataset, DSEC-3DOD, which includes\nground-truth 3D bounding boxes at 100 FPS, establishing the first benchmark for\nevent-based 3D detectors. The code and dataset are available at\nhttps://github.com/mickeykang16/Ev3DOD.\n", "date": "2025-02-26", "categories": []} {"id": "2403.05061", "title": "RadarDistill: Boosting Radar-based Object Detection Performance via\n Knowledge Distillation from LiDAR Features", "abstract": " The inherent noisy and sparse characteristics of radar data pose challenges\nin finding effective representations for 3D object detection. In this paper, we\npropose RadarDistill, a novel knowledge distillation (KD) method, which can\nimprove the representation of radar data by leveraging LiDAR data. RadarDistill\nsuccessfully transfers desirable characteristics of LiDAR features into radar\nfeatures using three key components: Cross-Modality Alignment (CMA),\nActivation-based Feature Distillation (AFD), and Proposal-based Feature\nDistillation (PFD). CMA enhances the density of radar features by employing\nmultiple layers of dilation operations, effectively addressing the challenge of\ninefficient knowledge transfer from LiDAR to radar. AFD selectively transfers\nknowledge based on regions of the LiDAR features, with a specific focus on\nareas where activation intensity exceeds a predefined threshold. PFD similarly\nguides the radar network to selectively mimic features from the LiDAR network\nwithin the object proposals. Our comparative analyses conducted on the nuScenes\ndatasets demonstrate that RadarDistill achieves state-of-the-art (SOTA)\nperformance for radar-only object detection task, recording 20.5% in mAP and\n43.7% in NDS. Also, RadarDistill significantly improves the performance of the\ncamera-radar fusion model.\n", "date": "2024-03-08", "categories": []} {"id": "2502.15342", "title": "PFSD: A Multi-Modal Pedestrian-Focus Scene Dataset for Rich Tasks in\n Semi-Structured Environments", "abstract": " Recent advancements in autonomous driving perception have revealed\nexceptional capabilities within structured environments dominated by vehicular\ntraffic. However, current perception models exhibit significant limitations in\nsemi-structured environments, where dynamic pedestrians with more diverse\nirregular movement and occlusion prevail. We attribute this shortcoming to the\nscarcity of high-quality datasets in semi-structured scenes, particularly\nconcerning pedestrian perception and prediction. In this work, we present the\nmulti-modal Pedestrian-Focused Scene Dataset(PFSD), rigorously annotated in\nsemi-structured scenes with the format of nuScenes. PFSD provides comprehensive\nmulti-modal data annotations with point cloud segmentation, detection, and\nobject IDs for tracking. It encompasses over 130,000 pedestrian instances\ncaptured across various scenarios with varying densities, movement patterns,\nand occlusions. Furthermore, to demonstrate the importance of addressing the\nchallenges posed by more diverse and complex semi-structured environments, we\npropose a novel Hybrid Multi-Scale Fusion Network (HMFN). Specifically, to\ndetect pedestrians in densely populated and occluded scenarios, our method\neffectively captures and fuses multi-scale features using a meticulously\ndesigned hybrid framework that integrates sparse and vanilla convolutions.\nExtensive experiments on PFSD demonstrate that HMFN attains improvement in mean\nAverage Precision (mAP) over existing methods, thereby underscoring its\nefficacy in addressing the challenges of 3D pedestrian detection in complex\nsemi-structured environments. Coding and benchmark are available.\n", "date": "2025-02-21", "categories": []} {"id": "2502.15516", "title": "Depth-aware Fusion Method based on Image and 4D Radar Spectrum for 3D\n Object Detection", "abstract": " Safety and reliability are crucial for the public acceptance of autonomous\ndriving. To ensure accurate and reliable environmental perception, intelligent\nvehicles must exhibit accuracy and robustness in various environments.\nMillimeter-wave radar, known for its high penetration capability, can operate\neffectively in adverse weather conditions such as rain, snow, and fog.\nTraditional 3D millimeter-wave radars can only provide range, Doppler, and\nazimuth information for objects. Although the recent emergence of 4D\nmillimeter-wave radars has added elevation resolution, the radar point clouds\nremain sparse due to Constant False Alarm Rate (CFAR) operations. In contrast,\ncameras offer rich semantic details but are sensitive to lighting and weather\nconditions. Hence, this paper leverages these two highly complementary and\ncost-effective sensors, 4D millimeter-wave radar and camera. By integrating 4D\nradar spectra with depth-aware camera images and employing attention\nmechanisms, we fuse texture-rich images with depth-rich radar data in the\nBird's Eye View (BEV) perspective, enhancing 3D object detection. Additionally,\nwe propose using GAN-based networks to generate depth images from radar spectra\nin the absence of depth sensors, further improving detection accuracy.\n", "date": "2025-02-21", "categories": []} {"id": "2409.15529", "title": "VaLID: Verification as Late Integration of Detections for LiDAR-Camera\n Fusion", "abstract": " Vehicle object detection benefits from both LiDAR and camera data, with LiDAR\noffering superior performance in many scenarios. Fusion of these modalities\nfurther enhances accuracy, but existing methods often introduce complexity or\ndataset-specific dependencies. In our study, we propose a model-adaptive\nlate-fusion method, VaLID, which validates whether each predicted bounding box\nis acceptable or not. Our method verifies the higher-performing, yet overly\noptimistic LiDAR model detections using camera detections that are obtained\nfrom either specially trained, general, or open-vocabulary models. VaLID uses a\nlightweight neural verification network trained with a high recall bias to\nreduce the false predictions made by the LiDAR detector, while still preserving\nthe true ones. Evaluating with multiple combinations of LiDAR and camera\ndetectors on the KITTI dataset, we reduce false positives by an average of\n63.9%, thus outperforming the individual detectors on 3D average precision\n(3DAP). Our approach is model-adaptive and demonstrates state-of-the-art\ncompetitive performance even when using generic camera detectors that were not\ntrained specifically for this dataset.\n", "date": "2024-09-23", "categories": []} {"id": "2502.14503", "title": "LXLv2: Enhanced LiDAR Excluded Lean 3D Object Detection with Fusion of\n 4D Radar and Camera", "abstract": " As the previous state-of-the-art 4D radar-camera fusion-based 3D object\ndetection method, LXL utilizes the predicted image depth distribution maps and\nradar 3D occupancy grids to assist the sampling-based image view\ntransformation. However, the depth prediction lacks accuracy and consistency,\nand the concatenation-based fusion in LXL impedes the model robustness. In this\nwork, we propose LXLv2, where modifications are made to overcome the\nlimitations and improve the performance. Specifically, considering the position\nerror in radar measurements, we devise a one-to-many depth supervision strategy\nvia radar points, where the radar cross section (RCS) value is further\nexploited to adjust the supervision area for object-level depth consistency.\nAdditionally, a channel and spatial attention-based fusion module named\nCSAFusion is introduced to improve feature adaptiveness. Experimental results\non the View-of-Delft and TJ4DRadSet datasets show that the proposed LXLv2 can\noutperform LXL in detection accuracy, inference speed and robustness,\ndemonstrating the effectiveness of the model.\n", "date": "2025-02-20", "categories": []} {"id": "2405.15688", "title": "UNION: Unsupervised 3D Object Detection using Object Appearance-based\n Pseudo-Classes", "abstract": " Unsupervised 3D object detection methods have emerged to leverage vast\namounts of data without requiring manual labels for training. Recent approaches\nrely on dynamic objects for learning to detect mobile objects but penalize the\ndetections of static instances during training. Multiple rounds of\nself-training are used to add detected static instances to the set of training\ntargets; this procedure to improve performance is computationally expensive. To\naddress this, we propose the method UNION. We use spatial clustering and\nself-supervised scene flow to obtain a set of static and dynamic object\nproposals from LiDAR. Subsequently, object proposals' visual appearances are\nencoded to distinguish static objects in the foreground and background by\nselecting static instances that are visually similar to dynamic objects. As a\nresult, static and dynamic mobile objects are obtained together, and existing\ndetectors can be trained with a single training. In addition, we extend 3D\nobject discovery to detection by using object appearance-based cluster labels\nas pseudo-class labels for training object classification. We conduct extensive\nexperiments on the nuScenes dataset and increase the state-of-the-art\nperformance for unsupervised 3D object discovery, i.e. UNION more than doubles\nthe average precision to 39.5. The code is available at\ngithub.com/TedLentsch/UNION.\n", "date": "2024-05-24", "categories": []} {"id": "2405.18295", "title": "Intent3D: 3D Object Detection in RGB-D Scans Based on Human Intention", "abstract": " In real-life scenarios, humans seek out objects in the 3D world to fulfill\ntheir daily needs or intentions. This inspires us to introduce 3D intention\ngrounding, a new task in 3D object detection employing RGB-D, based on human\nintention, such as \"I want something to support my back\". Closely related, 3D\nvisual grounding focuses on understanding human reference. To achieve detection\nbased on human intention, it relies on humans to observe the scene, reason out\nthe target that aligns with their intention (\"pillow\" in this case), and\nfinally provide a reference to the AI system, such as \"A pillow on the couch\".\nInstead, 3D intention grounding challenges AI agents to automatically observe,\nreason and detect the desired target solely based on human intention. To tackle\nthis challenge, we introduce the new Intent3D dataset, consisting of 44,990\nintention texts associated with 209 fine-grained classes from 1,042 scenes of\nthe ScanNet dataset. We also establish several baselines based on different\nlanguage-based 3D object detection models on our benchmark. Finally, we propose\nIntentNet, our unique approach, designed to tackle this intention-based\ndetection problem. It focuses on three key aspects: intention understanding,\nreasoning to identify object candidates, and cascaded adaptive learning that\nleverages the intrinsic priority logic of different losses for multiple\nobjective optimization. Project Page:\nhttps://weitaikang.github.io/Intent3D-webpage/\n", "date": "2024-05-28", "categories": []} {"id": "2502.13071", "title": "RobuRCDet: Enhancing Robustness of Radar-Camera Fusion in Bird's Eye\n View for 3D Object Detection", "abstract": " While recent low-cost radar-camera approaches have shown promising results in\nmulti-modal 3D object detection, both sensors face challenges from\nenvironmental and intrinsic disturbances. Poor lighting or adverse weather\nconditions degrade camera performance, while radar suffers from noise and\npositional ambiguity. Achieving robust radar-camera 3D object detection\nrequires consistent performance across varying conditions, a topic that has not\nyet been fully explored. In this work, we first conduct a systematic analysis\nof robustness in radar-camera detection on five kinds of noises and propose\nRobuRCDet, a robust object detection model in BEV. Specifically, we design a 3D\nGaussian Expansion (3DGE) module to mitigate inaccuracies in radar points,\nincluding position, Radar Cross-Section (RCS), and velocity. The 3DGE uses RCS\nand velocity priors to generate a deformable kernel map and variance for kernel\nsize adjustment and value distribution. Additionally, we introduce a\nweather-adaptive fusion module, which adaptively fuses radar and camera\nfeatures based on camera signal confidence. Extensive experiments on the\npopular benchmark, nuScenes, show that our model achieves competitive results\nin regular and noisy conditions.\n", "date": "2025-02-18", "categories": []} {"id": "2408.03677", "title": "L4DR: LiDAR-4DRadar Fusion for Weather-Robust 3D Object Detection", "abstract": " LiDAR-based vision systems are integral for 3D object detection, which is\ncrucial for autonomous navigation. However, they suffer from performance\ndegradation in adverse weather conditions due to the quality deterioration of\nLiDAR point clouds. Fusing LiDAR with the weather-robust 4D radar sensor is\nexpected to solve this problem. However, the fusion of LiDAR and 4D radar is\nchallenging because they differ significantly in terms of data quality and the\ndegree of degradation in adverse weather. To address these issues, we introduce\nL4DR, a weather-robust 3D object detection method that effectively achieves\nLiDAR and 4D Radar fusion. Our L4DR includes Multi-Modal Encoding (MME) and\nForeground-Aware Denoising (FAD) technique to reconcile sensor gaps, which is\nthe first exploration of the complementarity of early fusion between LiDAR and\n4D radar. Additionally, we design an Inter-Modal and Intra-Modal ({IM}2 )\nparallel feature extraction backbone coupled with a Multi-Scale Gated Fusion\n(MSGF) module to counteract the varying degrees of sensor degradation under\nadverse weather conditions. Experimental evaluation on a VoD dataset with\nsimulated fog proves that L4DR is more adaptable to changing weather\nconditions. It delivers a significant performance increase under different fog\nlevels, improving the 3D mAP by up to 20.0% over the traditional LiDAR-only\napproach. Moreover, the results on the K-Radar dataset validate the consistent\nperformance improvement of L4DR in real-world adverse weather conditions.\n", "date": "2024-08-07", "categories": []} {"id": "2410.01404", "title": "Gaussian-Det: Learning Closed-Surface Gaussians for 3D Object Detection", "abstract": " Skins wrapping around our bodies, leathers covering over the sofa, sheet\nmetal coating the car - it suggests that objects are enclosed by a series of\ncontinuous surfaces, which provides us with informative geometry prior for\nobjectness deduction. In this paper, we propose Gaussian-Det which leverages\nGaussian Splatting as surface representation for multi-view based 3D object\ndetection. Unlike existing monocular or NeRF-based methods which depict the\nobjects via discrete positional data, Gaussian-Det models the objects in a\ncontinuous manner by formulating the input Gaussians as feature descriptors on\na mass of partial surfaces. Furthermore, to address the numerous outliers\ninherently introduced by Gaussian splatting, we accordingly devise a Closure\nInferring Module (CIM) for the comprehensive surface-based objectness\ndeduction. CIM firstly estimates the probabilistic feature residuals for\npartial surfaces given the underdetermined nature of Gaussian Splatting, which\nare then coalesced into a holistic representation on the overall surface\nclosure of the object proposal. In this way, the surface information\nGaussian-Det exploits serves as the prior on the quality and reliability of\nobjectness and the information basis of proposal refinement. Experiments on\nboth synthetic and real-world datasets demonstrate that Gaussian-Det\noutperforms various existing approaches, in terms of both average precision and\nrecall.\n", "date": "2024-10-02", "categories": []} {"id": "2502.07822", "title": "PDM-SSD: Single-Stage Three-Dimensional Object Detector With Point\n Dilation", "abstract": " Current Point-based detectors can only learn from the provided points, with\nlimited receptive fields and insufficient global learning capabilities for such\ntargets. In this paper, we present a novel Point Dilation Mechanism for\nsingle-stage 3D detection (PDM-SSD) that takes advantage of these two\nrepresentations. Specifically, we first use a PointNet-style 3D backbone for\nefficient feature encoding. Then, a neck with Point Dilation Mechanism (PDM) is\nused to expand the feature space, which involves two key steps: point dilation\nand feature filling. The former expands points to a certain size grid centered\naround the sampled points in Euclidean space. The latter fills the unoccupied\ngrid with feature for backpropagation using spherical harmonic coefficients and\nGaussian density function in terms of direction and scale. Next, we associate\nmultiple dilation centers and fuse coefficients to obtain sparse grid features\nthrough height compression. Finally, we design a hybrid detection head for\njoint learning, where on one hand, the scene heatmap is predicted to complement\nthe voting point set for improved detection accuracy, and on the other hand,\nthe target probability of detected boxes are calibrated through feature fusion.\nOn the challenging Karlsruhe Institute of Technology and Toyota Technological\nInstitute (KITTI) dataset, PDM-SSD achieves state-of-the-art results for\nmulti-class detection among single-modal methods with an inference speed of 68\nframes. We also demonstrate the advantages of PDM-SSD in detecting sparse and\nincomplete objects through numerous object-level instances. Additionally, PDM\ncan serve as an auxiliary network to establish a connection between sampling\npoints and object centers, thereby improving the accuracy of the model without\nsacrificing inference speed. Our code will be available at\nhttps://github.com/AlanLiangC/PDM-SSD.git.\n", "date": "2025-02-10", "categories": []} {"id": "2502.01856", "title": "Reliability-Driven LiDAR-Camera Fusion for Robust 3D Object Detection", "abstract": " Accurate and robust 3D object detection is essential for autonomous driving,\nwhere fusing data from sensors like LiDAR and camera enhances detection\naccuracy. However, sensor malfunctions such as corruption or disconnection can\ndegrade performance, and existing fusion models often struggle to maintain\nreliability when one modality fails. To address this, we propose ReliFusion, a\nnovel LiDAR-camera fusion framework operating in the bird's-eye view (BEV)\nspace. ReliFusion integrates three key components: the Spatio-Temporal Feature\nAggregation (STFA) module, which captures dependencies across frames to\nstabilize predictions over time; the Reliability module, which assigns\nconfidence scores to quantify the dependability of each modality under\nchallenging conditions; and the Confidence-Weighted Mutual Cross-Attention\n(CW-MCA) module, which dynamically balances information from LiDAR and camera\nmodalities based on these confidence scores. Experiments on the nuScenes\ndataset show that ReliFusion significantly outperforms state-of-the-art\nmethods, achieving superior robustness and accuracy in scenarios with limited\nLiDAR fields of view and severe sensor malfunctions.\n", "date": "2025-02-03", "categories": []} {"id": "2502.02322", "title": "Improving Generalization Ability for 3D Object Detection by Learning\n Sparsity-invariant Features", "abstract": " In autonomous driving, 3D object detection is essential for accurately\nidentifying and tracking objects. Despite the continuous development of various\ntechnologies for this task, a significant drawback is observed in most of\nthem-they experience substantial performance degradation when detecting objects\nin unseen domains. In this paper, we propose a method to improve the\ngeneralization ability for 3D object detection on a single domain. We primarily\nfocus on generalizing from a single source domain to target domains with\ndistinct sensor configurations and scene distributions. To learn\nsparsity-invariant features from a single source domain, we selectively\nsubsample the source data to a specific beam, using confidence scores\ndetermined by the current detector to identify the density that holds utmost\nimportance for the detector. Subsequently, we employ the teacher-student\nframework to align the Bird's Eye View (BEV) features for different point\nclouds densities. We also utilize feature content alignment (FCA) and\ngraph-based embedding relationship alignment (GERA) to instruct the detector to\nbe domain-agnostic. Extensive experiments demonstrate that our method exhibits\nsuperior generalization capabilities compared to other baselines. Furthermore,\nour approach even outperforms certain domain adaptation methods that can access\nto the target domain data.\n", "date": "2025-02-04", "categories": []} {"id": "2502.00074", "title": "SpikingRTNH: Spiking Neural Network for 4D Radar Object Detection", "abstract": " Recently, 4D Radar has emerged as a crucial sensor for 3D object detection in\nautonomous vehicles, offering both stable perception in adverse weather and\nhigh-density point clouds for object shape recognition. However, processing\nsuch high-density data demands substantial computational resources and energy\nconsumption. We propose SpikingRTNH, the first spiking neural network (SNN) for\n3D object detection using 4D Radar data. By replacing conventional ReLU\nactivation functions with leaky integrate-and-fire (LIF) spiking neurons,\nSpikingRTNH achieves significant energy efficiency gains. Furthermore, inspired\nby human cognitive processes, we introduce biological top-down inference (BTI),\nwhich processes point clouds sequentially from higher to lower densities. This\napproach effectively utilizes points with lower noise and higher importance for\ndetection. Experiments on K-Radar dataset demonstrate that SpikingRTNH with BTI\nsignificantly reduces energy consumption by 78% while achieving comparable\ndetection performance to its ANN counterpart (51.1% AP 3D, 57.0% AP BEV). These\nresults establish the viability of SNNs for energy-efficient 4D Radar-based\nobject detection in autonomous driving systems. All codes are available at\nhttps://github.com/kaist-avelab/k-radar.\n", "date": "2025-01-31", "categories": []} {"id": "2502.00315", "title": "MonoDINO-DETR: Depth-Enhanced Monocular 3D Object Detection Using a\n Vision Foundation Model", "abstract": " This paper proposes novel methods to enhance the performance of monocular 3D\nobject detection models by leveraging the generalized feature extraction\ncapabilities of a vision foundation model. Unlike traditional CNN-based\napproaches, which often suffer from inaccurate depth estimation and rely on\nmulti-stage object detection pipelines, this study employs a Vision Transformer\n(ViT)-based foundation model as the backbone, which excels at capturing global\nfeatures for depth estimation. It integrates a detection transformer (DETR)\narchitecture to improve both depth estimation and object detection performance\nin a one-stage manner. Specifically, a hierarchical feature fusion block is\nintroduced to extract richer visual features from the foundation model, further\nenhancing feature extraction capabilities. Depth estimation accuracy is further\nimproved by incorporating a relative depth estimation model trained on\nlarge-scale data and fine-tuning it through transfer learning. Additionally,\nthe use of queries in the transformer's decoder, which consider reference\npoints and the dimensions of 2D bounding boxes, enhances recognition\nperformance. The proposed model outperforms recent state-of-the-art methods, as\ndemonstrated through quantitative and qualitative evaluations on the KITTI 3D\nbenchmark and a custom dataset collected from high-elevation racing\nenvironments. Code is available at https://github.com/JihyeokKim/MonoDINO-DETR.\n", "date": "2025-02-01", "categories": []} {"id": "2501.17977", "title": "TransRAD: Retentive Vision Transformer for Enhanced Radar Object\n Detection", "abstract": " Despite significant advancements in environment perception capabilities for\nautonomous driving and intelligent robotics, cameras and LiDARs remain\nnotoriously unreliable in low-light conditions and adverse weather, which\nlimits their effectiveness. Radar serves as a reliable and low-cost sensor that\ncan effectively complement these limitations. However, radar-based object\ndetection has been underexplored due to the inherent weaknesses of radar data,\nsuch as low resolution, high noise, and lack of visual information. In this\npaper, we present TransRAD, a novel 3D radar object detection model designed to\naddress these challenges by leveraging the Retentive Vision Transformer (RMT)\nto more effectively learn features from information-dense radar\nRange-Azimuth-Doppler (RAD) data. Our approach leverages the Retentive\nManhattan Self-Attention (MaSA) mechanism provided by RMT to incorporate\nexplicit spatial priors, thereby enabling more accurate alignment with the\nspatial saliency characteristics of radar targets in RAD data and achieving\nprecise 3D radar detection across Range-Azimuth-Doppler dimensions.\nFurthermore, we propose Location-Aware NMS to effectively mitigate the common\nissue of duplicate bounding boxes in deep radar object detection. The\nexperimental results demonstrate that TransRAD outperforms state-of-the-art\nmethods in both 2D and 3D radar detection tasks, achieving higher accuracy,\nfaster inference speed, and reduced computational complexity. Code is available\nat https://github.com/radar-lab/TransRAD\n", "date": "2025-01-29", "categories": []} {"id": "2501.17076", "title": "DINOSTAR: Deep Iterative Neural Object Detector Self-Supervised Training\n for Roadside LiDAR Applications", "abstract": " Recent advancements in deep-learning methods for object detection in\npoint-cloud data have enabled numerous roadside applications, fostering\nimprovements in transportation safety and management. However, the intricate\nnature of point-cloud data poses significant challenges for human-supervised\nlabeling, resulting in substantial expenditures of time and capital. This paper\naddresses the issue by developing an end-to-end, scalable, and self-supervised\nframework for training deep object detectors tailored for roadside point-cloud\ndata. The proposed framework leverages self-supervised, statistically modeled\nteachers to train off-the-shelf deep object detectors, thus circumventing the\nneed for human supervision. The teacher models follow fine-tuned set standard\npractices of background filtering, object clustering, bounding-box fitting, and\nclassification to generate noisy labels. It is presented that by training the\nstudent model over the combined noisy annotations from multitude of teachers\nenhances its capacity to discern background/foreground more effectively and\nforces it to learn diverse point-cloud-representations for object categories of\ninterest. The evaluations, involving publicly available roadside datasets and\nstate-of-art deep object detectors, demonstrate that the proposed framework\nachieves comparable performance to deep object detectors trained on\nhuman-annotated labels, despite not utilizing such human-annotations in its\ntraining process.\n", "date": "2025-01-28", "categories": []} {"id": "2501.14502", "title": "LiDAR-Based Vehicle Detection and Tracking for Autonomous Racing", "abstract": " Autonomous racing provides a controlled environment for testing the software\nand hardware of autonomous vehicles operating at their performance limits.\nCompetitive interactions between multiple autonomous racecars however introduce\nchallenging and potentially dangerous scenarios. Accurate and consistent\nvehicle detection and tracking is crucial for overtaking maneuvers, and\nlow-latency sensor processing is essential to respond quickly to hazardous\nsituations. This paper presents the LiDAR-based perception algorithms deployed\non Team PoliMOVE's autonomous racecar, which won multiple competitions in the\nIndy Autonomous Challenge series. Our Vehicle Detection and Tracking pipeline\nis composed of a novel fast Point Cloud Segmentation technique and a specific\nVehicle Pose Estimation methodology, together with a variable-step Multi-Target\nTracking algorithm. Experimental results demonstrate the algorithm's\nperformance, robustness, computational efficiency, and suitability for\nautonomous racing applications, enabling fully autonomous overtaking maneuvers\nat velocities exceeding 275 km/h.\n", "date": "2025-01-24", "categories": []} {"id": "2208.03974", "title": "Aerial Monocular 3D Object Detection", "abstract": " Drones equipped with cameras can significantly enhance human ability to\nperceive the world because of their remarkable maneuverability in 3D space.\nIronically, object detection for drones has always been conducted in the 2D\nimage space, which fundamentally limits their ability to understand 3D scenes.\nFurthermore, existing 3D object detection methods developed for autonomous\ndriving cannot be directly applied to drones due to the lack of deformation\nmodeling, which is essential for the distant aerial perspective with sensitive\ndistortion and small objects. To fill the gap, this work proposes a dual-view\ndetection system named DVDET to achieve aerial monocular object detection in\nboth the 2D image space and the 3D physical space. To address the severe view\ndeformation issue, we propose a novel trainable geo-deformable transformation\nmodule that can properly warp information from the drone's perspective to the\nBEV. Compared to the monocular methods for cars, our transformation includes a\nlearnable deformable network for explicitly revising the severe deviation. To\naddress the dataset challenge, we propose a new large-scale simulation dataset\nnamed AM3D-Sim, generated by the co-simulation of AirSIM and CARLA, and a new\nreal-world aerial dataset named AM3D-Real, collected by DJI Matrice 300 RTK, in\nboth datasets, high-quality annotations for 3D object detection are provided.\nExtensive experiments show that i) aerial monocular 3D object detection is\nfeasible; ii) the model pre-trained on the simulation dataset benefits\nreal-world performance, and iii) DVDET also benefits monocular 3D object\ndetection for cars. To encourage more researchers to investigate this area, we\nwill release the dataset and related code in https://github.com/PhyllisH/DVDET.\n", "date": "2022-08-08", "categories": []} {"id": "2412.09658", "title": "SEGT: A General Spatial Expansion Group Transformer for nuScenes\n Lidar-based Object Detection Task", "abstract": " In the technical report, we present a novel transformer-based framework for\nnuScenes lidar-based object detection task, termed Spatial Expansion Group\nTransformer (SEGT). To efficiently handle the irregular and sparse nature of\npoint cloud, we propose migrating the voxels into distinct specialized ordered\nfields with the general spatial expansion strategies, and employ group\nattention mechanisms to extract the exclusive feature maps within each field.\nSubsequently, we integrate the feature representations across different ordered\nfields by alternately applying diverse expansion strategies, thereby enhancing\nthe model's ability to capture comprehensive spatial information. The method\nwas evaluated on the nuScenes lidar-based object detection test dataset,\nachieving an NDS score of 73.9 without Test-Time Augmentation (TTA) and 74.5\nwith TTA, demonstrating the effectiveness of the proposed method. Notably, our\nmethod ranks the 1st place in the nuScenes lidar-based object detection task.\n", "date": "2024-12-12", "categories": []} {"id": "2501.11159", "title": "LiFT: Lightweight, FPGA-tailored 3D object detection based on LiDAR data", "abstract": " This paper presents LiFT, a lightweight, fully quantized 3D object detection\nalgorithm for LiDAR data, optimized for real-time inference on FPGA platforms.\nThrough an in-depth analysis of FPGA-specific limitations, we identify a set of\nFPGA-induced constraints that shape the algorithm's design. These include a\ncomputational complexity limit of 30 GMACs (billion multiply-accumulate\noperations), INT8 quantization for weights and activations, 2D cell-based\nprocessing instead of 3D voxels, and minimal use of skip connections. To meet\nthese constraints while maximizing performance, LiFT combines novel mechanisms\nwith state-of-the-art techniques such as reparameterizable convolutions and\nfully sparse architecture. Key innovations include the Dual-bound Pillar\nFeature Net, which boosts performance without increasing complexity, and an\nefficient scheme for INT8 quantization of input features. With a computational\ncost of just 20.73 GMACs, LiFT stands out as one of the few algorithms\ntargeting minimal-complexity 3D object detection. Among comparable methods,\nLiFT ranks first, achieving an mAP of 51.84% and an NDS of 61.01% on the\nchallenging NuScenes validation dataset. The code will be available at\nhttps://github.com/vision-agh/lift.\n", "date": "2025-01-19", "categories": []} {"id": "2412.12460", "title": "PromptDet: A Lightweight 3D Object Detection Framework with LiDAR\n Prompts", "abstract": " Multi-camera 3D object detection aims to detect and localize objects in 3D\nspace using multiple cameras, which has attracted more attention due to its\ncost-effectiveness trade-off. However, these methods often struggle with the\nlack of accurate depth estimation caused by the natural weakness of the camera\nin ranging. Recently, multi-modal fusion and knowledge distillation methods for\n3D object detection have been proposed to solve this problem, which are\ntime-consuming during the training phase and not friendly to memory cost. In\nlight of this, we propose PromptDet, a lightweight yet effective 3D object\ndetection framework motivated by the success of prompt learning in 2D\nfoundation model. Our proposed framework, PromptDet, comprises two integral\ncomponents: a general camera-based detection module, exemplified by models like\nBEVDet and BEVDepth, and a LiDAR-assisted prompter. The LiDAR-assisted prompter\nleverages the LiDAR points as a complementary signal, enriched with a minimal\nset of additional trainable parameters. Notably, our framework is flexible due\nto our prompt-like design, which can not only be used as a lightweight\nmulti-modal fusion method but also as a camera-only method for 3D object\ndetection during the inference phase. Extensive experiments on nuScenes\nvalidate the effectiveness of the proposed PromptDet. As a multi-modal\ndetector, PromptDet improves the mAP and NDS by at most 22.8\\% and 21.1\\% with\nfewer than 2\\% extra parameters compared with the camera-only baseline. Without\nLiDAR points, PromptDet still achieves an improvement of at most 2.4\\% mAP and\n4.0\\% NDS with almost no impact on camera detection inference time.\n", "date": "2024-12-17", "categories": []} {"id": "2501.06550", "title": "CoreNet: Conflict Resolution Network for Point-Pixel Misalignment and\n Sub-Task Suppression of 3D LiDAR-Camera Object Detection", "abstract": " Fusing multi-modality inputs from different sensors is an effective way to\nimprove the performance of 3D object detection. However, current methods\noverlook two important conflicts: point-pixel misalignment and sub-task\nsuppression. The former means a pixel feature from the opaque object is\nprojected to multiple point features of the same ray in the world space, and\nthe latter means the classification prediction and bounding box regression may\ncause mutual suppression. In this paper, we propose a novel method named\nConflict Resolution Network (CoreNet) to address the aforementioned issues.\nSpecifically, we first propose a dual-stream transformation module to tackle\npoint-pixel misalignment. It consists of ray-based and point-based 2D-to-BEV\ntransformations. Both of them achieve approximately unique mapping from the\nimage space to the world space. Moreover, we introduce a task-specific\npredictor to tackle sub-task suppression. It uses the dual-branch structure\nwhich adopts class-specific query and Bbox-specific query to corresponding\nsub-tasks. Each task-specific query is constructed of task-specific feature and\ngeneral feature, which allows the heads to adaptively select information of\ninterest based on different sub-tasks. Experiments on the large-scale nuScenes\ndataset demonstrate the superiority of our proposed CoreNet, by achieving\n75.6\\% NDS and 73.3\\% mAP on the nuScenes test set without test-time\naugmentation and model ensemble techniques. The ample ablation study also\ndemonstrates the effectiveness of each component. The code is released on\nhttps://github.com/liyih/CoreNet.\n", "date": "2025-01-11", "categories": []} {"id": "2501.04373", "title": "FGU3R: Fine-Grained Fusion via Unified 3D Representation for Multimodal\n 3D Object Detection", "abstract": " Multimodal 3D object detection has garnered considerable interest in\nautonomous driving. However, multimodal detectors suffer from dimension\nmismatches that derive from fusing 3D points with 2D pixels coarsely, which\nleads to sub-optimal fusion performance. In this paper, we propose a multimodal\nframework FGU3R to tackle the issue mentioned above via unified 3D\nrepresentation and fine-grained fusion, which consists of two important\ncomponents. First, we propose an efficient feature extractor for raw and pseudo\npoints, termed Pseudo-Raw Convolution (PRConv), which modulates multimodal\nfeatures synchronously and aggregates the features from different types of\npoints on key points based on multimodal interaction. Second, a Cross-Attention\nAdaptive Fusion (CAAF) is designed to fuse homogeneous 3D RoI (Region of\nInterest) features adaptively via a cross-attention variant in a fine-grained\nmanner. Together they make fine-grained fusion on unified 3D representation.\nThe experiments conducted on the KITTI and nuScenes show the effectiveness of\nour proposed method.\n", "date": "2025-01-08", "categories": []} {"id": "2310.15624", "title": "GUPNet++: Geometry Uncertainty Propagation Network for Monocular 3D\n Object Detection", "abstract": " Geometry plays a significant role in monocular 3D object detection. It can be\nused to estimate object depth by using the perspective projection between\nobject's physical size and 2D projection in the image plane, which can\nintroduce mathematical priors into deep models. However, this projection\nprocess also introduces error amplification, where the error of the estimated\nheight is amplified and reflected into the projected depth. It leads to\nunreliable depth inferences and also impairs training stability. To tackle this\nproblem, we propose a novel Geometry Uncertainty Propagation Network (GUPNet++)\nby modeling geometry projection in a probabilistic manner. This ensures depth\npredictions are well-bounded and associated with a reasonable uncertainty. The\nsignificance of introducing such geometric uncertainty is two-fold: (1). It\nmodels the uncertainty propagation relationship of the geometry projection\nduring training, improving the stability and efficiency of the end-to-end model\nlearning. (2). It can be derived to a highly reliable confidence to indicate\nthe quality of the 3D detection result, enabling more reliable detection\ninference. Experiments show that the proposed approach not only obtains\n(state-of-the-art) SOTA performance in image-based monocular 3D detection but\nalso demonstrates superiority in efficacy with a simplified framework.\n", "date": "2023-10-24", "categories": []} {"id": "2501.03700", "title": "AuxDepthNet: Real-Time Monocular 3D Object Detection with\n Depth-Sensitive Features", "abstract": " Monocular 3D object detection is a challenging task in autonomous systems due\nto the lack of explicit depth information in single-view images. Existing\nmethods often depend on external depth estimators or expensive sensors, which\nincrease computational complexity and hinder real-time performance. To overcome\nthese limitations, we propose AuxDepthNet, an efficient framework for real-time\nmonocular 3D object detection that eliminates the reliance on external depth\nmaps or pre-trained depth models. AuxDepthNet introduces two key components:\nthe Auxiliary Depth Feature (ADF) module, which implicitly learns\ndepth-sensitive features to improve spatial reasoning and computational\nefficiency, and the Depth Position Mapping (DPM) module, which embeds depth\npositional information directly into the detection process to enable accurate\nobject localization and 3D bounding box regression. Leveraging the DepthFusion\nTransformer architecture, AuxDepthNet globally integrates visual and\ndepth-sensitive features through depth-guided interactions, ensuring robust and\nefficient detection. Extensive experiments on the KITTI dataset show that\nAuxDepthNet achieves state-of-the-art performance, with $\\text{AP}_{3D}$ scores\nof 24.72\\% (Easy), 18.63\\% (Moderate), and 15.31\\% (Hard), and\n$\\text{AP}_{\\text{BEV}}$ scores of 34.11\\% (Easy), 25.18\\% (Moderate), and\n21.90\\% (Hard) at an IoU threshold of 0.7.\n", "date": "2025-01-07", "categories": []} {"id": "2501.02314", "title": "RadarNeXt: Real-Time and Reliable 3D Object Detector Based On 4D mmWave\n Imaging Radar", "abstract": " 3D object detection is crucial for Autonomous Driving (AD) and Advanced\nDriver Assistance Systems (ADAS). However, most 3D detectors prioritize\ndetection accuracy, often overlooking network inference speed in practical\napplications. In this paper, we propose RadarNeXt, a real-time and reliable 3D\nobject detector based on the 4D mmWave radar point clouds. It leverages the\nre-parameterizable neural networks to catch multi-scale features, reduce memory\ncost and accelerate the inference. Moreover, to highlight the irregular\nforeground features of radar point clouds and suppress background clutter, we\npropose a Multi-path Deformable Foreground Enhancement Network (MDFEN),\nensuring detection accuracy while minimizing the sacrifice of speed and\nexcessive number of parameters. Experimental results on View-of-Delft and\nTJ4DRadSet datasets validate the exceptional performance and efficiency of\nRadarNeXt, achieving 50.48 and 32.30 mAPs with the variant using our proposed\nMDFEN. Notably, our RadarNeXt variants achieve inference speeds of over 67.10\nFPS on the RTX A4000 GPU and 28.40 FPS on the Jetson AGX Orin. This research\ndemonstrates that RadarNeXt brings a novel and effective paradigm for 3D\nperception based on 4D mmWave radar.\n", "date": "2025-01-04", "categories": []} {"id": "2501.00220", "title": "DecoratingFusion: A LiDAR-Camera Fusion Network with the Combination of\n Point-level and Feature-level Fusion", "abstract": " Lidars and cameras play essential roles in autonomous driving, offering\ncomplementary information for 3D detection. The state-of-the-art fusion methods\nintegrate them at the feature level, but they mostly rely on the learned soft\nassociation between point clouds and images, which lacks interpretability and\nneglects the hard association between them. In this paper, we combine\nfeature-level fusion with point-level fusion, using hard association\nestablished by the calibration matrices to guide the generation of object\nqueries. Specifically, in the early fusion stage, we use the 2D CNN features of\nimages to decorate the point cloud data, and employ two independent sparse\nconvolutions to extract the decorated point cloud features. In the mid-level\nfusion stage, we initialize the queries with a center heatmap and embed the\npredicted class labels as auxiliary information into the queries, making the\ninitial positions closer to the actual centers of the targets. Extensive\nexperiments conducted on two popular datasets, i.e. KITTI, Waymo, demonstrate\nthe superiority of DecoratingFusion.\n", "date": "2024-12-31", "categories": []} {"id": "2404.13921", "title": "NeRF-DetS: Enhanced Adaptive Spatial-wise Sampling and View-wise Fusion\n Strategies for NeRF-based Indoor Multi-view 3D Object Detection", "abstract": " In indoor scenes, the diverse distribution of object locations and scales\nmakes the visual 3D perception task a big challenge.\n Previous works (e.g, NeRF-Det) have demonstrated that implicit representation\nhas the capacity to benefit the visual 3D perception task in indoor scenes with\nhigh amount of overlap between input images.\n However, previous works cannot fully utilize the advancement of implicit\nrepresentation because of fixed sampling and simple multi-view feature fusion.\n In this paper, inspired by sparse fashion method (e.g, DETR3D), we propose a\nsimple yet effective method, NeRF-DetS, to address above issues. NeRF-DetS\nincludes two modules: Progressive Adaptive Sampling Strategy (PASS) and\nDepth-Guided Simplified Multi-Head Attention Fusion (DS-MHA).\n Specifically,\n (1)PASS can automatically sample features of each layer within a dense 3D\ndetector, using offsets predicted by the previous layer.\n (2)DS-MHA can not only efficiently fuse multi-view features with strong\nocclusion awareness but also reduce computational cost.\n Extensive experiments on ScanNetV2 dataset demonstrate our NeRF-DetS\noutperforms NeRF-Det, by achieving +5.02% and +5.92% improvement in mAP under\nIoU25 and IoU50, respectively. Also, NeRF-DetS shows consistent improvements on\nARKITScenes.\n", "date": "2024-04-22", "categories": []} {"id": "2412.20911", "title": "TiGDistill-BEV: Multi-view BEV 3D Object Detection via Target\n Inner-Geometry Learning Distillation", "abstract": " Accurate multi-view 3D object detection is essential for applications such as\nautonomous driving. Researchers have consistently aimed to leverage LiDAR's\nprecise spatial information to enhance camera-based detectors through methods\nlike depth supervision and bird-eye-view (BEV) feature distillation. However,\nexisting approaches often face challenges due to the inherent differences\nbetween LiDAR and camera data representations. In this paper, we introduce the\nTiGDistill-BEV, a novel approach that effectively bridges this gap by\nleveraging the strengths of both sensors. Our method distills knowledge from\ndiverse modalities(e.g., LiDAR) as the teacher model to a camera-based student\ndetector, utilizing the Target Inner-Geometry learning scheme to enhance\ncamera-based BEV detectors through both depth and BEV features by leveraging\ndiverse modalities. Specially, we propose two key modules: an inner-depth\nsupervision module to learn the low-level relative depth relations within\nobjects which equips detectors with a deeper understanding of object-level\nspatial structures, and an inner-feature BEV distillation module to transfer\nhigh-level semantics of different key points within foreground targets. To\nfurther alleviate the domain gap, we incorporate both inter-channel and\ninter-keypoint distillation to model feature similarity. Extensive experiments\non the nuScenes benchmark demonstrate that TiGDistill-BEV significantly boosts\ncamera-based only detectors achieving a state-of-the-art with 62.8% NDS and\nsurpassing previous methods by a significant margin. The codes is available at:\nhttps://github.com/Public-BOTs/TiGDistill-BEV.git.\n", "date": "2024-12-30", "categories": []} {"id": "2409.01816", "title": "GeoBEV: Learning Geometric BEV Representation for Multi-view 3D Object\n Detection", "abstract": " Bird's-Eye-View (BEV) representation has emerged as a mainstream paradigm for\nmulti-view 3D object detection, demonstrating impressive perceptual\ncapabilities. However, existing methods overlook the geometric quality of BEV\nrepresentation, leaving it in a low-resolution state and failing to restore the\nauthentic geometric information of the scene. In this paper, we identify the\ndrawbacks of previous approaches that limit the geometric quality of BEV\nrepresentation and propose Radial-Cartesian BEV Sampling (RC-Sampling), which\noutperforms other feature transformation methods in efficiently generating\nhigh-resolution dense BEV representation to restore fine-grained geometric\ninformation. Additionally, we design a novel In-Box Label to substitute the\ntraditional depth label generated from the LiDAR points. This label reflects\nthe actual geometric structure of objects rather than just their surfaces,\ninjecting real-world geometric information into the BEV representation. In\nconjunction with the In-Box Label, Centroid-Aware Inner Loss (CAI Loss) is\ndeveloped to capture the inner geometric structure of objects. Finally, we\nintegrate the aforementioned modules into a novel multi-view 3D object\ndetector, dubbed GeoBEV, which achieves a state-of-the-art result of 66.2\\% NDS\non the nuScenes test set. The code is available at\nhttps://github.com/mengtan00/GeoBEV.git.\n", "date": "2024-09-03", "categories": []} {"id": "2412.16947", "title": "Separating Drone Point Clouds From Complex Backgrounds by Cluster Filter\n -- Technical Report for CVPR 2024 UG2 Challenge", "abstract": " The increasing deployment of small drones as tools of conflict and disruption\nhas amplified their threat, highlighting the urgent need for effective\nanti-drone measures. However, the compact size of most drones presents a\nsignificant challenge, as traditional supervised point cloud or image-based\nobject detection methods often fail to identify such small objects effectively.\nThis paper proposes a simple UAV detection method using an unsupervised\npipeline. It uses spatial-temporal sequence processing to fuse multiple lidar\ndatasets effectively, tracking and determining the position of UAVs, so as to\ndetect and track UAVs in challenging environments. Our method performs front\nand rear background segmentation of point clouds through a global-local\nsequence clusterer and parses point cloud data from both the spatial-temporal\ndensity and spatial-temporal voxels of the point cloud. Furthermore, a scoring\nmechanism for point cloud moving targets is proposed, using time series\ndetection to improve accuracy and efficiency. We used the MMAUD dataset, and\nour method achieved 4th place in the CVPR 2024 UG2+ Challenge, confirming the\neffectiveness of our method in practical applications.\n", "date": "2024-12-22", "categories": []} {"id": "2412.14571", "title": "SCKD: Semi-Supervised Cross-Modality Knowledge Distillation for 4D Radar\n Object Detection", "abstract": " 3D object detection is one of the fundamental perception tasks for autonomous\nvehicles. Fulfilling such a task with a 4D millimeter-wave radar is very\nattractive since the sensor is able to acquire 3D point clouds similar to Lidar\nwhile maintaining robust measurements under adverse weather. However, due to\nthe high sparsity and noise associated with the radar point clouds, the\nperformance of the existing methods is still much lower than expected. In this\npaper, we propose a novel Semi-supervised Cross-modality Knowledge Distillation\n(SCKD) method for 4D radar-based 3D object detection. It characterizes the\ncapability of learning the feature from a Lidar-radar-fused teacher network\nwith semi-supervised distillation. We first propose an adaptive fusion module\nin the teacher network to boost its performance. Then, two feature distillation\nmodules are designed to facilitate the cross-modality knowledge transfer.\nFinally, a semi-supervised output distillation is proposed to increase the\neffectiveness and flexibility of the distillation framework. With the same\nnetwork structure, our radar-only student trained by SCKD boosts the mAP by\n10.38% over the baseline and outperforms the state-of-the-art works on the VoD\ndataset. The experiment on ZJUODset also shows 5.12% mAP improvements on the\nmoderate difficulty level over the baseline when extra unlabeled data are\navailable. Code is available at https://github.com/Ruoyu-Xu/SCKD.\n", "date": "2024-12-19", "categories": []} {"id": "2412.12799", "title": "RCTrans: Radar-Camera Transformer via Radar Densifier and Sequential\n Decoder for 3D Object Detection", "abstract": " In radar-camera 3D object detection, the radar point clouds are sparse and\nnoisy, which causes difficulties in fusing camera and radar modalities. To\nsolve this, we introduce a novel query-based detection method named\nRadar-Camera Transformer (RCTrans). Specifically, we first design a Radar Dense\nEncoder to enrich the sparse valid radar tokens, and then concatenate them with\nthe image tokens. By doing this, we can fully explore the 3D information of\neach interest region and reduce the interference of empty tokens during the\nfusing stage. We then design a Pruning Sequential Decoder to predict 3D boxes\nbased on the obtained tokens and random initialized queries. To alleviate the\neffect of elevation ambiguity in radar point clouds, we gradually locate the\nposition of the object via a sequential fusion structure. It helps to get more\nprecise and flexible correspondences between tokens and queries. A pruning\ntraining strategy is adopted in the decoder, which can save much time during\ninference and inhibit queries from losing their distinctiveness. Extensive\nexperiments on the large-scale nuScenes dataset prove the superiority of our\nmethod, and we also achieve new state-of-the-art radar-camera 3D detection\nresults. Our implementation is available at https://github.com/liyih/RCTrans.\n", "date": "2024-12-17", "categories": []} {"id": "2412.13017", "title": "A New Adversarial Perspective for LiDAR-based 3D Object Detection", "abstract": " Autonomous vehicles (AVs) rely on LiDAR sensors for environmental perception\nand decision-making in driving scenarios. However, ensuring the safety and\nreliability of AVs in complex environments remains a pressing challenge. To\naddress this issue, we introduce a real-world dataset (ROLiD) comprising\nLiDAR-scanned point clouds of two random objects: water mist and smoke. In this\npaper, we introduce a novel adversarial perspective by proposing an attack\nframework that utilizes water mist and smoke to simulate environmental\ninterference. Specifically, we propose a point cloud sequence generation method\nusing a motion and content decomposition generative adversarial network named\nPCS-GAN to simulate the distribution of random objects. Furthermore, leveraging\nthe simulated LiDAR scanning characteristics implemented with Range Image, we\nexamine the effects of introducing random object perturbations at various\npositions on the target vehicle. Extensive experiments demonstrate that\nadversarial perturbations based on random objects effectively deceive vehicle\ndetection and reduce the recognition rate of 3D object detection models.\n", "date": "2024-12-17", "categories": []} {"id": "2412.11412", "title": "V-MIND: Building Versatile Monocular Indoor 3D Detector with Diverse 2D\n Annotations", "abstract": " The field of indoor monocular 3D object detection is gaining significant\nattention, fueled by the increasing demand in VR/AR and robotic applications.\nHowever, its advancement is impeded by the limited availability and diversity\nof 3D training data, owing to the labor-intensive nature of 3D data collection\nand annotation processes. In this paper, we present V-MIND (Versatile Monocular\nINdoor Detector), which enhances the performance of indoor 3D detectors across\na diverse set of object classes by harnessing publicly available large-scale 2D\ndatasets. By leveraging well-established monocular depth estimation techniques\nand camera intrinsic predictors, we can generate 3D training data by converting\nlarge-scale 2D images into 3D point clouds and subsequently deriving pseudo 3D\nbounding boxes. To mitigate distance errors inherent in the converted point\nclouds, we introduce a novel 3D self-calibration loss for refining the pseudo\n3D bounding boxes during training. Additionally, we propose a novel ambiguity\nloss to address the ambiguity that arises when introducing new classes from 2D\ndatasets. Finally, through joint training with existing 3D datasets and pseudo\n3D bounding boxes derived from 2D datasets, V-MIND achieves state-of-the-art\nobject detection performance across a wide range of classes on the Omni3D\nindoor dataset.\n", "date": "2024-12-16", "categories": []} {"id": "2412.11489", "title": "HGSFusion: Radar-Camera Fusion with Hybrid Generation and\n Synchronization for 3D Object Detection", "abstract": " Millimeter-wave radar plays a vital role in 3D object detection for\nautonomous driving due to its all-weather and all-lighting-condition\ncapabilities for perception. However, radar point clouds suffer from pronounced\nsparsity and unavoidable angle estimation errors. To address these limitations,\nincorporating a camera may partially help mitigate the shortcomings.\nNevertheless, the direct fusion of radar and camera data can lead to negative\nor even opposite effects due to the lack of depth information in images and\nlow-quality image features under adverse lighting conditions. Hence, in this\npaper, we present the radar-camera fusion network with Hybrid Generation and\nSynchronization (HGSFusion), designed to better fuse radar potentials and image\nfeatures for 3D object detection. Specifically, we propose the Radar Hybrid\nGeneration Module (RHGM), which fully considers the Direction-Of-Arrival (DOA)\nestimation errors in radar signal processing. This module generates denser\nradar points through different Probability Density Functions (PDFs) with the\nassistance of semantic information. Meanwhile, we introduce the Dual Sync\nModule (DSM), comprising spatial sync and modality sync, to enhance image\nfeatures with radar positional information and facilitate the fusion of\ndistinct characteristics in different modalities. Extensive experiments\ndemonstrate the effectiveness of our approach, outperforming the\nstate-of-the-art methods in the VoD and TJ4DRadSet datasets by $6.53\\%$ and\n$2.03\\%$ in RoI AP and BEV AP, respectively. The code is available at\nhttps://github.com/garfield-cpp/HGSFusion.\n", "date": "2024-12-16", "categories": []} {"id": "2412.08806", "title": "DALI: Domain Adaptive LiDAR Object Detection via Distribution-level and\n Instance-level Pseudo Label Denoising", "abstract": " Object detection using LiDAR point clouds relies on a large amount of\nhuman-annotated samples when training the underlying detectors' deep neural\nnetworks. However, generating 3D bounding box annotation for a large-scale\ndataset could be costly and time-consuming. Alternatively, unsupervised domain\nadaptation (UDA) enables a given object detector to operate on a novel new\ndata, with unlabeled training dataset, by transferring the knowledge learned\nfrom training labeled \\textit{source domain} data to the new unlabeled\n\\textit{target domain}. Pseudo label strategies, which involve training the 3D\nobject detector using target-domain predicted bounding boxes from a pre-trained\nmodel, are commonly used in UDA. However, these pseudo labels often introduce\nnoise, impacting performance. In this paper, we introduce the Domain Adaptive\nLIdar (DALI) object detection framework to address noise at both distribution\nand instance levels. Firstly, a post-training size normalization (PTSN)\nstrategy is developed to mitigate bias in pseudo label size distribution by\nidentifying an unbiased scale after network training. To address instance-level\nnoise between pseudo labels and corresponding point clouds, two pseudo point\nclouds generation (PPCG) strategies, ray-constrained and constraint-free, are\ndeveloped to generate pseudo point clouds for each instance, ensuring the\nconsistency between pseudo labels and pseudo points during training. We\ndemonstrate the effectiveness of our method on the publicly available and\npopular datasets KITTI, Waymo, and nuScenes. We show that the proposed DALI\nframework achieves state-of-the-art results and outperforms leading approaches\non most of the domain adaptation tasks. Our code is available at\n\\href{https://github.com/xiaohulugo/T-RO2024-DALI}{https://github.com/xiaohulugo/T-RO2024-DALI}.\n", "date": "2024-12-11", "categories": []} {"id": "2411.03013", "title": "CRT-Fusion: Camera, Radar, Temporal Fusion Using Motion Information for\n 3D Object Detection", "abstract": " Accurate and robust 3D object detection is a critical component in autonomous\nvehicles and robotics. While recent radar-camera fusion methods have made\nsignificant progress by fusing information in the bird's-eye view (BEV)\nrepresentation, they often struggle to effectively capture the motion of\ndynamic objects, leading to limited performance in real-world scenarios. In\nthis paper, we introduce CRT-Fusion, a novel framework that integrates temporal\ninformation into radar-camera fusion to address this challenge. Our approach\ncomprises three key modules: Multi-View Fusion (MVF), Motion Feature Estimator\n(MFE), and Motion Guided Temporal Fusion (MGTF). The MVF module fuses radar and\nimage features within both the camera view and bird's-eye view, thereby\ngenerating a more precise unified BEV representation. The MFE module conducts\ntwo simultaneous tasks: estimation of pixel-wise velocity information and BEV\nsegmentation. Based on the velocity and the occupancy score map obtained from\nthe MFE module, the MGTF module aligns and fuses feature maps across multiple\ntimestamps in a recurrent manner. By considering the motion of dynamic objects,\nCRT-Fusion can produce robust BEV feature maps, thereby improving detection\naccuracy and robustness. Extensive evaluations on the challenging nuScenes\ndataset demonstrate that CRT-Fusion achieves state-of-the-art performance for\nradar-camera-based 3D object detection. Our approach outperforms the previous\nbest method in terms of NDS by +1.7%, while also surpassing the leading\napproach in mAP by +1.4%. These significant improvements in both metrics\nshowcase the effectiveness of our proposed fusion strategy in enhancing the\nreliability and accuracy of 3D object detection.\n", "date": "2024-11-05", "categories": []} {"id": "2410.07475", "title": "Progressive Multi-Modal Fusion for Robust 3D Object Detection", "abstract": " Multi-sensor fusion is crucial for accurate 3D object detection in autonomous\ndriving, with cameras and LiDAR being the most commonly used sensors. However,\nexisting methods perform sensor fusion in a single view by projecting features\nfrom both modalities either in Bird's Eye View (BEV) or Perspective View (PV),\nthus sacrificing complementary information such as height or geometric\nproportions. To address this limitation, we propose ProFusion3D, a progressive\nfusion framework that combines features in both BEV and PV at both intermediate\nand object query levels. Our architecture hierarchically fuses local and global\nfeatures, enhancing the robustness of 3D object detection. Additionally, we\nintroduce a self-supervised mask modeling pre-training strategy to improve\nmulti-modal representation learning and data efficiency through three novel\nobjectives. Extensive experiments on nuScenes and Argoverse2 datasets\nconclusively demonstrate the efficacy of ProFusion3D. Moreover, ProFusion3D is\nrobust to sensor failure, demonstrating strong performance when only one\nmodality is available.\n", "date": "2024-10-09", "categories": []} {"id": "2412.07411", "title": "DSFEC: Efficient and Deployable Deep Radar Object Detection", "abstract": " Deploying radar object detection models on resource-constrained edge devices\nlike the Raspberry Pi poses significant challenges due to the large size of the\nmodel and the limited computational power and the memory of the Pi. In this\nwork, we explore the efficiency of Depthwise Separable Convolutions in radar\nobject detection networks and integrate them into our model. Additionally, we\nintroduce a novel Feature Enhancement and Compression (FEC) module to the\nPointPillars feature encoder to further improve the model performance. With\nthese innovations, we propose the DSFEC-L model and its two versions, which\noutperform the baseline (23.9 mAP of Car class, 20.72 GFLOPs) on nuScenes\ndataset: 1). An efficient DSFEC-M model with a 14.6% performance improvement\nand a 60% reduction in GFLOPs. 2). A deployable DSFEC-S model with a 3.76%\nperformance improvement and a remarkable 78.5% reduction in GFLOPs. Despite\nmarginal performance gains, our deployable model achieves an impressive 74.5%\nreduction in runtime on the Raspberry Pi compared to the baseline.\n", "date": "2024-12-10", "categories": []} {"id": "2412.07509", "title": "Enhancing 3D Object Detection in Autonomous Vehicles Based on Synthetic\n Virtual Environment Analysis", "abstract": " Autonomous Vehicles (AVs) use natural images and videos as input to\nunderstand the real world by overlaying and inferring digital elements,\nfacilitating proactive detection in an effort to assure safety. A crucial\naspect of this process is real-time, accurate object recognition through\nautomatic scene analysis. While traditional methods primarily concentrate on 2D\nobject detection, exploring 3D object detection, which involves projecting 3D\nbounding boxes into the three-dimensional environment, holds significance and\ncan be notably enhanced using the AR ecosystem. This study examines an AI\nmodel's ability to deduce 3D bounding boxes in the context of real-time scene\nanalysis while producing and evaluating the model's performance and processing\ntime, in the virtual domain, which is then applied to AVs. This work also\nemploys a synthetic dataset that includes artificially generated images\nmimicking various environmental, lighting, and spatiotemporal states. This\nevaluation is oriented in handling images featuring objects in diverse weather\nconditions, captured with varying camera settings. These variations pose more\nchallenging detection and recognition scenarios, which the outcomes of this\nwork can help achieve competitive results under most of the tested conditions.\n", "date": "2024-12-10", "categories": []} {"id": "2402.03235", "title": "ActiveAnno3D -- An Active Learning Framework for Multi-Modal 3D Object\n Detection", "abstract": " The curation of large-scale datasets is still costly and requires much time\nand resources. Data is often manually labeled, and the challenge of creating\nhigh-quality datasets remains. In this work, we fill the research gap using\nactive learning for multi-modal 3D object detection. We propose ActiveAnno3D,\nan active learning framework to select data samples for labeling that are of\nmaximum informativeness for training. We explore various continuous training\nmethods and integrate the most efficient method regarding computational demand\nand detection performance. Furthermore, we perform extensive experiments and\nablation studies with BEVFusion and PV-RCNN on the nuScenes and TUM Traffic\nIntersection dataset. We show that we can achieve almost the same performance\nwith PV-RCNN and the entropy-based query strategy when using only half of the\ntraining data (77.25 mAP compared to 83.50 mAP) of the TUM Traffic Intersection\ndataset. BEVFusion achieved an mAP of 64.31 when using half of the training\ndata and 75.0 mAP when using the complete nuScenes dataset. We integrate our\nactive learning framework into the proAnno labeling tool to enable AI-assisted\ndata selection and labeling and minimize the labeling costs. Finally, we\nprovide code, weights, and visualization results on our website:\nhttps://active3d-framework.github.io/active3d-framework.\n", "date": "2024-02-05", "categories": []} {"id": "2405.16034", "title": "DiffuBox: Refining 3D Object Detection with Point Diffusion", "abstract": " Ensuring robust 3D object detection and localization is crucial for many\napplications in robotics and autonomous driving. Recent models, however, face\ndifficulties in maintaining high performance when applied to domains with\ndiffering sensor setups or geographic locations, often resulting in poor\nlocalization accuracy due to domain shift. To overcome this challenge, we\nintroduce a novel diffusion-based box refinement approach. This method employs\na domain-agnostic diffusion model, conditioned on the LiDAR points surrounding\na coarse bounding box, to simultaneously refine the box's location, size, and\norientation. We evaluate this approach under various domain adaptation\nsettings, and our results reveal significant improvements across different\ndatasets, object classes and detectors. Our PyTorch implementation is available\nat\n\\href{https://github.com/cxy1997/DiffuBox}{https://github.com/cxy1997/DiffuBox}.\n", "date": "2024-05-25", "categories": []} {"id": "2412.05594", "title": "Real-Time 3D Object Detection Using InnovizOne LiDAR and Low-Power\n Hailo-8 AI Accelerator", "abstract": " Object detection is a significant field in autonomous driving. Popular\nsensors for this task include cameras and LiDAR sensors. LiDAR sensors offer\nseveral advantages, such as insensitivity to light changes, like in a dark\nsetting and the ability to provide 3D information in the form of point clouds,\nwhich include the ranges of objects. However, 3D detection methods, such as\nPointPillars, typically require high-power hardware. Additionally, most common\nspinning LiDARs are sparse and may not achieve the desired quality of object\ndetection in front of the car. In this paper, we present the feasibility of\nperforming real-time 3D object detection of cars using 3D point clouds from a\nLiDAR sensor, processed and deployed on a low-power Hailo-8 AI accelerator. The\nLiDAR sensor used in this study is the InnovizOne sensor, which captures\nobjects in higher quality compared to spinning LiDAR techniques, especially for\ndistant objects. We successfully achieved real-time inference at a rate of\napproximately 5Hz with a high accuracy of 0.91% F1 score, with only -0.2%\ndegradation compared to running the same model on an NVIDIA GeForce RTX 2080\nTi. This work demonstrates that effective real-time 3D object detection can be\nachieved on low-cost, low-power hardware, representing a significant step\ntowards more accessible autonomous driving technologies. The source code and\nthe pre-trained models are available at https://github.com/AIROTAU/\nPointPillarsHailoInnoviz/tree/main\n", "date": "2024-12-07", "categories": []} {"id": "2412.05154", "title": "Towards Flexible 3D Perception: Object-Centric Occupancy Completion\n Augments 3D Object Detection", "abstract": " While 3D object bounding box (bbox) representation has been widely used in\nautonomous driving perception, it lacks the ability to capture the precise\ndetails of an object's intrinsic geometry. Recently, occupancy has emerged as a\npromising alternative for 3D scene perception. However, constructing a\nhigh-resolution occupancy map remains infeasible for large scenes due to\ncomputational constraints. Recognizing that foreground objects only occupy a\nsmall portion of the scene, we introduce object-centric occupancy as a\nsupplement to object bboxes. This representation not only provides intricate\ndetails for detected objects but also enables higher voxel resolution in\npractical applications. We advance the development of object-centric occupancy\nperception from both data and algorithm perspectives. On the data side, we\nconstruct the first object-centric occupancy dataset from scratch using an\nautomated pipeline. From the algorithmic standpoint, we introduce a novel\nobject-centric occupancy completion network equipped with an implicit shape\ndecoder that manages dynamic-size occupancy generation. This network accurately\npredicts the complete object-centric occupancy volume for inaccurate object\nproposals by leveraging temporal information from long sequences. Our method\ndemonstrates robust performance in completing object shapes under noisy\ndetection and tracking conditions. Additionally, we show that our occupancy\nfeatures significantly enhance the detection results of state-of-the-art 3D\nobject detectors, especially for incomplete or distant objects in the Waymo\nOpen Dataset.\n", "date": "2024-12-06", "categories": []} {"id": "2412.04337", "title": "Reflective Teacher: Semi-Supervised Multimodal 3D Object Detection in\n Bird's-Eye-View via Uncertainty Measure", "abstract": " Applying pseudo labeling techniques has been found to be advantageous in\nsemi-supervised 3D object detection (SSOD) in Bird's-Eye-View (BEV) for\nautonomous driving, particularly where labeled data is limited. In the\nliterature, Exponential Moving Average (EMA) has been used for adjustments of\nthe weights of teacher network by the student network. However, the same\ninduces catastrophic forgetting in the teacher network. In this work, we\naddress this issue by introducing a novel concept of Reflective Teacher where\nthe student is trained by both labeled and pseudo labeled data while its\nknowledge is progressively passed to the teacher through a regularizer to\nensure retention of previous knowledge. Additionally, we propose Geometry Aware\nBEV Fusion (GA-BEVFusion) for efficient alignment of multi-modal BEV features,\nthus reducing the disparity between the modalities - camera and LiDAR. This\nhelps to map the precise geometric information embedded among LiDAR points\nreliably with the spatial priors for extraction of semantic information from\ncamera images. Our experiments on the nuScenes and Waymo datasets demonstrate:\n1) improved performance over state-of-the-art methods in both fully supervised\nand semi-supervised settings; 2) Reflective Teacher achieves equivalent\nperformance with only 25% and 22% of labeled data for nuScenes and Waymo\ndatasets respectively, in contrast to other fully supervised methods that\nutilize the full labeled dataset.\n", "date": "2024-12-05", "categories": []} {"id": "2412.04458", "title": "Cubify Anything: Scaling Indoor 3D Object Detection", "abstract": " We consider indoor 3D object detection with respect to a single RGB(-D) frame\nacquired from a commodity handheld device. We seek to significantly advance the\nstatus quo with respect to both data and modeling. First, we establish that\nexisting datasets have significant limitations to scale, accuracy, and\ndiversity of objects. As a result, we introduce the Cubify-Anything 1M (CA-1M)\ndataset, which exhaustively labels over 400K 3D objects on over 1K highly\naccurate laser-scanned scenes with near-perfect registration to over 3.5K\nhandheld, egocentric captures. Next, we establish Cubify Transformer (CuTR), a\nfully Transformer 3D object detection baseline which rather than operating in\n3D on point or voxel-based representations, predicts 3D boxes directly from 2D\nfeatures derived from RGB(-D) inputs. While this approach lacks any 3D\ninductive biases, we show that paired with CA-1M, CuTR outperforms point-based\nmethods - accurately recalling over 62% of objects in 3D, and is significantly\nmore capable at handling noise and uncertainty present in commodity\nLiDAR-derived depth maps while also providing promising RGB only performance\nwithout architecture changes. Furthermore, by pre-training on CA-1M, CuTR can\noutperform point-based methods on a more diverse variant of SUN RGB-D -\nsupporting the notion that while inductive biases in 3D are useful at the\nsmaller sizes of existing datasets, they fail to scale to the data-rich regime\nof CA-1M. Overall, this dataset and baseline model provide strong evidence that\nwe are moving towards models which can effectively Cubify Anything.\n", "date": "2024-12-05", "categories": []} {"id": "2408.16200", "title": "PolarBEVDet: Exploring Polar Representation for Multi-View 3D Object\n Detection in Bird's-Eye-View", "abstract": " Recently, LSS-based multi-view 3D object detection provides an economical and\ndeployment-friendly solution for autonomous driving. However, all the existing\nLSS-based methods transform multi-view image features into a Cartesian\nBird's-Eye-View(BEV) representation, which does not take into account the\nnon-uniform image information distribution and hardly exploits the view\nsymmetry. In this paper, in order to adapt the image information distribution\nand preserve the view symmetry by regular convolution, we propose to employ the\npolar BEV representation to substitute the Cartesian BEV representation. To\nachieve this, we elaborately tailor three modules: a polar view transformer to\ngenerate the polar BEV representation, a polar temporal fusion module for\nfusing historical polar BEV features and a polar detection head to predict the\npolar-parameterized representation of the object. In addition, we design a 2D\nauxiliary detection head and a spatial attention enhancement module to improve\nthe quality of feature extraction in perspective view and BEV, respectively.\nFinally, we integrate the above improvements into a novel multi-view 3D object\ndetector, PolarBEVDet. Experiments on nuScenes show that PolarBEVDet achieves\nthe superior performance. The code is available at\nhttps://github.com/Yzichen/PolarBEVDet.git.(This work has been submitted to the\nIEEE for possible publication. Copyright may be transferred without notice,\nafter which this version may no longer be accessible)\n", "date": "2024-08-29", "categories": []} {"id": "2412.02855", "title": "Optimized CNNs for Rapid 3D Point Cloud Object Recognition", "abstract": " This study introduces a method for efficiently detecting objects within 3D\npoint clouds using convolutional neural networks (CNNs). Our approach adopts a\nunique feature-centric voting mechanism to construct convolutional layers that\ncapitalize on the typical sparsity observed in input data. We explore the\ntrade-off between accuracy and speed across diverse network architectures and\nadvocate for integrating an $\\mathcal{L}_1$ penalty on filter activations to\naugment sparsity within intermediate layers. This research pioneers the\nproposal of sparse convolutional layers combined with $\\mathcal{L}_1$\nregularization to effectively handle large-scale 3D data processing. Our\nmethod's efficacy is demonstrated on the MVTec 3D-AD object detection\nbenchmark. The Vote3Deep models, with just three layers, outperform the\nprevious state-of-the-art in both laser-only approaches and combined\nlaser-vision methods. Additionally, they maintain competitive processing\nspeeds. This underscores our approach's capability to substantially enhance\ndetection performance while ensuring computational efficiency suitable for\nreal-time applications.\n", "date": "2024-12-03", "categories": []} {"id": "2404.03015", "title": "DPFT: Dual Perspective Fusion Transformer for Camera-Radar-based Object\n Detection", "abstract": " The perception of autonomous vehicles has to be efficient, robust, and\ncost-effective. However, cameras are not robust against severe weather\nconditions, lidar sensors are expensive, and the performance of radar-based\nperception is still inferior to the others. Camera-radar fusion methods have\nbeen proposed to address this issue, but these are constrained by the typical\nsparsity of radar point clouds and often designed for radars without elevation\ninformation. We propose a novel camera-radar fusion approach called Dual\nPerspective Fusion Transformer (DPFT), designed to overcome these limitations.\nOur method leverages lower-level radar data (the radar cube) instead of the\nprocessed point clouds to preserve as much information as possible and employs\nprojections in both the camera and ground planes to effectively use radars with\nelevation information and simplify the fusion with camera data. As a result,\nDPFT has demonstrated state-of-the-art performance on the K-Radar dataset while\nshowing remarkable robustness against adverse weather conditions and\nmaintaining a low inference time. The code is made available as open-source\nsoftware under https://github.com/TUMFTM/DPFT.\n", "date": "2024-04-03", "categories": []} {"id": "2411.14869", "title": "BIP3D: Bridging 2D Images and 3D Perception for Embodied Intelligence", "abstract": " In embodied intelligence systems, a key component is 3D perception algorithm,\nwhich enables agents to understand their surrounding environments. Previous\nalgorithms primarily rely on point cloud, which, despite offering precise\ngeometric information, still constrain perception performance due to inherent\nsparsity, noise, and data scarcity. In this work, we introduce a novel\nimage-centric 3D perception model, BIP3D, which leverages expressive image\nfeatures with explicit 3D position encoding to overcome the limitations of\npoint-centric methods. Specifically, we leverage pre-trained 2D vision\nfoundation models to enhance semantic understanding, and introduce a spatial\nenhancer module to improve spatial understanding. Together, these modules\nenable BIP3D to achieve multi-view, multi-modal feature fusion and end-to-end\n3D perception. In our experiments, BIP3D outperforms current state-of-the-art\nresults on the EmbodiedScan benchmark, achieving improvements of 5.69% in the\n3D detection task and 15.25% in the 3D visual grounding task.\n", "date": "2024-11-22", "categories": []} {"id": "2405.15176", "title": "MonoDETRNext: Next-Generation Accurate and Efficient Monocular 3D Object\n Detector", "abstract": " Monocular 3D object detection has vast application potential across various\nfields. DETR-type models have shown remarkable performance in different areas,\nbut there is still considerable room for improvement in monocular 3D detection,\nespecially with the existing DETR-based method, MonoDETR. After addressing the\nquery initialization issues in MonoDETR, we explored several performance\nenhancement strategies, such as incorporating a more efficient encoder and\nutilizing a more powerful depth estimator. Ultimately, we proposed\nMonoDETRNext, a model that comes in two variants based on the choice of depth\nestimator: MonoDETRNext-E, which prioritizes speed, and MonoDETRNext-A, which\nfocuses on accuracy. We posit that MonoDETRNext establishes a new benchmark in\nmonocular 3D object detection and opens avenues for future research. We\nconducted an exhaustive evaluation demonstrating the model's superior\nperformance against existing solutions. Notably, MonoDETRNext-A demonstrated a\n3.52$\\%$ improvement in the $AP_{3D}$ metric on the KITTI test benchmark over\nMonoDETR, while MonoDETRNext-E showed a 2.35$\\%$ increase. Additionally, the\ncomputational efficiency of MonoDETRNext-E slightly exceeds that of its\npredecessor.\n", "date": "2024-05-24", "categories": []} {"id": "2407.17197", "title": "ALPI: Auto-Labeller with Proxy Injection for 3D Object Detection using\n 2D Labels Only", "abstract": " 3D object detection plays a crucial role in various applications such as\nautonomous vehicles, robotics and augmented reality. However, training 3D\ndetectors requires a costly precise annotation, which is a hindrance to scaling\nannotation to large datasets. To address this challenge, we propose a weakly\nsupervised 3D annotator that relies solely on 2D bounding box annotations from\nimages, along with size priors. One major problem is that supervising a 3D\ndetection model using only 2D boxes is not reliable due to ambiguities between\ndifferent 3D poses and their identical 2D projection. We introduce a simple yet\neffective and generic solution: we build 3D proxy objects with annotations by\nconstruction and add them to the training dataset. Our method requires only\nsize priors to adapt to new classes. To better align 2D supervision with 3D\ndetection, our method ensures depth invariance with a novel expression of the\n2D losses. Finally, to detect more challenging instances, our annotator follows\nan offline pseudo-labelling scheme which gradually improves its 3D\npseudo-labels. Extensive experiments on the KITTI dataset demonstrate that our\nmethod not only performs on-par or above previous works on the Car category,\nbut also achieves performance close to fully supervised methods on more\nchallenging classes. We further demonstrate the effectiveness and robustness of\nour method by being the first to experiment on the more challenging nuScenes\ndataset. We additionally propose a setting where weak labels are obtained from\na 2D detector pre-trained on MS-COCO instead of human annotations. The code is\navailable at https://github.com/CEA-LIST/ALPI\n", "date": "2024-07-24", "categories": []} {"id": "2411.15657", "title": "Training an Open-Vocabulary Monocular 3D Object Detection Model without\n 3D Data", "abstract": " Open-vocabulary 3D object detection has recently attracted considerable\nattention due to its broad applications in autonomous driving and robotics,\nwhich aims to effectively recognize novel classes in previously unseen domains.\nHowever, existing point cloud-based open-vocabulary 3D detection models are\nlimited by their high deployment costs. In this work, we propose a novel\nopen-vocabulary monocular 3D object detection framework, dubbed OVM3D-Det,\nwhich trains detectors using only RGB images, making it both cost-effective and\nscalable to publicly available data. Unlike traditional methods, OVM3D-Det does\nnot require high-precision LiDAR or 3D sensor data for either input or\ngenerating 3D bounding boxes. Instead, it employs open-vocabulary 2D models and\npseudo-LiDAR to automatically label 3D objects in RGB images, fostering the\nlearning of open-vocabulary monocular 3D detectors. However, training 3D models\nwith labels directly derived from pseudo-LiDAR is inadequate due to imprecise\nboxes estimated from noisy point clouds and severely occluded objects. To\naddress these issues, we introduce two innovative designs: adaptive\npseudo-LiDAR erosion and bounding box refinement with prior knowledge from\nlarge language models. These techniques effectively calibrate the 3D labels and\nenable RGB-only training for 3D detectors. Extensive experiments demonstrate\nthe superiority of OVM3D-Det over baselines in both indoor and outdoor\nscenarios. The code will be released.\n", "date": "2024-11-23", "categories": []} {"id": "2411.13628", "title": "MambaDETR: Query-based Temporal Modeling using State Space Model for\n Multi-View 3D Object Detection", "abstract": " Utilizing temporal information to improve the performance of 3D detection has\nmade great progress recently in the field of autonomous driving. Traditional\ntransformer-based temporal fusion methods suffer from quadratic computational\ncost and information decay as the length of the frame sequence increases. In\nthis paper, we propose a novel method called MambaDETR, whose main idea is to\nimplement temporal fusion in the efficient state space. Moreover, we design a\nMotion Elimination module to remove the relatively static objects for temporal\nfusion. On the standard nuScenes benchmark, our proposed MambaDETR achieves\nremarkable result in the 3D object detection task, exhibiting state-of-the-art\nperformance among existing temporal fusion methods.\n", "date": "2024-11-20", "categories": []} {"id": "2411.13186", "title": "VADet: Multi-frame LiDAR 3D Object Detection using Variable Aggregation", "abstract": " Input aggregation is a simple technique used by state-of-the-art LiDAR 3D\nobject detectors to improve detection. However, increasing aggregation is known\nto have diminishing returns and even performance degradation, due to objects\nresponding differently to the number of aggregated frames. To address this\nlimitation, we propose an efficient adaptive method, which we call Variable\nAggregation Detection (VADet). Instead of aggregating the entire scene using a\nfixed number of frames, VADet performs aggregation per object, with the number\nof frames determined by an object's observed properties, such as speed and\npoint density. VADet thus reduces the inherent trade-offs of fixed aggregation\nand is not architecture specific. To demonstrate its benefits, we apply VADet\nto three popular single-stage detectors and achieve state-of-the-art\nperformance on the Waymo dataset.\n", "date": "2024-11-20", "categories": []} {"id": "2307.08850", "title": "LiDAR-BEVMTN: Real-Time LiDAR Bird's-Eye View Multi-Task Perception\n Network for Autonomous Driving", "abstract": " LiDAR is crucial for robust 3D scene perception in autonomous driving. LiDAR\nperception has the largest body of literature after camera perception. However,\nmulti-task learning across tasks like detection, segmentation, and motion\nestimation using LiDAR remains relatively unexplored, especially on\nautomotive-grade embedded platforms. We present a real-time multi-task\nconvolutional neural network for LiDAR-based object detection, semantics, and\nmotion segmentation. The unified architecture comprises a shared encoder and\ntask-specific decoders, enabling joint representation learning. We propose a\nnovel Semantic Weighting and Guidance (SWAG) module to transfer semantic\nfeatures for improved object detection selectively. Our heterogeneous training\nscheme combines diverse datasets and exploits complementary cues between tasks.\nThe work provides the first embedded implementation unifying these key\nperception tasks from LiDAR point clouds achieving 3ms latency on the embedded\nNVIDIA Xavier platform. We achieve state-of-the-art results for two tasks,\nsemantic and motion segmentation, and close to state-of-the-art performance for\n3D object detection. By maximizing hardware efficiency and leveraging\nmulti-task synergies, our method delivers an accurate and efficient solution\ntailored for real-world automated driving deployment. Qualitative results can\nbe seen at https://youtu.be/H-hWRzv2lIY.\n", "date": "2023-07-17", "categories": []} {"id": "2407.14923", "title": "RayFormer: Improving Query-Based Multi-Camera 3D Object Detection via\n Ray-Centric Strategies", "abstract": " The recent advances in query-based multi-camera 3D object detection are\nfeatured by initializing object queries in the 3D space, and then sampling\nfeatures from perspective-view images to perform multi-round query refinement.\nIn such a framework, query points near the same camera ray are likely to sample\nsimilar features from very close pixels, resulting in ambiguous query features\nand degraded detection accuracy. To this end, we introduce RayFormer, a\ncamera-ray-inspired query-based 3D object detector that aligns the\ninitialization and feature extraction of object queries with the optical\ncharacteristics of cameras. Specifically, RayFormer transforms perspective-view\nimage features into bird's eye view (BEV) via the lift-splat-shoot method and\nsegments the BEV map to sectors based on the camera rays. Object queries are\nuniformly and sparsely initialized along each camera ray, facilitating the\nprojection of different queries onto different areas in the image to extract\ndistinct features. Besides, we leverage the instance information of images to\nsupplement the uniformly initialized object queries by further involving\nadditional queries along the ray from 2D object detection boxes. To extract\nunique object-level features that cater to distinct queries, we design a ray\nsampling method that suitably organizes the distribution of feature sampling\npoints on both images and bird's eye view. Extensive experiments are conducted\non the nuScenes dataset to validate our proposed ray-inspired model design. The\nproposed RayFormer achieves superior performance of 55.5% mAP and 63.3% NDS,\nrespectively.\n", "date": "2024-07-20", "categories": []} {"id": "2411.06173", "title": "LSSInst: Improving Geometric Modeling in LSS-Based BEV Perception with\n Instance Representation", "abstract": " With the attention gained by camera-only 3D object detection in autonomous\ndriving, methods based on Bird-Eye-View (BEV) representation especially derived\nfrom the forward view transformation paradigm, i.e., lift-splat-shoot (LSS),\nhave recently seen significant progress. The BEV representation formulated by\nthe frustum based on depth distribution prediction is ideal for learning the\nroad structure and scene layout from multi-view images. However, to retain\ncomputational efficiency, the compressed BEV representation such as in\nresolution and axis is inevitably weak in retaining the individual geometric\ndetails, undermining the methodological generality and applicability. With this\nin mind, to compensate for the missing details and utilize multi-view geometry\nconstraints, we propose LSSInst, a two-stage object detector incorporating BEV\nand instance representations in tandem. The proposed detector exploits\nfine-grained pixel-level features that can be flexibly integrated into existing\nLSS-based BEV networks. Having said that, due to the inherent gap between two\nrepresentation spaces, we design the instance adaptor for the BEV-to-instance\nsemantic coherence rather than pass the proposal naively. Extensive experiments\ndemonstrated that our proposed framework is of excellent generalization ability\nand performance, which boosts the performances of modern LSS-based BEV\nperception methods without bells and whistles and outperforms current LSS-based\nstate-of-the-art works on the large-scale nuScenes benchmark.\n", "date": "2024-11-09", "categories": []} {"id": "2410.24001", "title": "ImOV3D: Learning Open-Vocabulary Point Clouds 3D Object Detection from\n Only 2D Images", "abstract": " Open-vocabulary 3D object detection (OV-3Det) aims to generalize beyond the\nlimited number of base categories labeled during the training phase. The\nbiggest bottleneck is the scarcity of annotated 3D data, whereas 2D image\ndatasets are abundant and richly annotated. Consequently, it is intuitive to\nleverage the wealth of annotations in 2D images to alleviate the inherent data\nscarcity in OV-3Det. In this paper, we push the task setup to its limits by\nexploring the potential of using solely 2D images to learn OV-3Det. The major\nchallenges for this setup is the modality gap between training images and\ntesting point clouds, which prevents effective integration of 2D knowledge into\nOV-3Det. To address this challenge, we propose a novel framework ImOV3D to\nleverage pseudo multimodal representation containing both images and point\nclouds (PC) to close the modality gap. The key of ImOV3D lies in flexible\nmodality conversion where 2D images can be lifted into 3D using monocular depth\nestimation and can also be derived from 3D scenes through rendering. This\nallows unifying both training images and testing point clouds into a common\nimage-PC representation, encompassing a wealth of 2D semantic information and\nalso incorporating the depth and structural characteristics of 3D spatial data.\nWe carefully conduct such conversion to minimize the domain gap between\ntraining and test cases. Extensive experiments on two benchmark datasets,\nSUNRGBD and ScanNet, show that ImOV3D significantly outperforms existing\nmethods, even in the absence of ground truth 3D training data. With the\ninclusion of a minimal amount of real 3D data for fine-tuning, the performance\nalso significantly surpasses previous state-of-the-art. Codes and pre-trained\nmodels are released on the https://github.com/yangtiming/ImOV3D.\n", "date": "2024-10-31", "categories": []} {"id": "2408.07999", "title": "Co-Fix3D: Enhancing 3D Object Detection with Collaborative Refinement", "abstract": " 3D object detection in driving scenarios faces the challenge of complex road\nenvironments, which can lead to the loss or incompleteness of key features,\nthereby affecting perception performance. To address this issue, we propose an\nadvanced detection framework called Co-Fix3D. Co-Fix3D integrates Local and\nGlobal Enhancement (LGE) modules to refine Bird's Eye View (BEV) features. The\nLGE module uses Discrete Wavelet Transform (DWT) for pixel-level local\noptimization and incorporates an attention mechanism for global optimization.\nTo handle varying detection difficulties, we adopt multi-head LGE modules,\nenabling each module to focus on targets with different levels of detection\ncomplexity, thus further enhancing overall perception capability. Experimental\nresults show that on the nuScenes dataset's LiDAR benchmark, Co-Fix3D achieves\n69.4\\% mAP and 73.5\\% NDS, while on the multimodal benchmark, it achieves\n72.3\\% mAP and 74.7\\% NDS. The source code is publicly available at\n\\href{https://github.com/rubbish001/Co-Fix3d}{https://github.com/rubbish001/Co-Fix3d}.\n", "date": "2024-08-15", "categories": []} {"id": "2408.08092", "title": "SC3D: Label-Efficient Outdoor 3D Object Detection via Single Click\n Annotation", "abstract": " LiDAR-based outdoor 3D object detection has received widespread attention.\nHowever, training 3D detectors from the LiDAR point cloud typically relies on\nexpensive bounding box annotations. This paper presents SC3D, an innovative\nlabel-efficient method requiring only a single coarse click on the bird's eye\nview of the 3D point cloud for each frame. A key challenge here is the absence\nof complete geometric descriptions of the target objects from such simple click\nannotations. To address this issue, our proposed SC3D adopts a progressive\npipeline. Initially, we design a mixed pseudo-label generation module that\nexpands limited click annotations into a mixture of bounding box and semantic\nmask supervision. Next, we propose a mix-supervised teacher model, enabling the\ndetector to learn mixed supervision information. Finally, we introduce a\nmixed-supervised student network that leverages the teacher model's\ngeneralization ability to learn unclicked instances.Experimental results on the\nwidely used nuScenes and KITTI datasets demonstrate that our SC3D with only\ncoarse clicks, which requires only 0.2% annotation cost, achieves\nstate-of-the-art performance compared to weakly-supervised 3D detection\nmethods.The code will be made publicly available.\n", "date": "2024-08-15", "categories": []} {"id": "2405.09828", "title": "*: Improving the 3D detector by introducing Voxel2Pillar feature\n encoding and extracting multi-scale features", "abstract": " The multi-line LiDAR is widely used in autonomous vehicles, so point\ncloud-based 3D detectors are essential for autonomous driving. Extracting rich\nmulti-scale features is crucial for point cloud-based 3D detectors in\nautonomous driving due to significant differences in the size of different\ntypes of objects. However, because of the real-time requirements, large-size\nconvolution kernels are rarely used to extract large-scale features in the\nbackbone. Current 3D detectors commonly use feature pyramid networks to obtain\nlarge-scale features; however, some objects containing fewer point clouds are\nfurther lost during down-sampling, resulting in degraded performance. Since\npillar-based schemes require much less computation than voxel-based schemes,\nthey are more suitable for constructing real-time 3D detectors. Hence, we\npropose the *, a pillar-based scheme. We redesigned the feature encoding, the\nbackbone, and the neck of the 3D detector. We propose the Voxel2Pillar feature\nencoding, which uses a sparse convolution constructor to construct pillars with\nricher point cloud features, especially height features. The Voxel2Pillar adds\nmore learnable parameters to the feature encoding, enabling the initial pillars\nto have higher performance ability. We extract multi-scale and large-scale\nfeatures in the proposed fully sparse backbone, which does not utilize\nlarge-size convolutional kernels; the backbone consists of the proposed\nmulti-scale feature extraction module. The neck consists of the proposed sparse\nConvNeXt, whose simple structure significantly improves the performance. We\nvalidate the effectiveness of the proposed * on the Waymo Open Dataset, and the\nobject detection accuracy for vehicles, pedestrians, and cyclists is improved.\nWe also verify the effectiveness of each proposed module in detail through\nablation studies.\n", "date": "2024-05-16", "categories": []} {"id": "2406.17654", "title": "MDHA: Multi-Scale Deformable Transformer with Hybrid Anchors for\n Multi-View 3D Object Detection", "abstract": " Multi-view 3D object detection is a crucial component of autonomous driving\nsystems. Contemporary query-based methods primarily depend either on\ndataset-specific initialization of 3D anchors, introducing bias, or utilize\ndense attention mechanisms, which are computationally inefficient and\nunscalable. To overcome these issues, we present MDHA, a novel sparse\nquery-based framework, which constructs adaptive 3D output proposals using\nhybrid anchors from multi-view, multi-scale image input. Fixed 2D anchors are\ncombined with depth predictions to form 2.5D anchors, which are projected to\nobtain 3D proposals. To ensure high efficiency, our proposed Anchor Encoder\nperforms sparse refinement and selects the top-$k$ anchors and features.\nMoreover, while existing multi-view attention mechanisms rely on projecting\nreference points to multiple images, our novel Circular Deformable Attention\nmechanism only projects to a single image but allows reference points to\nseamlessly attend to adjacent images, improving efficiency without compromising\non performance. On the nuScenes val set, it achieves 46.4\\% mAP and 55.0\\% NDS\nwith a ResNet101 backbone. MDHA significantly outperforms the baseline where\nanchor proposals are modelled as learnable embeddings. Code is available at\nhttps://github.com/NaomiEX/MDHA.\n", "date": "2024-06-25", "categories": []} {"id": "2411.05292", "title": "SimpleBEV: Improved LiDAR-Camera Fusion Architecture for 3D Object\n Detection", "abstract": " More and more research works fuse the LiDAR and camera information to improve\nthe 3D object detection of the autonomous driving system. Recently, a simple\nyet effective fusion framework has achieved an excellent detection performance,\nfusing the LiDAR and camera features in a unified bird's-eye-view (BEV) space.\nIn this paper, we propose a LiDAR-camera fusion framework, named SimpleBEV, for\naccurate 3D object detection, which follows the BEV-based fusion framework and\nimproves the camera and LiDAR encoders, respectively. Specifically, we perform\nthe camera-based depth estimation using a cascade network and rectify the depth\nresults with the depth information derived from the LiDAR points. Meanwhile, an\nauxiliary branch that implements the 3D object detection using only the\ncamera-BEV features is introduced to exploit the camera information during the\ntraining phase. Besides, we improve the LiDAR feature extractor by fusing the\nmulti-scaled sparse convolutional features. Experimental results demonstrate\nthe effectiveness of our proposed method. Our method achieves 77.6\\% NDS\naccuracy on the nuScenes dataset, showcasing superior performance in the 3D\nobject detection track.\n", "date": "2024-11-08", "categories": []} {"id": "2403.17387", "title": "Decoupled Pseudo-labeling for Semi-Supervised Monocular 3D Object\n Detection", "abstract": " We delve into pseudo-labeling for semi-supervised monocular 3D object\ndetection (SSM3OD) and discover two primary issues: a misalignment between the\nprediction quality of 3D and 2D attributes and the tendency of depth\nsupervision derived from pseudo-labels to be noisy, leading to significant\noptimization conflicts with other reliable forms of supervision. We introduce a\nnovel decoupled pseudo-labeling (DPL) approach for SSM3OD. Our approach\nfeatures a Decoupled Pseudo-label Generation (DPG) module, designed to\nefficiently generate pseudo-labels by separately processing 2D and 3D\nattributes. This module incorporates a unique homography-based method for\nidentifying dependable pseudo-labels in BEV space, specifically for 3D\nattributes. Additionally, we present a DepthGradient Projection (DGP) module to\nmitigate optimization conflicts caused by noisy depth supervision of\npseudo-labels, effectively decoupling the depth gradient and removing\nconflicting gradients. This dual decoupling strategy-at both the pseudo-label\ngeneration and gradient levels-significantly improves the utilization of\npseudo-labels in SSM3OD. Our comprehensive experiments on the KITTI benchmark\ndemonstrate the superiority of our method over existing approaches.\n", "date": "2024-03-26", "categories": []} {"id": "2411.01584", "title": "One for All: Multi-Domain Joint Training for Point Cloud Based 3D Object\n Detection", "abstract": " The current trend in computer vision is to utilize one universal model to\naddress all various tasks. Achieving such a universal model inevitably requires\nincorporating multi-domain data for joint training to learn across multiple\nproblem scenarios. In point cloud based 3D object detection, however, such\nmulti-domain joint training is highly challenging, because large domain gaps\namong point clouds from different datasets lead to the severe\ndomain-interference problem. In this paper, we propose \\textbf{OneDet3D}, a\nuniversal one-for-all model that addresses 3D detection across different\ndomains, including diverse indoor and outdoor scenes, within the \\emph{same}\nframework and only \\emph{one} set of parameters. We propose the domain-aware\npartitioning in scatter and context, guided by a routing mechanism, to address\nthe data interference issue, and further incorporate the text modality for a\nlanguage-guided classification to unify the multi-dataset label spaces and\nmitigate the category interference issue. The fully sparse structure and\nanchor-free head further accommodate point clouds with significant scale\ndisparities. Extensive experiments demonstrate the strong universal ability of\nOneDet3D to utilize only one trained model for addressing almost all 3D object\ndetection tasks.\n", "date": "2024-11-03", "categories": []} {"id": "2410.03918", "title": "STONE: A Submodular Optimization Framework for Active 3D Object\n Detection", "abstract": " 3D object detection is fundamentally important for various emerging\napplications, including autonomous driving and robotics. A key requirement for\ntraining an accurate 3D object detector is the availability of a large amount\nof LiDAR-based point cloud data. Unfortunately, labeling point cloud data is\nextremely challenging, as accurate 3D bounding boxes and semantic labels are\nrequired for each potential object. This paper proposes a unified active 3D\nobject detection framework, for greatly reducing the labeling cost of training\n3D object detectors. Our framework is based on a novel formulation of\nsubmodular optimization, specifically tailored to the problem of active 3D\nobject detection. In particular, we address two fundamental challenges\nassociated with active 3D object detection: data imbalance and the need to\ncover the distribution of the data, including LiDAR-based point cloud data of\nvarying difficulty levels. Extensive experiments demonstrate that our method\nachieves state-of-the-art performance with high computational efficiency\ncompared to existing active learning methods. The code is available at\nhttps://github.com/RuiyuM/STONE.\n", "date": "2024-10-04", "categories": []} {"id": "2411.00340", "title": "GAFusion: Adaptive Fusing LiDAR and Camera with Multiple Guidance for 3D\n Object Detection", "abstract": " Recent years have witnessed the remarkable progress of 3D multi-modality\nobject detection methods based on the Bird's-Eye-View (BEV) perspective.\nHowever, most of them overlook the complementary interaction and guidance\nbetween LiDAR and camera. In this work, we propose a novel multi-modality 3D\nobjection detection method, named GAFusion, with LiDAR-guided global\ninteraction and adaptive fusion. Specifically, we introduce sparse depth\nguidance (SDG) and LiDAR occupancy guidance (LOG) to generate 3D features with\nsufficient depth information. In the following, LiDAR-guided adaptive fusion\ntransformer (LGAFT) is developed to adaptively enhance the interaction of\ndifferent modal BEV features from a global perspective. Meanwhile, additional\ndownsampling with sparse height compression and multi-scale dual-path\ntransformer (MSDPT) are designed to enlarge the receptive fields of different\nmodal features. Finally, a temporal fusion module is introduced to aggregate\nfeatures from previous frames. GAFusion achieves state-of-the-art 3D object\ndetection results with 73.6$\\%$ mAP and 74.9$\\%$ NDS on the nuScenes test set.\n", "date": "2024-11-01", "categories": []} {"id": "2410.23910", "title": "Uncertainty Estimation for 3D Object Detection via Evidential Learning", "abstract": " 3D object detection is an essential task for computer vision applications in\nautonomous vehicles and robotics. However, models often struggle to quantify\ndetection reliability, leading to poor performance on unfamiliar scenes. We\nintroduce a framework for quantifying uncertainty in 3D object detection by\nleveraging an evidential learning loss on Bird's Eye View representations in\nthe 3D detector. These uncertainty estimates require minimal computational\noverhead and are generalizable across different architectures. We demonstrate\nboth the efficacy and importance of these uncertainty estimates on identifying\nout-of-distribution scenes, poorly localized objects, and missing (false\nnegative) detections; our framework consistently improves over baselines by\n10-20% on average. Finally, we integrate this suite of tasks into a system\nwhere a 3D object detector auto-labels driving scenes and our uncertainty\nestimates verify label correctness before the labels are used to train a second\nmodel. Here, our uncertainty-driven verification results in a 1% improvement in\nmAP and a 1-2% improvement in NDS.\n", "date": "2024-10-31", "categories": []} {"id": "2410.22461", "title": "Unified Domain Generalization and Adaptation for Multi-View 3D Object\n Detection", "abstract": " Recent advances in 3D object detection leveraging multi-view cameras have\ndemonstrated their practical and economical value in various challenging vision\ntasks. However, typical supervised learning approaches face challenges in\nachieving satisfactory adaptation toward unseen and unlabeled target datasets\n(\\ie, direct transfer) due to the inevitable geometric misalignment between the\nsource and target domains. In practice, we also encounter constraints on\nresources for training models and collecting annotations for the successful\ndeployment of 3D object detectors. In this paper, we propose Unified Domain\nGeneralization and Adaptation (UDGA), a practical solution to mitigate those\ndrawbacks. We first propose Multi-view Overlap Depth Constraint that leverages\nthe strong association between multi-view, significantly alleviating geometric\ngaps due to perspective view changes. Then, we present a Label-Efficient Domain\nAdaptation approach to handle unfamiliar targets with significantly fewer\namounts of labels (\\ie, 1$\\%$ and 5$\\%)$, while preserving well-defined source\nknowledge for training efficiency. Overall, UDGA framework enables stable\ndetection performance in both source and target domains, effectively bridging\ninevitable domain gaps, while demanding fewer annotations. We demonstrate the\nrobustness of UDGA with large-scale benchmarks: nuScenes, Lyft, and Waymo,\nwhere our framework outperforms the current state-of-the-art methods.\n", "date": "2024-10-29", "categories": []} {"id": "2410.21566", "title": "MVSDet: Multi-View Indoor 3D Object Detection via Efficient Plane Sweeps", "abstract": " The key challenge of multi-view indoor 3D object detection is to infer\naccurate geometry information from images for precise 3D detection. Previous\nmethod relies on NeRF for geometry reasoning. However, the geometry extracted\nfrom NeRF is generally inaccurate, which leads to sub-optimal detection\nperformance. In this paper, we propose MVSDet which utilizes plane sweep for\ngeometry-aware 3D object detection. To circumvent the requirement for a large\nnumber of depth planes for accurate depth prediction, we design a probabilistic\nsampling and soft weighting mechanism to decide the placement of pixel features\non the 3D volume. We select multiple locations that score top in the\nprobability volume for each pixel and use their probability score to indicate\nthe confidence. We further apply recent pixel-aligned Gaussian Splatting to\nregularize depth prediction and improve detection performance with little\ncomputation overhead. Extensive experiments on ScanNet and ARKitScenes datasets\nare conducted to show the superiority of our model. Our code is available at\nhttps://github.com/Pixie8888/MVSDet.\n", "date": "2024-10-28", "categories": []} {"id": "2410.19872", "title": "Radar and Camera Fusion for Object Detection and Tracking: A\n Comprehensive Survey", "abstract": " Multi-modal fusion is imperative to the implementation of reliable object\ndetection and tracking in complex environments. Exploiting the synergy of\nheterogeneous modal information endows perception systems the ability to\nachieve more comprehensive, robust, and accurate performance. As a nucleus\nconcern in wireless-vision collaboration, radar-camera fusion has prompted\nprospective research directions owing to its extensive applicability,\ncomplementarity, and compatibility. Nonetheless, there still lacks a systematic\nsurvey specifically focusing on deep fusion of radar and camera for object\ndetection and tracking. To fill this void, we embark on an endeavor to\ncomprehensively review radar-camera fusion in a holistic way. First, we\nelaborate on the fundamental principles, methodologies, and applications of\nradar-camera fusion perception. Next, we delve into the key techniques\nconcerning sensor calibration, modal representation, data alignment, and fusion\noperation. Furthermore, we provide a detailed taxonomy covering the research\ntopics related to object detection and tracking in the context of radar and\ncamera technologies.Finally, we discuss the emerging perspectives in the field\nof radar-camera fusion perception and highlight the potential areas for future\nresearch.\n", "date": "2024-10-24", "categories": []} {"id": "2310.10353", "title": "Multimodal Object Query Initialization for 3D Object Detection", "abstract": " 3D object detection models that exploit both LiDAR and camera sensor features\nare top performers in large-scale autonomous driving benchmarks. A transformer\nis a popular network architecture used for this task, in which so-called object\nqueries act as candidate objects. Initializing these object queries based on\ncurrent sensor inputs is a common practice. For this, existing methods strongly\nrely on LiDAR data however, and do not fully exploit image features. Besides,\nthey introduce significant latency. To overcome these limitations we propose\nEfficientQ3M, an efficient, modular, and multimodal solution for object query\ninitialization for transformer-based 3D object detection models. The proposed\ninitialization method is combined with a \"modality-balanced\" transformer\ndecoder where the queries can access all sensor modalities throughout the\ndecoder. In experiments, we outperform the state of the art in\ntransformer-based LiDAR object detection on the competitive nuScenes benchmark\nand showcase the benefits of input-dependent multimodal query initialization,\nwhile being more efficient than the available alternatives for LiDAR-camera\ninitialization. The proposed method can be applied with any combination of\nsensor modalities as input, demonstrating its modularity.\n", "date": "2023-10-16", "categories": []} {"id": "2403.14626", "title": "ODTFormer: Efficient Obstacle Detection and Tracking with Stereo Cameras\n Based on Transformer", "abstract": " Obstacle detection and tracking represent a critical component in robot\nautonomous navigation. In this paper, we propose ODTFormer, a Transformer-based\nmodel to address both obstacle detection and tracking problems. For the\ndetection task, our approach leverages deformable attention to construct a 3D\ncost volume, which is decoded progressively in the form of voxel occupancy\ngrids. We further track the obstacles by matching the voxels between\nconsecutive frames. The entire model can be optimized in an end-to-end manner.\nThrough extensive experiments on DrivingStereo and KITTI benchmarks, our model\nachieves state-of-the-art performance in the obstacle detection task. We also\nreport comparable accuracy to state-of-the-art obstacle tracking models while\nrequiring only a fraction of their computation cost, typically ten-fold to\ntwenty-fold less. The code and model weights will be publicly released.\n", "date": "2024-03-21", "categories": []} {"id": "2403.17633", "title": "UADA3D: Unsupervised Adversarial Domain Adaptation for 3D Object\n Detection with Sparse LiDAR and Large Domain Gaps", "abstract": " In this study, we address a gap in existing unsupervised domain adaptation\napproaches on LiDAR-based 3D object detection, which have predominantly\nconcentrated on adapting between established, high-density autonomous driving\ndatasets. We focus on sparser point clouds, capturing scenarios from different\nperspectives: not just from vehicles on the road but also from mobile robots on\nsidewalks, which encounter significantly different environmental conditions and\nsensor configurations. We introduce Unsupervised Adversarial Domain Adaptation\nfor 3D Object Detection (UADA3D). UADA3D does not depend on pre-trained source\nmodels or teacher-student architectures. Instead, it uses an adversarial\napproach to directly learn domain-invariant features. We demonstrate its\nefficacy in various adaptation scenarios, showing significant improvements in\nboth self-driving car and mobile robot domains. Our code is open-source and\nwill be available soon.\n", "date": "2024-03-26", "categories": []} {"id": "2408.11958", "title": "CARLA Drone: Monocular 3D Object Detection from a Different Perspective", "abstract": " Existing techniques for monocular 3D detection have a serious restriction.\nThey tend to perform well only on a limited set of benchmarks, faring well\neither on ego-centric car views or on traffic camera views, but rarely on both.\nTo encourage progress, this work advocates for an extended evaluation of 3D\ndetection frameworks across different camera perspectives. We make two key\ncontributions. First, we introduce the CARLA Drone dataset, CDrone. Simulating\ndrone views, it substantially expands the diversity of camera perspectives in\nexisting benchmarks. Despite its synthetic nature, CDrone represents a\nreal-world challenge. To show this, we confirm that previous techniques\nstruggle to perform well both on CDrone and a real-world 3D drone dataset.\nSecond, we develop an effective data augmentation pipeline called GroundMix.\nIts distinguishing element is the use of the ground for creating 3D-consistent\naugmentation of a training image. GroundMix significantly boosts the detection\naccuracy of a lightweight one-stage detector. In our expanded evaluation, we\nachieve the average precision on par with or substantially higher than the\nprevious state of the art across all tested datasets.\n", "date": "2024-08-21", "categories": []} {"id": "2410.07758", "title": "HeightFormer: A Semantic Alignment Monocular 3D Object Detection Method\n from Roadside Perspective", "abstract": " The on-board 3D object detection technology has received extensive attention\nas a critical technology for autonomous driving, while few studies have focused\non applying roadside sensors in 3D traffic object detection. Existing studies\nachieve the projection of 2D image features to 3D features through height\nestimation based on the frustum. However, they did not consider the height\nalignment and the extraction efficiency of bird's-eye-view features. We propose\na novel 3D object detection framework integrating Spatial Former and Voxel\nPooling Former to enhance 2D-to-3D projection based on height estimation.\nExtensive experiments were conducted using the Rope3D and DAIR-V2X-I dataset,\nand the results demonstrated the outperformance of the proposed algorithm in\nthe detection of both vehicles and cyclists. These results indicate that the\nalgorithm is robust and generalized under various detection scenarios.\nImproving the accuracy of 3D object detection on the roadside is conducive to\nbuilding a safe and trustworthy intelligent transportation system of\nvehicle-road coordination and promoting the large-scale application of\nautonomous driving. The code and pre-trained models will be released on\nhttps://anonymous.4open.science/r/HeightFormer.\n", "date": "2024-10-10", "categories": []} {"id": "2410.14799", "title": "Deep Generic Dynamic Object Detection Based on Dynamic Grid Maps", "abstract": " This paper describes a method to detect generic dynamic objects for automated\ndriving. First, a LiDAR-based dynamic grid is generated online. Second, a deep\nlearning-based detector is trained on the dynamic grid to infer the presence of\ndynamic objects of any type, which is a prerequisite for safe automated\nvehicles in arbitrary, edge-case scenarios. The Rotation-equivariant Detector\n(ReDet) - originally designed for oriented object detection on aerial images -\nwas chosen due to its high detection performance. Experiments are conducted\nbased on real sensor data and the benefits in comparison to classic dynamic\ncell clustering strategies are highlighted. The false positive object detection\nrate is strongly reduced by the proposed approach.\n", "date": "2024-10-18", "categories": []} {"id": "2410.15814", "title": "Kaninfradet3D:A Road-side Camera-LiDAR Fusion 3D Perception Model based\n on Nonlinear Feature Extraction and Intrinsic Correlation", "abstract": " With the development of AI-assisted driving, numerous methods have emerged\nfor ego-vehicle 3D perception tasks, but there has been limited research on\nroadside perception. With its ability to provide a global view and a broader\nsensing range, the roadside perspective is worth developing. LiDAR provides\nprecise three-dimensional spatial information, while cameras offer semantic\ninformation. These two modalities are complementary in 3D detection. However,\nadding camera data does not increase accuracy in some studies since the\ninformation extraction and fusion procedure is not sufficiently reliable.\nRecently, Kolmogorov-Arnold Networks (KANs) have been proposed as replacements\nfor MLPs, which are better suited for high-dimensional, complex data. Both the\ncamera and the LiDAR provide high-dimensional information, and employing KANs\nshould enhance the extraction of valuable features to produce better fusion\noutcomes. This paper proposes Kaninfradet3D, which optimizes the feature\nextraction and fusion modules. To extract features from complex\nhigh-dimensional data, the model's encoder and fuser modules were improved\nusing KAN Layers. Cross-attention was applied to enhance feature fusion, and\nvisual comparisons verified that camera features were more evenly integrated.\nThis addressed the issue of camera features being abnormally concentrated,\nnegatively impacting fusion. Compared to the benchmark, our approach shows\nimprovements of +9.87 mAP and +10.64 mAP in the two viewpoints of the TUMTraf\nIntersection Dataset and an improvement of +1.40 mAP in the roadside end of the\nTUMTraf V2X Cooperative Perception Dataset. The results indicate that\nKaninfradet3D can effectively fuse features, demonstrating the potential of\napplying KANs in roadside perception tasks.\n", "date": "2024-10-21", "categories": []} {"id": "2410.11125", "title": "UAV3D: A Large-scale 3D Perception Benchmark for Unmanned Aerial\n Vehicles", "abstract": " Unmanned Aerial Vehicles (UAVs), equipped with cameras, are employed in\nnumerous applications, including aerial photography, surveillance, and\nagriculture. In these applications, robust object detection and tracking are\nessential for the effective deployment of UAVs. However, existing benchmarks\nfor UAV applications are mainly designed for traditional 2D perception tasks,\nrestricting the development of real-world applications that require a 3D\nunderstanding of the environment. Furthermore, despite recent advancements in\nsingle-UAV perception, limited views of a single UAV platform significantly\nconstrain its perception capabilities over long distances or in occluded areas.\nTo address these challenges, we introduce UAV3D, a benchmark designed to\nadvance research in both 3D and collaborative 3D perception tasks with UAVs.\nUAV3D comprises 1,000 scenes, each of which has 20 frames with fully annotated\n3D bounding boxes on vehicles. We provide the benchmark for four 3D perception\ntasks: single-UAV 3D object detection, single-UAV object tracking,\ncollaborative-UAV 3D object detection, and collaborative-UAV object tracking.\nOur dataset and code are available at\nhttps://huiyegit.github.io/UAV3D_Benchmark/.\n", "date": "2024-10-14", "categories": []} {"id": "2410.11211", "title": "CVCP-Fusion: On Implicit Depth Estimation for 3D Bounding Box Prediction", "abstract": " Combining LiDAR and Camera-view data has become a common approach for 3D\nObject Detection. However, previous approaches combine the two input streams at\na point-level, throwing away semantic information derived from camera features.\nIn this paper we propose Cross-View Center Point-Fusion, a state-of-the-art\nmodel to perform 3D object detection by combining camera and LiDAR-derived\nfeatures in the BEV space to preserve semantic density from the camera stream\nwhile incorporating spacial data from the LiDAR stream. Our architecture\nutilizes aspects from previously established algorithms, Cross-View\nTransformers and CenterPoint, and runs their backbones in parallel, allowing\nefficient computation for real-time processing and application. In this paper\nwe find that while an implicitly calculated depth-estimate may be sufficiently\naccurate in a 2D map-view representation, explicitly calculated geometric and\nspacial information is needed for precise bounding box prediction in the 3D\nworld-view space.\n", "date": "2024-10-15", "categories": []} {"id": "2410.12394", "title": "Real-time Stereo-based 3D Object Detection for Streaming Perception", "abstract": " The ability to promptly respond to environmental changes is crucial for the\nperception system of autonomous driving. Recently, a new task called streaming\nperception was proposed. It jointly evaluate the latency and accuracy into a\nsingle metric for video online perception. In this work, we introduce\nStreamDSGN, the first real-time stereo-based 3D object detection framework\ndesigned for streaming perception. StreamDSGN is an end-to-end framework that\ndirectly predicts the 3D properties of objects in the next moment by leveraging\nhistorical information, thereby alleviating the accuracy degradation of\nstreaming perception. Further, StreamDSGN applies three strategies to enhance\nthe perception accuracy: (1) A feature-flow-based fusion method, which\ngenerates a pseudo-next feature at the current moment to address the\nmisalignment issue between feature and ground truth. (2) An extra regression\nloss for explicit supervision of object motion consistency in consecutive\nframes. (3) A large kernel backbone with a large receptive field for\neffectively capturing long-range spatial contextual features caused by changes\nin object positions. Experiments on the KITTI Tracking dataset show that,\ncompared with the strong baseline, StreamDSGN significantly improves the\nstreaming average precision by up to 4.33%. Our code is available at\nhttps://github.com/weiyangdaren/streamDSGN-pytorch.\n", "date": "2024-10-16", "categories": []} {"id": "2410.12592", "title": "Cocoon: Robust Multi-Modal Perception with Uncertainty-Aware Sensor\n Fusion", "abstract": " An important paradigm in 3D object detection is the use of multiple\nmodalities to enhance accuracy in both normal and challenging conditions,\nparticularly for long-tail scenarios. To address this, recent studies have\nexplored two directions of adaptive approaches: MoE-based adaptive fusion,\nwhich struggles with uncertainties arising from distinct object configurations,\nand late fusion for output-level adaptive fusion, which relies on separate\ndetection pipelines and limits comprehensive understanding. In this work, we\nintroduce Cocoon, an object- and feature-level uncertainty-aware fusion\nframework. The key innovation lies in uncertainty quantification for\nheterogeneous representations, enabling fair comparison across modalities\nthrough the introduction of a feature aligner and a learnable surrogate ground\ntruth, termed feature impression. We also define a training objective to ensure\nthat their relationship provides a valid metric for uncertainty quantification.\nCocoon consistently outperforms existing static and adaptive methods in both\nnormal and challenging conditions, including those with natural and artificial\ncorruptions. Furthermore, we show the validity and efficacy of our uncertainty\nmetric across diverse datasets.\n", "date": "2024-10-16", "categories": []} {"id": "2405.07696", "title": "MonoMAE: Enhancing Monocular 3D Detection through Depth-Aware Masked\n Autoencoders", "abstract": " Monocular 3D object detection aims for precise 3D localization and\nidentification of objects from a single-view image. Despite its recent\nprogress, it often struggles while handling pervasive object occlusions that\ntend to complicate and degrade the prediction of object dimensions, depths, and\norientations. We design MonoMAE, a monocular 3D detector inspired by Masked\nAutoencoders that addresses the object occlusion issue by masking and\nreconstructing objects in the feature space. MonoMAE consists of two novel\ndesigns. The first is depth-aware masking that selectively masks certain parts\nof non-occluded object queries in the feature space for simulating occluded\nobject queries for network training. It masks non-occluded object queries by\nbalancing the masked and preserved query portions adaptively according to the\ndepth information. The second is lightweight query completion that works with\nthe depth-aware masking to learn to reconstruct and complete the masked object\nqueries. With the proposed object occlusion and completion, MonoMAE learns\nenriched 3D representations that achieve superior monocular 3D detection\nperformance qualitatively and quantitatively for both occluded and non-occluded\nobjects. Additionally, MonoMAE learns generalizable representations that can\nwork well in new domains.\n", "date": "2024-05-13", "categories": []} {"id": "2406.10115", "title": "Shelf-Supervised Cross-Modal Pre-Training for 3D Object Detection", "abstract": " State-of-the-art 3D object detectors are often trained on massive labeled\ndatasets. However, annotating 3D bounding boxes remains prohibitively expensive\nand time-consuming, particularly for LiDAR. Instead, recent works demonstrate\nthat self-supervised pre-training with unlabeled data can improve detection\naccuracy with limited labels. Contemporary methods adapt best-practices for\nself-supervised learning from the image domain to point clouds (such as\ncontrastive learning). However, publicly available 3D datasets are considerably\nsmaller and less diverse than those used for image-based self-supervised\nlearning, limiting their effectiveness. We do note, however, that such 3D data\nis naturally collected in a multimodal fashion, often paired with images.\nRather than pre-training with only self-supervised objectives, we argue that it\nis better to bootstrap point cloud representations using image-based foundation\nmodels trained on internet-scale data. Specifically, we propose a\nshelf-supervised approach (e.g. supervised with off-the-shelf image foundation\nmodels) for generating zero-shot 3D bounding boxes from paired RGB and LiDAR\ndata. Pre-training 3D detectors with such pseudo-labels yields significantly\nbetter semi-supervised detection accuracy than prior self-supervised pretext\ntasks. Importantly, we show that image-based shelf-supervision is helpful for\ntraining LiDAR-only, RGB-only and multi-modal (RGB + LiDAR) detectors. We\ndemonstrate the effectiveness of our approach on nuScenes and WOD,\nsignificantly improving over prior work in limited data settings. Our code is\navailable at https://github.com/meharkhurana03/cm3d\n", "date": "2024-06-14", "categories": []} {"id": "2410.11358", "title": "SeaDATE: Remedy Dual-Attention Transformer with Semantic Alignment via\n Contrast Learning for Multimodal Object Detection", "abstract": " Multimodal object detection leverages diverse modal information to enhance\nthe accuracy and robustness of detectors. By learning long-term dependencies,\nTransformer can effectively integrate multimodal features in the feature\nextraction stage, which greatly improves the performance of multimodal object\ndetection. However, current methods merely stack Transformer-guided fusion\ntechniques without exploring their capability to extract features at various\ndepth layers of network, thus limiting the improvements in detection\nperformance. In this paper, we introduce an accurate and efficient object\ndetection method named SeaDATE. Initially, we propose a novel dual attention\nFeature Fusion (DTF) module that, under Transformer's guidance, integrates\nlocal and global information through a dual attention mechanism, strengthening\nthe fusion of modal features from orthogonal perspectives using spatial and\nchannel tokens. Meanwhile, our theoretical analysis and empirical validation\ndemonstrate that the Transformer-guided fusion method, treating images as\nsequences of pixels for fusion, performs better on shallow features' detail\ninformation compared to deep semantic information. To address this, we designed\na contrastive learning (CL) module aimed at learning features of multimodal\nsamples, remedying the shortcomings of Transformer-guided fusion in extracting\ndeep semantic features, and effectively utilizing cross-modal information.\nExtensive experiments and ablation studies on the FLIR, LLVIP, and M3FD\ndatasets have proven our method to be effective, achieving state-of-the-art\ndetection performance.\n", "date": "2024-10-15", "categories": []} {"id": "2304.11906", "title": "Transformer-based stereo-aware 3D object detection from binocular images", "abstract": " Transformers have shown promising progress in various visual object detection\ntasks, including monocular 2D/3D detection and surround-view 3D detection. More\nimportantly, the attention mechanism in the Transformer model and the 3D\ninformation extraction in binocular stereo are both similarity-based. However,\ndirectly applying existing Transformer-based detectors to binocular stereo 3D\nobject detection leads to slow convergence and significant precision drops. We\nargue that a key cause of that defect is that existing Transformers ignore the\nbinocular-stereo-specific image correspondence information. In this paper, we\nexplore the model design of Transformers in binocular 3D object detection,\nfocusing particularly on extracting and encoding task-specific image\ncorrespondence information. To achieve this goal, we present TS3D, a\nTransformer-based Stereo-aware 3D object detector. In the TS3D, a\nDisparity-Aware Positional Encoding (DAPE) module is proposed to embed the\nimage correspondence information into stereo features. The correspondence is\nencoded as normalized sub-pixel-level disparity and is used in conjunction with\nsinusoidal 2D positional encoding to provide the 3D location information of the\nscene. To enrich multi-scale stereo features, we propose a Stereo Preserving\nFeature Pyramid Network (SPFPN). The SPFPN is designed to preserve the\ncorrespondence information while fusing intra-scale and aggregating cross-scale\nstereo features. Our proposed TS3D achieves a 41.29% Moderate Car detection\naverage precision on the KITTI test set and takes 88 ms to detect objects from\neach binocular image pair. It is competitive with advanced counterparts in\nterms of both precision and inference speed.\n", "date": "2023-04-24", "categories": []} {"id": "2312.00633", "title": "Towards Efficient 3D Object Detection in Bird's-Eye-View Space for\n Autonomous Driving: A Convolutional-Only Approach", "abstract": " 3D object detection in Bird's-Eye-View (BEV) space has recently emerged as a\nprevalent approach in the field of autonomous driving. Despite the demonstrated\nimprovements in accuracy and velocity estimation compared to perspective view\nmethods, the deployment of BEV-based techniques in real-world autonomous\nvehicles remains challenging. This is primarily due to their reliance on\nvision-transformer (ViT) based architectures, which introduce quadratic\ncomplexity with respect to the input resolution. To address this issue, we\npropose an efficient BEV-based 3D detection framework called BEVENet, which\nleverages a convolutional-only architectural design to circumvent the\nlimitations of ViT models while maintaining the effectiveness of BEV-based\nmethods. Our experiments show that BEVENet is 3$\\times$ faster than\ncontemporary state-of-the-art (SOTA) approaches on the NuScenes challenge,\nachieving a mean average precision (mAP) of 0.456 and a nuScenes detection\nscore (NDS) of 0.555 on the NuScenes validation dataset, with an inference\nspeed of 47.6 frames per second. To the best of our knowledge, this study\nstands as the first to achieve such significant efficiency improvements for\nBEV-based methods, highlighting their enhanced feasibility for real-world\nautonomous driving applications.\n", "date": "2023-12-01", "categories": []} {"id": "2408.00619", "title": "Harnessing Uncertainty-aware Bounding Boxes for Unsupervised 3D Object\n Detection", "abstract": " Unsupervised 3D object detection aims to identify objects of interest from\nunlabeled raw data, such as LiDAR points. Recent approaches usually adopt\npseudo 3D bounding boxes (3D bboxes) from clustering algorithm to initialize\nthe model training. However, pseudo bboxes inevitably contain noise, and such\ninaccuracies accumulate to the final model, compromising the performance.\nTherefore, in an attempt to mitigate the negative impact of inaccurate pseudo\nbboxes, we introduce a new uncertainty-aware framework for unsupervised 3D\nobject detection, dubbed UA3D. In particular, our method consists of two\nphases: uncertainty estimation and uncertainty regularization. (1) In the\nuncertainty estimation phase, we incorporate an extra auxiliary detection\nbranch alongside the original primary detector. The prediction disparity\nbetween the primary and auxiliary detectors could reflect fine-grained\nuncertainty at the box coordinate level. (2) Based on the assessed uncertainty,\nwe adaptively adjust the weight of every 3D bbox coordinate via uncertainty\nregularization, refining the training process on pseudo bboxes. For pseudo bbox\ncoordinate with high uncertainty, we assign a relatively low loss weight.\nExtensive experiments verify that the proposed method is robust against the\nnoisy pseudo bboxes, yielding substantial improvements on nuScenes and Lyft\ncompared to existing approaches, with increases of +6.9% AP$_{BEV}$ and +2.5%\nAP$_{3D}$ on nuScenes, and +4.1% AP$_{BEV}$ and +2.0% AP$_{3D}$ on Lyft.\n", "date": "2024-08-01", "categories": []} {"id": "2410.06516", "title": "QuadBEV: An Efficient Quadruple-Task Perception Framework via\n Bird's-Eye-View Representation", "abstract": " Bird's-Eye-View (BEV) perception has become a vital component of autonomous\ndriving systems due to its ability to integrate multiple sensor inputs into a\nunified representation, enhancing performance in various downstream tasks.\nHowever, the computational demands of BEV models pose challenges for real-world\ndeployment in vehicles with limited resources. To address these limitations, we\npropose QuadBEV, an efficient multitask perception framework that leverages the\nshared spatial and contextual information across four key tasks: 3D object\ndetection, lane detection, map segmentation, and occupancy prediction. QuadBEV\nnot only streamlines the integration of these tasks using a shared backbone and\ntask-specific heads but also addresses common multitask learning challenges\nsuch as learning rate sensitivity and conflicting task objectives. Our\nframework reduces redundant computations, thereby enhancing system efficiency,\nmaking it particularly suited for embedded systems. We present comprehensive\nexperiments that validate the effectiveness and robustness of QuadBEV,\ndemonstrating its suitability for real-world applications.\n", "date": "2024-10-09", "categories": []} {"id": "2410.01270", "title": "Panopticus: Omnidirectional 3D Object Detection on Resource-constrained\n Edge Devices", "abstract": " 3D object detection with omnidirectional views enables safety-critical\napplications such as mobile robot navigation. Such applications increasingly\noperate on resource-constrained edge devices, facilitating reliable processing\nwithout privacy concerns or network delays. To enable cost-effective\ndeployment, cameras have been widely adopted as a low-cost alternative to LiDAR\nsensors. However, the compute-intensive workload to achieve high performance of\ncamera-based solutions remains challenging due to the computational limitations\nof edge devices. In this paper, we present Panopticus, a carefully designed\nsystem for omnidirectional and camera-based 3D detection on edge devices.\nPanopticus employs an adaptive multi-branch detection scheme that accounts for\nspatial complexities. To optimize the accuracy within latency limits,\nPanopticus dynamically adjusts the model's architecture and operations based on\navailable edge resources and spatial characteristics. We implemented Panopticus\non three edge devices and conducted experiments across real-world environments\nbased on the public self-driving dataset and our mobile 360{\\deg} camera\ndataset. Experiment results showed that Panopticus improves accuracy by 62% on\naverage given the strict latency objective of 33ms. Also, Panopticus achieves a\n2.1{\\times} latency reduction on average compared to baselines.\n", "date": "2024-10-02", "categories": []} {"id": "2410.01319", "title": "Finetuning Pre-trained Model with Limited Data for LiDAR-based 3D Object\n Detection by Bridging Domain Gaps", "abstract": " LiDAR-based 3D object detectors have been largely utilized in various\napplications, including autonomous vehicles or mobile robots. However,\nLiDAR-based detectors often fail to adapt well to target domains with different\nsensor configurations (e.g., types of sensors, spatial resolution, or FOVs) and\nlocation shifts. Collecting and annotating datasets in a new setup is commonly\nrequired to reduce such gaps, but it is often expensive and time-consuming.\nRecent studies suggest that pre-trained backbones can be learned in a\nself-supervised manner with large-scale unlabeled LiDAR frames. However,\ndespite their expressive representations, they remain challenging to generalize\nwell without substantial amounts of data from the target domain. Thus, we\npropose a novel method, called Domain Adaptive Distill-Tuning (DADT), to adapt\na pre-trained model with limited target data (approximately 100 LiDAR frames),\nretaining its representation power and preventing it from overfitting.\nSpecifically, we use regularizers to align object-level and context-level\nrepresentations between the pre-trained and finetuned models in a\nteacher-student architecture. Our experiments with driving benchmarks, i.e.,\nWaymo Open dataset and KITTI, confirm that our method effectively finetunes a\npre-trained model, achieving significant gains in accuracy.\n", "date": "2024-10-02", "categories": []} {"id": "2406.18129", "title": "CTS: Sim-to-Real Unsupervised Domain Adaptation on 3D Detection", "abstract": " Simulation data can be accurately labeled and have been expected to improve\nthe performance of data-driven algorithms, including object detection. However,\ndue to the various domain inconsistencies from simulation to reality\n(sim-to-real),cross-domain object detection algorithms usually suffer from\ndramatic performance drops. While numerous unsupervised domain adaptation (UDA)\nmethods have been developed to address cross-domain tasks between real-world\ndatasets, progress in sim-to-real remains limited. This paper presents a novel\nComplex-to-Simple (CTS) framework to transfer models from labeled simulation\n(source) to unlabeled reality (target) domains. Based on a two-stage detector,\nthe novelty of this work is threefold: 1) developing fixed-size anchor heads\nand RoI augmentation to address size bias and feature diversity between two\ndomains, thereby improving the quality of pseudo-label; 2) developing a novel\ncorner-format representation of aleatoric uncertainty (AU) for the bounding\nbox, to uniformly quantify pseudo-label quality; 3) developing a noise-aware\nmean teacher domain adaptation method based on AU, as well as object-level and\nframe-level sampling strategies, to migrate the impact of noisy labels.\nExperimental results demonstrate that our proposed approach significantly\nenhances the sim-to-real domain adaptation capability of 3D object detection\nmodels, outperforming state-of-the-art cross-domain algorithms, which are\nusually developed for real-to-real UDA tasks.\n", "date": "2024-06-26", "categories": []} {"id": "2409.20558", "title": "Uni$^2$Det: Unified and Universal Framework for Prompt-Guided\n Multi-dataset 3D Detection", "abstract": " We present Uni$^2$Det, a brand new framework for unified and universal\nmulti-dataset training on 3D detection, enabling robust performance across\ndiverse domains and generalization to unseen domains. Due to substantial\ndisparities in data distribution and variations in taxonomy across diverse\ndomains, training such a detector by simply merging datasets poses a\nsignificant challenge. Motivated by this observation, we introduce multi-stage\nprompting modules for multi-dataset 3D detection, which leverages prompts based\non the characteristics of corresponding datasets to mitigate existing\ndifferences. This elegant design facilitates seamless plug-and-play integration\nwithin various advanced 3D detection frameworks in a unified manner, while also\nallowing straightforward adaptation for universal applicability across\ndatasets. Experiments are conducted across multiple dataset consolidation\nscenarios involving KITTI, Waymo, and nuScenes, demonstrating that our\nUni$^2$Det outperforms existing methods by a large margin in multi-dataset\ntraining. Notably, results on zero-shot cross-dataset transfer validate the\ngeneralization capability of our proposed method.\n", "date": "2024-09-30", "categories": []} {"id": "2310.00920", "title": "Every Dataset Counts: Scaling up Monocular 3D Object Detection with\n Joint Datasets Training", "abstract": " Monocular 3D object detection plays a crucial role in autonomous driving.\nHowever, existing monocular 3D detection algorithms depend on 3D labels derived\nfrom LiDAR measurements, which are costly to acquire for new datasets and\nchallenging to deploy in novel environments. Specifically, this study\ninvestigates the pipeline for training a monocular 3D object detection model on\na diverse collection of 3D and 2D datasets. The proposed framework comprises\nthree components: (1) a robust monocular 3D model capable of functioning across\nvarious camera settings, (2) a selective-training strategy to accommodate\ndatasets with differing class annotations, and (3) a pseudo 3D training\napproach using 2D labels to enhance detection performance in scenes containing\nonly 2D labels. With this framework, we could train models on a joint set of\nvarious open 3D/2D datasets to obtain models with significantly stronger\ngeneralization capability and enhanced performance on new dataset with only 2D\nlabels. We conduct extensive experiments on\nKITTI/nuScenes/ONCE/Cityscapes/BDD100K datasets to demonstrate the scaling\nability of the proposed method.\n", "date": "2023-10-02", "categories": []} {"id": "2404.04629", "title": "DifFUSER: Diffusion Model for Robust Multi-Sensor Fusion in 3D Object\n Detection and BEV Segmentation", "abstract": " Diffusion models have recently gained prominence as powerful deep generative\nmodels, demonstrating unmatched performance across various domains. However,\ntheir potential in multi-sensor fusion remains largely unexplored. In this\nwork, we introduce DifFUSER, a novel approach that leverages diffusion models\nfor multi-modal fusion in 3D object detection and BEV map segmentation.\nBenefiting from the inherent denoising property of diffusion, DifFUSER is able\nto refine or even synthesize sensor features in case of sensor malfunction,\nthereby improving the quality of the fused output. In terms of architecture,\nour DifFUSER blocks are chained together in a hierarchical BiFPN fashion,\ntermed cMini-BiFPN, offering an alternative architecture for latent diffusion.\nWe further introduce a Gated Self-conditioned Modulated (GSM) latent diffusion\nmodule together with a Progressive Sensor Dropout Training (PSDT) paradigm,\ndesigned to add stronger conditioning to the diffusion process and robustness\nto sensor failures. Our extensive evaluations on the Nuscenes dataset reveal\nthat DifFUSER not only achieves state-of-the-art performance with a 70.04% mIOU\nin BEV map segmentation tasks but also competes effectively with leading\ntransformer-based fusion techniques in 3D object detection.\n", "date": "2024-04-06", "categories": []} {"id": "2310.10391", "title": "Open-CRB: Towards Open World Active Learning for 3D Object Detection", "abstract": " LiDAR-based 3D object detection has recently seen significant advancements\nthrough active learning (AL), attaining satisfactory performance by training on\na small fraction of strategically selected point clouds. However, in real-world\ndeployments where streaming point clouds may include unknown or novel objects,\nthe ability of current AL methods to capture such objects remains unexplored.\nThis paper investigates a more practical and challenging research task: Open\nWorld Active Learning for 3D Object Detection (OWAL-3D), aimed at acquiring\ninformative point clouds with new concepts. To tackle this challenge, we\npropose a simple yet effective strategy called Open Label Conciseness (OLC),\nwhich mines novel 3D objects with minimal annotation costs. Our empirical\nresults show that OLC successfully adapts the 3D detection model to the open\nworld scenario with just a single round of selection. Any generic AL policy can\nthen be integrated with the proposed OLC to efficiently address the OWAL-3D\nproblem. Based on this, we introduce the Open-CRB framework, which seamlessly\nintegrates OLC with our preliminary AL method, CRB, designed specifically for\n3D object detection. We develop a comprehensive codebase for easy reproducing\nand future research, supporting 15 baseline methods (\\textit{i.e.}, active\nlearning, out-of-distribution detection and open world detection), 2 types of\nmodern 3D detectors (\\textit{i.e.}, one-stage SECOND and two-stage PV-RCNN) and\n3 benchmark 3D datasets (\\textit{i.e.}, KITTI, nuScenes and Waymo). Extensive\nexperiments evidence that the proposed Open-CRB demonstrates superiority and\nflexibility in recognizing both novel and known classes with very limited\nlabeling costs, compared to state-of-the-art baselines. Source code is\navailable at \\url{https://github.com/Luoyadan/CRB-active-3Ddet/tree/Open-CRB}.\n", "date": "2023-10-16", "categories": []} {"id": "2402.18573", "title": "Towards Unified 3D Object Detection via Algorithm and Data Unification", "abstract": " Realizing unified 3D object detection, including both indoor and outdoor\nscenes, holds great importance in applications like robot navigation. However,\ninvolving various scenarios of data to train models poses challenges due to\ntheir significantly distinct characteristics, \\eg, diverse geometry properties\nand heterogeneous domain distributions. In this work, we propose to address the\nchallenges from two perspectives, the algorithm perspective and data\nperspective. In terms of the algorithm perspective, we first build a monocular\n3D object detector based on the bird's-eye-view (BEV) detection paradigm, where\nthe explicit feature projection is beneficial to addressing the geometry\nlearning ambiguity. In this detector, we split the classical BEV detection\narchitecture into two stages and propose an uneven BEV grid design to handle\nthe convergence instability caused by geometry difference between scenarios.\nBesides, we develop a sparse BEV feature projection strategy to reduce the\ncomputational cost and a unified domain alignment method to handle\nheterogeneous domains. From the data perspective, we propose to incorporate\ndepth information to improve training robustness. Specifically, we build the\nfirst unified multi-modal 3D object detection benchmark MM-Omni3D and extend\nthe aforementioned monocular detector to its multi-modal version, which is the\nfirst unified multi-modal 3D object detector. We name the designed monocular\nand multi-modal detectors as UniMODE and MM-UniMODE, respectively. The\nexperimental results reveal several insightful findings highlighting the\nbenefits of multi-modal data and confirm the effectiveness of all the proposed\nstrategies.\n", "date": "2024-02-28", "categories": []} {"id": "2403.05817", "title": "SAFDNet: A Simple and Effective Network for Fully Sparse 3D Object\n Detection", "abstract": " LiDAR-based 3D object detection plays an essential role in autonomous\ndriving. Existing high-performing 3D object detectors usually build dense\nfeature maps in the backbone network and prediction head. However, the\ncomputational costs introduced by the dense feature maps grow quadratically as\nthe perception range increases, making these models hard to scale up to\nlong-range detection. Some recent works have attempted to construct fully\nsparse detectors to solve this issue; nevertheless, the resulting models either\nrely on a complex multi-stage pipeline or exhibit inferior performance. In this\nwork, we propose SAFDNet, a straightforward yet highly effective architecture,\ntailored for fully sparse 3D object detection. In SAFDNet, an adaptive feature\ndiffusion strategy is designed to address the center feature missing problem.\nWe conducted extensive experiments on Waymo Open, nuScenes, and Argoverse2\ndatasets. SAFDNet performed slightly better than the previous SOTA on the first\ntwo datasets but much better on the last dataset, which features long-range\ndetection, verifying the efficacy of SAFDNet in scenarios where long-range\ndetection is required. Notably, on Argoverse2, SAFDNet surpassed the previous\nbest hybrid detector HEDNet by 2.6% mAP while being 2.1x faster, and yielded\n2.1% mAP gains over the previous best sparse detector FSDv2 while being 1.3x\nfaster. The code will be available at https://github.com/zhanggang001/HEDNet.\n", "date": "2024-03-09", "categories": []} {"id": "2403.09212", "title": "PoIFusion: Multi-Modal 3D Object Detection via Fusion at Points of\n Interest", "abstract": " In this work, we present PoIFusion, a conceptually simple yet effective\nmulti-modal 3D object detection framework to fuse the information of RGB images\nand LiDAR point clouds at the points of interest (PoIs). Different from the\nmost accurate methods to date that transform multi-sensor data into a unified\nview or leverage the global attention mechanism to facilitate fusion, our\napproach maintains the view of each modality and obtains multi-modal features\nby computation-friendly projection and interpolation. In particular, our\nPoIFusion follows the paradigm of query-based object detection, formulating\nobject queries as dynamic 3D boxes and generating a set of PoIs based on each\nquery box. The PoIs serve as the keypoints to represent a 3D object and play\nthe role of the basic units in multi-modal fusion. Specifically, we project\nPoIs into the view of each modality to sample the corresponding feature and\nintegrate the multi-modal features at each PoI through a dynamic fusion block.\nFurthermore, the features of PoIs derived from the same query box are\naggregated together to update the query feature. Our approach prevents\ninformation loss caused by view transformation and eliminates the\ncomputation-intensive global attention, making the multi-modal 3D object\ndetector more applicable. We conducted extensive experiments on nuScenes and\nArgoverse2 datasets to evaluate our approach. Remarkably, the proposed approach\nachieves state-of-the-art results on both datasets without any bells and\nwhistles, \\emph{i.e.}, 74.9\\% NDS and 73.4\\% mAP on nuScenes, and 31.6\\% CDS\nand 40.6\\% mAP on Argoverse2. The code will be made available at\n\\url{https://djiajunustc.github.io/projects/poifusion}.\n", "date": "2024-03-14", "categories": []} {"id": "2409.06583", "title": "Semi-Supervised 3D Object Detection with Channel Augmentation using\n Transformation Equivariance", "abstract": " Accurate 3D object detection is crucial for autonomous vehicles and robots to\nnavigate and interact with the environment safely and effectively. Meanwhile,\nthe performance of 3D detector relies on the data size and annotation which is\nexpensive. Consequently, the demand of training with limited labeled data is\ngrowing. We explore a novel teacher-student framework employing channel\naugmentation for 3D semi-supervised object detection. The teacher-student SSL\ntypically adopts a weak augmentation and strong augmentation to teacher and\nstudent, respectively. In this work, we apply multiple channel augmentations to\nboth networks using the transformation equivariance detector (TED). The TED\nallows us to explore different combinations of augmentation on point clouds and\nefficiently aggregates multi-channel transformation equivariance features. In\nprinciple, by adopting fixed channel augmentations for the teacher network, the\nstudent can train stably on reliable pseudo-labels. Adopting strong channel\naugmentations can enrich the diversity of data, fostering robustness to\ntransformations and enhancing generalization performance of the student\nnetwork. We use SOTA hierarchical supervision as a baseline and adapt its\ndual-threshold to TED, which is called channel IoU consistency. We evaluate our\nmethod with KITTI dataset, and achieved a significant performance leap,\nsurpassing SOTA 3D semi-supervised object detection models.\n", "date": "2024-09-10", "categories": []} {"id": "2409.10699", "title": "CoMamba: Real-time Cooperative Perception Unlocked with State Space\n Models", "abstract": " Cooperative perception systems play a vital role in enhancing the safety and\nefficiency of vehicular autonomy. Although recent studies have highlighted the\nefficacy of vehicle-to-everything (V2X) communication techniques in autonomous\ndriving, a significant challenge persists: how to efficiently integrate\nmultiple high-bandwidth features across an expanding network of connected\nagents such as vehicles and infrastructure. In this paper, we introduce\nCoMamba, a novel cooperative 3D detection framework designed to leverage\nstate-space models for real-time onboard vehicle perception. Compared to prior\nstate-of-the-art transformer-based models, CoMamba enjoys being a more scalable\n3D model using bidirectional state space models, bypassing the quadratic\ncomplexity pain-point of attention mechanisms. Through extensive\nexperimentation on V2X/V2V datasets, CoMamba achieves superior performance\ncompared to existing methods while maintaining real-time processing\ncapabilities. The proposed framework not only enhances object detection\naccuracy but also significantly reduces processing time, making it a promising\nsolution for next-generation cooperative perception systems in intelligent\ntransportation networks.\n", "date": "2024-09-16", "categories": []} {"id": "2409.14751", "title": "UniBEVFusion: Unified Radar-Vision BEVFusion for 3D Object Detection", "abstract": " 4D millimeter-wave (MMW) radar, which provides both height information and\ndense point cloud data over 3D MMW radar, has become increasingly popular in 3D\nobject detection. In recent years, radar-vision fusion models have demonstrated\nperformance close to that of LiDAR-based models, offering advantages in terms\nof lower hardware costs and better resilience in extreme conditions. However,\nmany radar-vision fusion models treat radar as a sparse LiDAR, underutilizing\nradar-specific information. Additionally, these multi-modal networks are often\nsensitive to the failure of a single modality, particularly vision. To address\nthese challenges, we propose the Radar Depth Lift-Splat-Shoot (RDL) module,\nwhich integrates radar-specific data into the depth prediction process,\nenhancing the quality of visual Bird-Eye View (BEV) features. We further\nintroduce a Unified Feature Fusion (UFF) approach that extracts BEV features\nacross different modalities using shared module. To assess the robustness of\nmulti-modal models, we develop a novel Failure Test (FT) ablation experiment,\nwhich simulates vision modality failure by injecting Gaussian noise. We conduct\nextensive experiments on the View-of-Delft (VoD) and TJ4D datasets. The results\ndemonstrate that our proposed Unified BEVFusion (UniBEVFusion) network\nsignificantly outperforms state-of-the-art models on the TJ4D dataset, with\nimprovements of 1.44 in 3D and 1.72 in BEV object detection accuracy.\n", "date": "2024-09-23", "categories": []} {"id": "2404.02517", "title": "HENet: Hybrid Encoding for End-to-end Multi-task 3D Perception from\n Multi-view Cameras", "abstract": " Three-dimensional perception from multi-view cameras is a crucial component\nin autonomous driving systems, which involves multiple tasks like 3D object\ndetection and bird's-eye-view (BEV) semantic segmentation. To improve\nperception precision, large image encoders, high-resolution images, and\nlong-term temporal inputs have been adopted in recent 3D perception models,\nbringing remarkable performance gains. However, these techniques are often\nincompatible in training and inference scenarios due to computational resource\nconstraints. Besides, modern autonomous driving systems prefer to adopt an\nend-to-end framework for multi-task 3D perception, which can simplify the\noverall system architecture and reduce the implementation complexity. However,\nconflict between tasks often arises when optimizing multiple tasks jointly\nwithin an end-to-end 3D perception model. To alleviate these issues, we present\nan end-to-end framework named HENet for multi-task 3D perception in this paper.\nSpecifically, we propose a hybrid image encoding network, using a large image\nencoder for short-term frames and a small image encoder for long-term temporal\nframes. Then, we introduce a temporal feature integration module based on the\nattention mechanism to fuse the features of different frames extracted by the\ntwo aforementioned hybrid image encoders. Finally, according to the\ncharacteristics of each perception task, we utilize BEV features of different\ngrid sizes, independent BEV encoders, and task decoders for different tasks.\nExperimental results show that HENet achieves state-of-the-art end-to-end\nmulti-task 3D perception results on the nuScenes benchmark, including 3D object\ndetection and BEV semantic segmentation. The source code and models will be\nreleased at https://github.com/VDIGPKU/HENet.\n", "date": "2024-04-03", "categories": []} {"id": "2409.11542", "title": "VALO: A Versatile Anytime Framework for LiDAR-based Object Detection\n Deep Neural Networks", "abstract": " This work addresses the challenge of adapting dynamic deadline requirements\nfor LiDAR object detection deep neural networks (DNNs). The computing latency\nof object detection is critically important to ensure safe and efficient\nnavigation. However, state-of-the-art LiDAR object detection DNNs often exhibit\nsignificant latency, hindering their real-time performance on\nresource-constrained edge platforms. Therefore, a tradeoff between detection\naccuracy and latency should be dynamically managed at runtime to achieve\noptimum results.\n In this paper, we introduce VALO (Versatile Anytime algorithm for LiDAR\nObject detection), a novel data-centric approach that enables anytime computing\nof 3D LiDAR object detection DNNs. VALO employs a deadline-aware scheduler to\nselectively process input regions, making execution time and accuracy tradeoffs\nwithout architectural modifications. Additionally, it leverages efficient\nforecasting of past detection results to mitigate possible loss of accuracy due\nto partial processing of input. Finally, it utilizes a novel input reduction\ntechnique within its detection heads to significantly accelerate execution\nwithout sacrificing accuracy.\n We implement VALO on state-of-the-art 3D LiDAR object detection networks,\nnamely CenterPoint and VoxelNext, and demonstrate its dynamic adaptability to a\nwide range of time constraints while achieving higher accuracy than the prior\nstate-of-the-art. Code is available\nathttps://github.com/CSL-KU/VALO}{github.com/CSL-KU/VALO.\n", "date": "2024-09-17", "categories": []} {"id": "2409.10901", "title": "TrajSSL: Trajectory-Enhanced Semi-Supervised 3D Object Detection", "abstract": " Semi-supervised 3D object detection is a common strategy employed to\ncircumvent the challenge of manually labeling large-scale autonomous driving\nperception datasets. Pseudo-labeling approaches to semi-supervised learning\nadopt a teacher-student framework in which machine-generated pseudo-labels on a\nlarge unlabeled dataset are used in combination with a small manually-labeled\ndataset for training. In this work, we address the problem of improving\npseudo-label quality through leveraging long-term temporal information captured\nin driving scenes. More specifically, we leverage pre-trained\nmotion-forecasting models to generate object trajectories on pseudo-labeled\ndata to further enhance the student model training. Our approach improves\npseudo-label quality in two distinct manners: first, we suppress false positive\npseudo-labels through establishing consistency across multiple frames of motion\nforecasting outputs. Second, we compensate for false negative detections by\ndirectly inserting predicted object tracks into the pseudo-labeled scene.\nExperiments on the nuScenes dataset demonstrate the effectiveness of our\napproach, improving the performance of standard semi-supervised approaches in a\nvariety of settings.\n", "date": "2024-09-17", "categories": []} {"id": "2310.02815", "title": "CoBEV: Elevating Roadside 3D Object Detection with Depth and Height\n Complementarity", "abstract": " Roadside camera-driven 3D object detection is a crucial task in intelligent\ntransportation systems, which extends the perception range beyond the\nlimitations of vision-centric vehicles and enhances road safety. While previous\nstudies have limitations in using only depth or height information, we find\nboth depth and height matter and they are in fact complementary. The depth\nfeature encompasses precise geometric cues, whereas the height feature is\nprimarily focused on distinguishing between various categories of height\nintervals, essentially providing semantic context. This insight motivates the\ndevelopment of Complementary-BEV (CoBEV), a novel end-to-end monocular 3D\nobject detection framework that integrates depth and height to construct robust\nBEV representations. In essence, CoBEV estimates each pixel's depth and height\ndistribution and lifts the camera features into 3D space for lateral fusion\nusing the newly proposed two-stage complementary feature selection (CFS)\nmodule. A BEV feature distillation framework is also seamlessly integrated to\nfurther enhance the detection accuracy from the prior knowledge of the\nfusion-modal CoBEV teacher. We conduct extensive experiments on the public 3D\ndetection benchmarks of roadside camera-based DAIR-V2X-I and Rope3D, as well as\nthe private Supremind-Road dataset, demonstrating that CoBEV not only achieves\nthe accuracy of the new state-of-the-art, but also significantly advances the\nrobustness of previous methods in challenging long-distance scenarios and noisy\ncamera disturbance, and enhances generalization by a large margin in\nheterologous settings with drastic changes in scene and camera parameters. For\nthe first time, the vehicle AP score of a camera model reaches 80% on\nDAIR-V2X-I in terms of easy mode. The source code will be made publicly\navailable at https://github.com/MasterHow/CoBEV.\n", "date": "2023-10-04", "categories": []} {"id": "2403.05402", "title": "DualBEV: Unifying Dual View Transformation with Probabilistic\n Correspondences", "abstract": " Camera-based Bird's-Eye-View (BEV) perception often struggles between\nadopting 3D-to-2D or 2D-to-3D view transformation (VT). The 3D-to-2D VT\ntypically employs resource-intensive Transformer to establish robust\ncorrespondences between 3D and 2D features, while the 2D-to-3D VT utilizes the\nLift-Splat-Shoot (LSS) pipeline for real-time application, potentially missing\ndistant information. To address these limitations, we propose DualBEV, a\nunified framework that utilizes a shared feature transformation incorporating\nthree probabilistic measurements for both strategies. By considering dual-view\ncorrespondences in one stage, DualBEV effectively bridges the gap between these\nstrategies, harnessing their individual strengths. Our method achieves\nstate-of-the-art performance without Transformer, delivering comparable\nefficiency to the LSS approach, with 55.2% mAP and 63.4% NDS on the nuScenes\ntest set. Code is available at \\url{https://github.com/PeidongLi/DualBEV}\n", "date": "2024-03-08", "categories": []} {"id": "2409.06945", "title": "FSMDet: Vision-guided feature diffusion for fully sparse 3D detector", "abstract": " Fully sparse 3D detection has attracted an increasing interest in the recent\nyears. However, the sparsity of the features in these frameworks challenges the\ngeneration of proposals because of the limited diffusion process. In addition,\nthe quest for efficiency has led to only few work on vision-assisted fully\nsparse models. In this paper, we propose FSMDet (Fully Sparse Multi-modal\nDetection), which use visual information to guide the LiDAR feature diffusion\nprocess while still maintaining the efficiency of the pipeline. Specifically,\nmost of fully sparse works focus on complex customized center fusion\ndiffusion/regression operators. However, we observed that if the adequate\nobject completion is performed, even the simplest interpolation operator leads\nto satisfactory results. Inspired by this observation, we split the\nvision-guided diffusion process into two modules: a Shape Recover Layer\n(SRLayer) and a Self Diffusion Layer (SDLayer). The former uses RGB information\nto recover the shape of the visible part of an object, and the latter uses a\nvisual prior to further spread the features to the center region. Experiments\ndemonstrate that our approach successfully improves the performance of previous\nfully sparse models that use LiDAR only and reaches SOTA performance in\nmultimodal models. At the same time, thanks to the sparse architecture, our\nmethod can be up to 5 times more efficient than previous SOTA methods in the\ninference process.\n", "date": "2024-09-11", "categories": []} {"id": "2409.04601", "title": "Multi-scale Feature Fusion with Point Pyramid for 3D Object Detection", "abstract": " Effective point cloud processing is crucial to LiDARbased autonomous driving\nsystems. The capability to understand features at multiple scales is required\nfor object detection of intelligent vehicles, where road users may appear in\ndifferent sizes. Recent methods focus on the design of the feature aggregation\noperators, which collect features at different scales from the encoder backbone\nand assign them to the points of interest. While efforts are made into the\naggregation modules, the importance of how to fuse these multi-scale features\nhas been overlooked. This leads to insufficient feature communication across\nscales. To address this issue, this paper proposes the Point Pyramid RCNN\n(POP-RCNN), a feature pyramid-based framework for 3D object detection on point\nclouds. POP-RCNN consists of a Point Pyramid Feature Enhancement (PPFE) module\nto establish connections across spatial scales and semantic depths for\ninformation exchange. The PPFE module effectively fuses multi-scale features\nfor rich information without the increased complexity in feature aggregation.\nTo remedy the impact of inconsistent point densities, a point density\nconfidence module is deployed. This design integration enables the use of a\nlightweight feature aggregator, and the emphasis on both shallow and deep\nsemantics, realising a detection framework for 3D object detection. With great\nadaptability, the proposed method can be applied to a variety of existing\nframeworks to increase feature richness, especially for long-distance\ndetection. By adopting the PPFE in the voxel-based and point-voxel-based\nbaselines, experimental results on KITTI and Waymo Open Dataset show that the\nproposed method achieves remarkable performance even with limited computational\nheadroom.\n", "date": "2024-09-06", "categories": []} {"id": "2409.04979", "title": "RCBEVDet++: Toward High-accuracy Radar-Camera Fusion 3D Perception\n Network", "abstract": " Perceiving the surrounding environment is a fundamental task in autonomous\ndriving. To obtain highly accurate perception results, modern autonomous\ndriving systems typically employ multi-modal sensors to collect comprehensive\nenvironmental data. Among these, the radar-camera multi-modal perception system\nis especially favored for its excellent sensing capabilities and\ncost-effectiveness. However, the substantial modality differences between radar\nand camera sensors pose challenges in fusing information. To address this\nproblem, this paper presents RCBEVDet, a radar-camera fusion 3D object\ndetection framework. Specifically, RCBEVDet is developed from an existing\ncamera-based 3D object detector, supplemented by a specially designed radar\nfeature extractor, RadarBEVNet, and a Cross-Attention Multi-layer Fusion (CAMF)\nmodule. Firstly, RadarBEVNet encodes sparse radar points into a dense\nbird's-eye-view (BEV) feature using a dual-stream radar backbone and a Radar\nCross Section aware BEV encoder. Secondly, the CAMF module utilizes a\ndeformable attention mechanism to align radar and camera BEV features and\nadopts channel and spatial fusion layers to fuse them. To further enhance\nRCBEVDet's capabilities, we introduce RCBEVDet++, which advances the CAMF\nthrough sparse fusion, supports query-based multi-view camera perception\nmodels, and adapts to a broader range of perception tasks. Extensive\nexperiments on the nuScenes show that our method integrates seamlessly with\nexisting camera-based 3D perception models and improves their performance\nacross various perception tasks. Furthermore, our method achieves\nstate-of-the-art radar-camera fusion results in 3D object detection, BEV\nsemantic segmentation, and 3D multi-object tracking tasks. Notably, with ViT-L\nas the image backbone, RCBEVDet++ achieves 72.73 NDS and 67.34 mAP in 3D object\ndetection without test-time augmentation or model ensembling.\n", "date": "2024-09-08", "categories": []} {"id": "2409.04980", "title": "Multi-V2X: A Large Scale Multi-modal Multi-penetration-rate Dataset for\n Cooperative Perception", "abstract": " Cooperative perception through vehicle-to-everything (V2X) has garnered\nsignificant attention in recent years due to its potential to overcome\nocclusions and enhance long-distance perception. Great achievements have been\nmade in both datasets and algorithms. However, existing real-world datasets are\nlimited by the presence of few communicable agents, while synthetic datasets\ntypically cover only vehicles. More importantly, the penetration rate of\nconnected and autonomous vehicles (CAVs) , a critical factor for the deployment\nof cooperative perception technologies, has not been adequately addressed. To\ntackle these issues, we introduce Multi-V2X, a large-scale, multi-modal,\nmulti-penetration-rate dataset for V2X perception. By co-simulating SUMO and\nCARLA, we equip a substantial number of cars and roadside units (RSUs) in\nsimulated towns with sensor suites, and collect comprehensive sensing data.\nDatasets with specified CAV penetration rates can be obtained by masking some\nequipped cars as normal vehicles. In total, our Multi-V2X dataset comprises\n549k RGB frames, 146k LiDAR frames, and 4,219k annotated 3D bounding boxes\nacross six categories. The highest possible CAV penetration rate reaches\n86.21%, with up to 31 agents in communication range, posing new challenges in\nselecting agents to collaborate with. We provide comprehensive benchmarks for\ncooperative 3D object detection tasks. Our data and code are available at\nhttps://github.com/RadetzkyLi/Multi-V2X .\n", "date": "2024-09-08", "categories": []} {"id": "2409.05564", "title": "LEROjD: Lidar Extended Radar-Only Object Detection", "abstract": " Accurate 3D object detection is vital for automated driving. While lidar\nsensors are well suited for this task, they are expensive and have limitations\nin adverse weather conditions. 3+1D imaging radar sensors offer a\ncost-effective, robust alternative but face challenges due to their low\nresolution and high measurement noise. Existing 3+1D imaging radar datasets\ninclude radar and lidar data, enabling cross-modal model improvements. Although\nlidar should not be used during inference, it can aid the training of\nradar-only object detectors. We explore two strategies to transfer knowledge\nfrom the lidar to the radar domain and radar-only object detectors: 1.\nmulti-stage training with sequential lidar point cloud thin-out, and 2.\ncross-modal knowledge distillation. In the multi-stage process, three thin-out\nmethods are examined. Our results show significant performance gains of up to\n4.2 percentage points in mean Average Precision with multi-stage training and\nup to 3.9 percentage points with knowledge distillation by initializing the\nstudent with the teacher's weights. The main benefit of these approaches is\ntheir applicability to other 3D object detection networks without altering\ntheir architecture, as we show by analyzing it on two different object\ndetectors. Our code is available at https://github.com/rst-tu-dortmund/lerojd\n", "date": "2024-09-09", "categories": []} {"id": "2409.04234", "title": "UniDet3D: Multi-dataset Indoor 3D Object Detection", "abstract": " Growing customer demand for smart solutions in robotics and augmented reality\nhas attracted considerable attention to 3D object detection from point clouds.\nYet, existing indoor datasets taken individually are too small and\ninsufficiently diverse to train a powerful and general 3D object detection\nmodel. In the meantime, more general approaches utilizing foundation models are\nstill inferior in quality to those based on supervised training for a specific\ntask. In this work, we propose \\ours{}, a simple yet effective 3D object\ndetection model, which is trained on a mixture of indoor datasets and is\ncapable of working in various indoor environments. By unifying different label\nspaces, \\ours{} enables learning a strong representation across multiple\ndatasets through a supervised joint training scheme. The proposed network\narchitecture is built upon a vanilla transformer encoder, making it easy to\nrun, customize and extend the prediction pipeline for practical use. Extensive\nexperiments demonstrate that \\ours{} obtains significant gains over existing 3D\nobject detection methods in 6 indoor benchmarks: ScanNet (+1.1 mAP50),\nARKitScenes (+19.4 mAP25), S3DIS (+9.1 mAP50), MultiScan (+9.3 mAP50), 3RScan\n(+3.2 mAP50), and ScanNet++ (+2.7 mAP50). Code is available at\nhttps://github.com/filapro/unidet3d .\n", "date": "2024-09-06", "categories": []} {"id": "2409.04390", "title": "Future Does Matter: Boosting 3D Object Detection with Temporal Motion\n Estimation in Point Cloud Sequences", "abstract": " Accurate and robust LiDAR 3D object detection is essential for comprehensive\nscene understanding in autonomous driving. Despite its importance, LiDAR\ndetection performance is limited by inherent constraints of point cloud data,\nparticularly under conditions of extended distances and occlusions. Recently,\ntemporal aggregation has been proven to significantly enhance detection\naccuracy by fusing multi-frame viewpoint information and enriching the spatial\nrepresentation of objects. In this work, we introduce a novel LiDAR 3D object\ndetection framework, namely LiSTM, to facilitate spatial-temporal feature\nlearning with cross-frame motion forecasting information. We aim to improve the\nspatial-temporal interpretation capabilities of the LiDAR detector by\nincorporating a dynamic prior, generated from a non-learnable motion estimation\nmodel. Specifically, Motion-Guided Feature Aggregation (MGFA) is proposed to\nutilize the object trajectory from previous and future motion states to model\nspatial-temporal correlations into gaussian heatmap over a driving sequence.\nThis motion-based heatmap then guides the temporal feature fusion, enriching\nthe proposed object features. Moreover, we design a Dual Correlation Weighting\nModule (DCWM) that effectively facilitates the interaction between past and\nprospective frames through scene- and channel-wise feature abstraction. In the\nend, a cascade cross-attention-based decoder is employed to refine the 3D\nprediction. We have conducted experiments on the Waymo and nuScenes datasets to\ndemonstrate that the proposed framework achieves superior 3D detection\nperformance with effective spatial-temporal feature learning.\n", "date": "2024-09-06", "categories": []} {"id": "2402.07680", "title": "AYDIV: Adaptable Yielding 3D Object Detection via Integrated Contextual\n Vision Transformer", "abstract": " Combining LiDAR and camera data has shown potential in enhancing\nshort-distance object detection in autonomous driving systems. Yet, the fusion\nencounters difficulties with extended distance detection due to the contrast\nbetween LiDAR's sparse data and the dense resolution of cameras. Besides,\ndiscrepancies in the two data representations further complicate fusion\nmethods. We introduce AYDIV, a novel framework integrating a tri-phase\nalignment process specifically designed to enhance long-distance detection even\namidst data discrepancies. AYDIV consists of the Global Contextual Fusion\nAlignment Transformer (GCFAT), which improves the extraction of camera features\nand provides a deeper understanding of large-scale patterns; the Sparse Fused\nFeature Attention (SFFA), which fine-tunes the fusion of LiDAR and camera\ndetails; and the Volumetric Grid Attention (VGA) for a comprehensive spatial\ndata fusion. AYDIV's performance on the Waymo Open Dataset (WOD) with an\nimprovement of 1.24% in mAPH value(L2 difficulty) and the Argoverse2 Dataset\nwith a performance improvement of 7.40% in AP value demonstrates its efficacy\nin comparison to other existing fusion-based methods. Our code is publicly\navailable at https://github.com/sanjay-810/AYDIV2\n", "date": "2024-02-12", "categories": []} {"id": "2409.00633", "title": "Make Your ViT-based Multi-view 3D Detectors Faster via Token Compression", "abstract": " Slow inference speed is one of the most crucial concerns for deploying\nmulti-view 3D detectors to tasks with high real-time requirements like\nautonomous driving. Although many sparse query-based methods have already\nattempted to improve the efficiency of 3D detectors, they neglect to consider\nthe backbone, especially when using Vision Transformers (ViT) for better\nperformance. To tackle this problem, we explore the efficient ViT backbones for\nmulti-view 3D detection via token compression and propose a simple yet\neffective method called TokenCompression3D (ToC3D). By leveraging history\nobject queries as foreground priors of high quality, modeling 3D motion\ninformation in them, and interacting them with image tokens through the\nattention mechanism, ToC3D can effectively determine the magnitude of\ninformation densities of image tokens and segment the salient foreground\ntokens. With the introduced dynamic router design, ToC3D can weigh more\ncomputing resources to important foreground tokens while compressing the\ninformation loss, leading to a more efficient ViT-based multi-view 3D detector.\nExtensive results on the large-scale nuScenes dataset show that our method can\nnearly maintain the performance of recent SOTA with up to 30% inference\nspeedup, and the improvements are consistent after scaling up the ViT and input\nresolution. The code will be made at https://github.com/DYZhang09/ToC3D.\n", "date": "2024-09-01", "categories": []} {"id": "2409.00673", "title": "Study of Dropout in PointPillars with 3D Object Detection", "abstract": " 3D object detection is critical for autonomous driving, leveraging deep\nlearning techniques to interpret LiDAR data. The PointPillars architecture is a\nprominent model in this field, distinguished by its efficient use of LiDAR\ndata. This study provides an analysis of enhancing the performance of\nPointPillars model under various dropout rates to address overfitting and\nimprove model generalization. Dropout, a regularization technique, involves\nrandomly omitting neurons during training, compelling the network to learn\nrobust and diverse features. We systematically compare the effects of different\nenhancement techniques on the model's regression performance during training\nand its accuracy, measured by Average Precision (AP) and Average Orientation\nSimilarity (AOS). Our findings offer insights into the optimal enhancements,\ncontributing to improved 3D object detection in autonomous driving\napplications.\n", "date": "2024-09-01", "categories": []} {"id": "2409.00690", "title": "Decoupled and Interactive Regression Modeling for High-performance\n One-stage 3D Object Detection", "abstract": " Inadequate bounding box modeling in regression tasks constrains the\nperformance of one-stage 3D object detection. Our study reveals that the\nprimary reason lies in two aspects: (1) The limited center-offset prediction\nseriously impairs the bounding box localization since many highest response\npositions significantly deviate from object centers. (2) The low-quality sample\nignored in regression tasks significantly impacts the bounding box prediction\nsince it produces unreliable quality (IoU) rectification. To tackle these\nproblems, we propose Decoupled and Interactive Regression Modeling (DIRM) for\none-stage detection. Specifically, Decoupled Attribute Regression (DAR) is\nimplemented to facilitate long regression range modeling for the center\nattribute through an adaptive multi-sample assignment strategy that deeply\ndecouples bounding box attributes. On the other hand, to enhance the\nreliability of IoU predictions for low-quality results, Interactive Quality\nPrediction (IQP) integrates the classification task, proficient in modeling\nnegative samples, with quality prediction for joint optimization. Extensive\nexperiments on Waymo and ONCE datasets demonstrate that DIRM significantly\nimproves the performance of several state-of-the-art methods with minimal\nadditional inference latency. Notably, DIRM achieves state-of-the-art detection\nperformance on both the Waymo and ONCE datasets.\n", "date": "2024-09-01", "categories": []} {"id": "2403.15026", "title": "VRSO: Visual-Centric Reconstruction for Static Object Annotation", "abstract": " As a part of the perception results of intelligent driving systems, static\nobject detection (SOD) in 3D space provides crucial cues for driving\nenvironment understanding. With the rapid deployment of deep neural networks\nfor SOD tasks, the demand for high-quality training samples soars. The\ntraditional, also reliable, way is manual labelling over the dense LiDAR point\nclouds and reference images. Though most public driving datasets adopt this\nstrategy to provide SOD ground truth (GT), it is still expensive and\ntime-consuming in practice. This paper introduces VRSO, a visual-centric\napproach for static object annotation. Experiments on the Waymo Open Dataset\nshow that the mean reprojection error from VRSO annotation is only 2.6 pixels,\naround four times lower than the Waymo Open Dataset labels (10.6 pixels). VRSO\nis distinguished in low cost, high efficiency, and high quality: (1) It\nrecovers static objects in 3D space with only camera images as input, and (2)\nmanual annotation is barely involved since GT for SOD tasks is generated based\non an automatic reconstruction and annotation pipeline.\n", "date": "2024-03-22", "categories": []} {"id": "2408.17036", "title": "CP-VoteNet: Contrastive Prototypical VoteNet for Few-Shot Point Cloud\n Object Detection", "abstract": " Few-shot point cloud 3D object detection (FS3D) aims to identify and localise\nobjects of novel classes from point clouds, using knowledge learnt from\nannotated base classes and novel classes with very few annotations. Thus far,\nthis challenging task has been approached using prototype learning, but the\nperformance remains far from satisfactory. We find that in existing methods,\nthe prototypes are only loosely constrained and lack of fine-grained awareness\nof the semantic and geometrical correlation embedded within the point cloud\nspace. To mitigate these issues, we propose to leverage the inherent\ncontrastive relationship within the semantic and geometrical subspaces to learn\nmore refined and generalisable prototypical representations. To this end, we\nfirst introduce contrastive semantics mining, which enables the network to\nextract discriminative categorical features by constructing positive and\nnegative pairs within training batches. Meanwhile, since point features\nrepresenting local patterns can be clustered into geometric components, we\nfurther propose to impose contrastive relationship at the primitive level.\nThrough refined primitive geometric structures, the transferability of feature\nencoding from base to novel classes is significantly enhanced. The above\ndesigns and insights lead to our novel Contrastive Prototypical VoteNet\n(CP-VoteNet). Extensive experiments on two FS3D benchmarks FS-ScanNet and\nFS-SUNRGBD demonstrate that CP-VoteNet surpasses current state-of-the-art\nmethods by considerable margins across different FS3D settings. Further\nablation studies conducted corroborate the rationale and effectiveness of our\ndesigns.\n", "date": "2024-08-30", "categories": []} {"id": "2408.15643", "title": "RIDE: Boosting 3D Object Detection for LiDAR Point Clouds via\n Rotation-Invariant Analysis", "abstract": " The rotation robustness property has drawn much attention to point cloud\nanalysis, whereas it still poses a critical challenge in 3D object detection.\nWhen subjected to arbitrary rotation, most existing detectors fail to produce\nexpected outputs due to the poor rotation robustness. In this paper, we present\nRIDE, a pioneering exploration of Rotation-Invariance for the 3D\nLiDAR-point-based object DEtector, with the key idea of designing\nrotation-invariant features from LiDAR scenes and then effectively\nincorporating them into existing 3D detectors. Specifically, we design a\nbi-feature extractor that extracts (i) object-aware features though sensitive\nto rotation but preserve geometry well, and (ii) rotation-invariant features,\nwhich lose geometric information to a certain extent but are robust to\nrotation. These two kinds of features complement each other to decode 3D\nproposals that are robust to arbitrary rotations. Particularly, our RIDE is\ncompatible and easy to plug into the existing one-stage and two-stage 3D\ndetectors, and boosts both detection performance and rotation robustness.\nExtensive experiments on the standard benchmarks showcase that the mean average\nprecision (mAP) and rotation robustness can be significantly boosted by\nintegrating with our RIDE, with +5.6% mAP and 53% rotation robustness\nimprovement on KITTI, +5.1% and 28% improvement correspondingly on nuScenes.\nThe code will be available soon.\n", "date": "2024-08-28", "categories": []} {"id": "2408.16530", "title": "A Comprehensive Review of 3D Object Detection in Autonomous Driving:\n Technological Advances and Future Directions", "abstract": " In recent years, 3D object perception has become a crucial component in the\ndevelopment of autonomous driving systems, providing essential environmental\nawareness. However, as perception tasks in autonomous driving evolve, their\nvariants have increased, leading to diverse insights from industry and\nacademia. Currently, there is a lack of comprehensive surveys that collect and\nsummarize these perception tasks and their developments from a broader\nperspective. This review extensively summarizes traditional 3D object detection\nmethods, focusing on camera-based, LiDAR-based, and fusion detection\ntechniques. We provide a comprehensive analysis of the strengths and\nlimitations of each approach, highlighting advancements in accuracy and\nrobustness. Furthermore, we discuss future directions, including methods to\nimprove accuracy such as temporal perception, occupancy grids, and end-to-end\nlearning frameworks. We also explore cooperative perception methods that extend\nthe perception range through collaborative communication. By providing a\nholistic view of the current state and future developments in 3D object\nperception, we aim to offer a more comprehensive understanding of perception\ntasks for autonomous driving. Additionally, we have established an active\nrepository to provide continuous updates on the latest advancements in this\nfield, accessible at:\nhttps://github.com/Fishsoup0/Autonomous-Driving-Perception.\n", "date": "2024-08-28", "categories": []} {"id": "2408.15428", "title": "HEAD: A Bandwidth-Efficient Cooperative Perception Approach for\n Heterogeneous Connected and Autonomous Vehicles", "abstract": " In cooperative perception studies, there is often a trade-off between\ncommunication bandwidth and perception performance. While current feature\nfusion solutions are known for their excellent object detection performance,\ntransmitting the entire sets of intermediate feature maps requires substantial\nbandwidth. Furthermore, these fusion approaches are typically limited to\nvehicles that use identical detection models. Our goal is to develop a solution\nthat supports cooperative perception across vehicles equipped with different\nmodalities of sensors. This method aims to deliver improved perception\nperformance compared to late fusion techniques, while achieving precision\nsimilar to the state-of-art intermediate fusion, but requires an order of\nmagnitude less bandwidth. We propose HEAD, a method that fuses features from\nthe classification and regression heads in 3D object detection networks. Our\nmethod is compatible with heterogeneous detection networks such as LiDAR\nPointPillars, SECOND, VoxelNet, and camera Bird's-eye View (BEV) Encoder. Given\nthe naturally smaller feature size in the detection heads, we design a\nself-attention mechanism to fuse the classification head and a complementary\nfeature fusion layer to fuse the regression head. Our experiments,\ncomprehensively evaluated on the V2V4Real and OPV2V datasets, demonstrate that\nHEAD is a fusion method that effectively balances communication bandwidth and\nperception performance.\n", "date": "2024-08-27", "categories": []} {"id": "2408.15637", "title": "Transfer Learning from Simulated to Real Scenes for Monocular 3D Object\n Detection", "abstract": " Accurately detecting 3D objects from monocular images in dynamic roadside\nscenarios remains a challenging problem due to varying camera perspectives and\nunpredictable scene conditions. This paper introduces a two-stage training\nstrategy to address these challenges. Our approach initially trains a model on\nthe large-scale synthetic dataset, RoadSense3D, which offers a diverse range of\nscenarios for robust feature learning. Subsequently, we fine-tune the model on\na combination of real-world datasets to enhance its adaptability to practical\nconditions. Experimental results of the Cube R-CNN model on challenging public\nbenchmarks show a remarkable improvement in detection performance, with a mean\naverage precision rising from 0.26 to 12.76 on the TUM Traffic A9 Highway\ndataset and from 2.09 to 6.60 on the DAIR-V2X-I dataset when performing\ntransfer learning. Code, data, and qualitative video results are available on\nthe project website: https://roadsense3d.github.io.\n", "date": "2024-08-28", "categories": []} {"id": "2408.02088", "title": "KAN-RCBEVDepth: A multi-modal fusion algorithm in object detection for\n autonomous driving", "abstract": " Accurate 3D object detection in autonomous driving is critical yet\nchallenging due to occlusions, varying object sizes, and complex urban\nenvironments. This paper introduces the KAN-RCBEVDepth method, an innovative\napproach aimed at enhancing 3D object detection by fusing multimodal sensor\ndata from cameras, LiDAR, and millimeter-wave radar. Our unique Bird's Eye\nView-based approach significantly improves detection accuracy and efficiency by\nseamlessly integrating diverse sensor inputs, refining spatial relationship\nunderstanding, and optimizing computational procedures. Experimental results\nshow that the proposed method outperforms existing techniques across multiple\ndetection metrics, achieving a higher Mean Distance AP (0.389, 23\\%\nimprovement), a better ND Score (0.485, 17.1\\% improvement), and a faster\nEvaluation Time (71.28s, 8\\% faster). Additionally, the KAN-RCBEVDepth method\nsignificantly reduces errors compared to BEVDepth, with lower Transformation\nError (0.6044, 13.8\\% improvement), Scale Error (0.2780, 2.6\\% improvement),\nOrientation Error (0.5830, 7.6\\% improvement), Velocity Error (0.4244, 28.3\\%\nimprovement), and Attribute Error (0.2129, 3.2\\% improvement). These findings\nsuggest that our method offers enhanced accuracy, reliability, and efficiency,\nmaking it well-suited for dynamic and demanding autonomous driving scenarios.\nThe code will be released in \\url{https://github.com/laitiamo/RCBEVDepth-KAN}.\n", "date": "2024-08-04", "categories": []} {"id": "2408.14600", "title": "PVAFN: Point-Voxel Attention Fusion Network with Multi-Pooling Enhancing\n for 3D Object Detection", "abstract": " The integration of point and voxel representations is becoming more common in\nLiDAR-based 3D object detection. However, this combination often struggles with\ncapturing semantic information effectively. Moreover, relying solely on point\nfeatures within regions of interest can lead to information loss and\nlimitations in local feature representation. To tackle these challenges, we\npropose a novel two-stage 3D object detector, called Point-Voxel Attention\nFusion Network (PVAFN). PVAFN leverages an attention mechanism to improve\nmulti-modal feature fusion during the feature extraction phase. In the\nrefinement stage, it utilizes a multi-pooling strategy to integrate both\nmulti-scale and region-specific information effectively. The point-voxel\nattention mechanism adaptively combines point cloud and voxel-based\nBird's-Eye-View (BEV) features, resulting in richer object representations that\nhelp to reduce false detections. Additionally, a multi-pooling enhancement\nmodule is introduced to boost the model's perception capabilities. This module\nemploys cluster pooling and pyramid pooling techniques to efficiently capture\nkey geometric details and fine-grained shape structures, thereby enhancing the\nintegration of local and global features. Extensive experiments on the KITTI\nand Waymo datasets demonstrate that the proposed PVAFN achieves competitive\nperformance. The code and models will be available.\n", "date": "2024-08-26", "categories": []} {"id": "2408.14941", "title": "BOX3D: Lightweight Camera-LiDAR Fusion for 3D Object Detection and\n Localization", "abstract": " Object detection and global localization play a crucial role in robotics,\nspanning across a great spectrum of applications from autonomous cars to\nmulti-layered 3D Scene Graphs for semantic scene understanding. This article\nproposes BOX3D, a novel multi-modal and lightweight scheme for localizing\nobjects of interest by fusing the information from RGB camera and 3D LiDAR.\nBOX3D is structured around a three-layered architecture, building up from the\nlocal perception of the incoming sequential sensor data to the global\nperception refinement that covers for outliers and the general consistency of\neach object's observation. More specifically, the first layer handles the\nlow-level fusion of camera and LiDAR data for initial 3D bounding box\nextraction. The second layer converts each LiDAR's scan 3D bounding boxes to\nthe world coordinate frame and applies a spatial pairing and merging mechanism\nto maintain the uniqueness of objects observed from different viewpoints.\nFinally, BOX3D integrates the third layer that supervises the consistency of\nthe results on the global map iteratively, using a point-to-voxel comparison\nfor identifying all points in the global map that belong to the object.\nBenchmarking results of the proposed novel architecture are showcased in\nmultiple experimental trials on public state-of-the-art large-scale dataset of\nurban environments.\n", "date": "2024-08-27", "categories": []} {"id": "2404.09431", "title": "VFMM3D: Releasing the Potential of Image by Vision Foundation Model for\n Monocular 3D Object Detection", "abstract": " Due to its cost-effectiveness and widespread availability, monocular 3D\nobject detection, which relies solely on a single camera during inference,\nholds significant importance across various applications, including autonomous\ndriving and robotics. Nevertheless, directly predicting the coordinates of\nobjects in 3D space from monocular images poses challenges. Therefore, an\neffective solution involves transforming monocular images into LiDAR-like\nrepresentations and employing a LiDAR-based 3D object detector to predict the\n3D coordinates of objects. The key step in this method is accurately converting\nthe monocular image into a reliable point cloud form. In this paper, we present\nVFMM3D, an innovative framework that leverages the capabilities of Vision\nFoundation Models (VFMs) to accurately transform single-view images into LiDAR\npoint cloud representations. VFMM3D utilizes the Segment Anything Model (SAM)\nand Depth Anything Model (DAM) to generate high-quality pseudo-LiDAR data\nenriched with rich foreground information. Specifically, the Depth Anything\nModel (DAM) is employed to generate dense depth maps. Subsequently, the Segment\nAnything Model (SAM) is utilized to differentiate foreground and background\nregions by predicting instance masks. These predicted instance masks and depth\nmaps are then combined and projected into 3D space to generate pseudo-LiDAR\npoints. Finally, any object detectors based on point clouds can be utilized to\npredict the 3D coordinates of objects. Comprehensive experiments are conducted\non two challenging 3D object detection datasets, KITTI and Waymo. Our VFMM3D\nestablishes a new state-of-the-art performance on both datasets. Additionally,\nexperimental results demonstrate the generality of VFMM3D, showcasing its\nseamless integration into various LiDAR-based 3D object detectors.\n", "date": "2024-04-15", "categories": []} {"id": "2406.11311", "title": "Syn-to-Real Unsupervised Domain Adaptation for Indoor 3D Object\n Detection", "abstract": " The use of synthetic data in indoor 3D object detection offers the potential\nof greatly reducing the manual labor involved in 3D annotations and training\neffective zero-shot detectors. However, the complicated domain shifts across\nsyn-to-real indoor datasets remains underexplored. In this paper, we propose a\nnovel Object-wise Hierarchical Domain Alignment (OHDA) framework for\nsyn-to-real unsupervised domain adaptation in indoor 3D object detection. Our\napproach includes an object-aware augmentation strategy to effectively\ndiversify the source domain data, and we introduce a two-branch adaptation\nframework consisting of an adversarial training branch and a pseudo labeling\nbranch, in order to simultaneously reach holistic-level and class-level domain\nalignment. The pseudo labeling is further refined through two proposed schemes\nspecifically designed for indoor UDA. Our adaptation results from synthetic\ndataset 3D-FRONT to real-world datasets ScanNetV2 and SUN RGB-D demonstrate\nremarkable mAP25 improvements of 9.7% and 9.1% over Source-Only baselines,\nrespectively, and consistently outperform the methods adapted from 2D and 3D\noutdoor scenarios. The code will be publicly available upon paper acceptance.\n", "date": "2024-06-17", "categories": []} {"id": "2408.13798", "title": "Selectively Dilated Convolution for Accuracy-Preserving Sparse\n Pillar-based Embedded 3D Object Detection", "abstract": " Pillar-based 3D object detection has gained traction in self-driving\ntechnology due to its speed and accuracy facilitated by the artificial\ndensification of pillars for GPU-friendly processing. However, dense pillar\nprocessing fundamentally wastes computation since it ignores the inherent\nsparsity of pillars derived from scattered point cloud data. Motivated by\nrecent embedded accelerators with native sparsity support, sparse pillar\nconvolution methods like submanifold convolution (SubM-Conv) aimed to reduce\nthese redundant computations by applying convolution only on active pillars but\nsuffered considerable accuracy loss. Our research identifies that this accuracy\nloss is due to the restricted fine-grained spatial information flow (fSIF) of\nSubM-Conv in sparse pillar networks. To overcome this restriction, we propose a\nselectively dilated (SD-Conv) convolution that evaluates the importance of\nencoded pillars and selectively dilates the convolution output, enhancing the\nreceptive field for critical pillars and improving object detection accuracy.\nTo facilitate actual acceleration with this novel convolution approach, we\ndesigned SPADE+ as a cost-efficient augmentation to existing embedded sparse\nconvolution accelerators. This design supports the SD-Conv without significant\ndemands in area and SRAM size, realizing superior trade-off between the speedup\nand model accuracy. This strategic enhancement allows our method to achieve\nextreme pillar sparsity, leading to up to 18.1x computational savings and 16.2x\nspeedup on the embedded accelerators, without compromising object detection\naccuracy.\n", "date": "2024-08-25", "categories": []} {"id": "2408.13902", "title": "TraIL-Det: Transformation-Invariant Local Feature Networks for 3D LiDAR\n Object Detection with Unsupervised Pre-Training", "abstract": " 3D point clouds are essential for perceiving outdoor scenes, especially\nwithin the realm of autonomous driving. Recent advances in 3D LiDAR Object\nDetection focus primarily on the spatial positioning and distribution of points\nto ensure accurate detection. However, despite their robust performance in\nvariable conditions, these methods are hindered by their sole reliance on\ncoordinates and point intensity, resulting in inadequate isometric invariance\nand suboptimal detection outcomes. To tackle this challenge, our work\nintroduces Transformation-Invariant Local (TraIL) features and the associated\nTraIL-Det architecture. Our TraIL features exhibit rigid transformation\ninvariance and effectively adapt to variations in point density, with a design\nfocus on capturing the localized geometry of neighboring structures. They\nutilize the inherent isotropic radiation of LiDAR to enhance local\nrepresentation, improve computational efficiency, and boost detection\nperformance. To effectively process the geometric relations among points within\neach proposal, we propose a Multi-head self-Attention Encoder (MAE) with\nasymmetric geometric features to encode high-dimensional TraIL features into\nmanageable representations. Our method outperforms contemporary self-supervised\n3D object detection approaches in terms of mAP on KITTI (67.8, 20% label,\nmoderate) and Waymo (68.9, 20% label, moderate) datasets under various label\nratios (20%, 50%, and 100%).\n", "date": "2024-08-25", "categories": []} {"id": "2408.13936", "title": "OpenNav: Efficient Open Vocabulary 3D Object Detection for Smart\n Wheelchair Navigation", "abstract": " Open vocabulary 3D object detection (OV3D) allows precise and extensible\nobject recognition crucial for adapting to diverse environments encountered in\nassistive robotics. This paper presents OpenNav, a zero-shot 3D object\ndetection pipeline based on RGB-D images for smart wheelchairs. Our pipeline\nintegrates an open-vocabulary 2D object detector with a mask generator for\nsemantic segmentation, followed by depth isolation and point cloud construction\nto create 3D bounding boxes. The smart wheelchair exploits these 3D bounding\nboxes to identify potential targets and navigate safely. We demonstrate\nOpenNav's performance through experiments on the Replica dataset and we report\npreliminary results with a real wheelchair. OpenNav improves state-of-the-art\nsignificantly on the Replica dataset at mAP25 (+9pts) and mAP50 (+5pts) with\nmarginal improvement at mAP. The code is publicly available at this link:\nhttps://github.com/EasyWalk-PRIN/OpenNav.\n", "date": "2024-08-25", "categories": []} {"id": "2408.12708", "title": "Revisiting Cross-Domain Problem for LiDAR-based 3D Object Detection", "abstract": " Deep learning models such as convolutional neural networks and transformers\nhave been widely applied to solve 3D object detection problems in the domain of\nautonomous driving. While existing models have achieved outstanding performance\non most open benchmarks, the generalization ability of these deep networks is\nstill in doubt. To adapt models to other domains including different cities,\ncountries, and weather, retraining with the target domain data is currently\nnecessary, which hinders the wide application of autonomous driving. In this\npaper, we deeply analyze the cross-domain performance of the state-of-the-art\nmodels. We observe that most models will overfit the training domains and it is\nchallenging to adapt them to other domains directly. Existing domain adaptation\nmethods for 3D object detection problems are actually shifting the models'\nknowledge domain instead of improving their generalization ability. We then\npropose additional evaluation metrics -- the side-view and front-view AP -- to\nbetter analyze the core issues of the methods' heavy drops in accuracy levels.\nBy using the proposed metrics and further evaluating the cross-domain\nperformance in each dimension, we conclude that the overfitting problem happens\nmore obviously on the front-view surface and the width dimension which usually\nfaces the sensor and has more 3D points surrounding it. Meanwhile, our\nexperiments indicate that the density of the point cloud data also\nsignificantly influences the models' cross-domain performance.\n", "date": "2024-08-22", "categories": []} {"id": "2408.12747", "title": "CatFree3D: Category-agnostic 3D Object Detection with Diffusion", "abstract": " Image-based 3D object detection is widely employed in applications such as\nautonomous vehicles and robotics, yet current systems struggle with\ngeneralisation due to complex problem setup and limited training data. We\nintroduce a novel pipeline that decouples 3D detection from 2D detection and\ndepth prediction, using a diffusion-based approach to improve accuracy and\nsupport category-agnostic detection. Additionally, we introduce the Normalised\nHungarian Distance (NHD) metric for an accurate evaluation of 3D detection\nresults, addressing the limitations of traditional IoU and GIoU metrics.\nExperimental results demonstrate that our method achieves state-of-the-art\naccuracy and strong generalisation across various object categories and\ndatasets.\n", "date": "2024-08-22", "categories": []} {"id": "2309.00928", "title": "S$^3$-MonoDETR: Supervised Shape&Scale-perceptive Deformable Transformer\n for Monocular 3D Object Detection", "abstract": " Recently, transformer-based methods have shown exceptional performance in\nmonocular 3D object detection, which can predict 3D attributes from a single 2D\nimage. These methods typically use visual and depth representations to generate\nquery points on objects, whose quality plays a decisive role in the detection\naccuracy. However, current unsupervised attention mechanisms without any\ngeometry appearance awareness in transformers are susceptible to producing\nnoisy features for query points, which severely limits the network performance\nand also makes the model have a poor ability to detect multi-category objects\nin a single training process. To tackle this problem, this paper proposes a\nnovel ``Supervised Shape&Scale-perceptive Deformable Attention'' (S$^3$-DA)\nmodule for monocular 3D object detection. Concretely, S$^3$-DA utilizes visual\nand depth features to generate diverse local features with various shapes and\nscales and predict the corresponding matching distribution simultaneously to\nimpose valuable shape&scale perception for each query. Benefiting from this,\nS$^3$-DA effectively estimates receptive fields for query points belonging to\nany category, enabling them to generate robust query features. Besides, we\npropose a Multi-classification-based Shape&Scale Matching (MSM) loss to\nsupervise the above process. Extensive experiments on KITTI and Waymo Open\ndatasets demonstrate that S$^3$-DA significantly improves the detection\naccuracy, yielding state-of-the-art performance of single-category and\nmulti-category 3D object detection in a single training process compared to the\nexisting approaches. The source code will be made publicly available at\nhttps://github.com/mikasa3lili/S3-MonoDETR.\n", "date": "2023-09-02", "categories": []} {"id": "2312.07530", "title": "Weakly Supervised 3D Object Detection via Multi-Level Visual Guidance", "abstract": " Weakly supervised 3D object detection aims to learn a 3D detector with lower\nannotation cost, e.g., 2D labels. Unlike prior work which still relies on few\naccurate 3D annotations, we propose a framework to study how to leverage\nconstraints between 2D and 3D domains without requiring any 3D labels.\nSpecifically, we employ visual data from three perspectives to establish\nconnections between 2D and 3D domains. First, we design a feature-level\nconstraint to align LiDAR and image features based on object-aware regions.\nSecond, the output-level constraint is developed to enforce the overlap between\n2D and projected 3D box estimations. Finally, the training-level constraint is\nutilized by producing accurate and consistent 3D pseudo-labels that align with\nthe visual data. We conduct extensive experiments on the KITTI dataset to\nvalidate the effectiveness of the proposed three constraints. Without using any\n3D labels, our method achieves favorable performance against state-of-the-art\napproaches and is competitive with the method that uses 500-frame 3D\nannotations. Code will be made publicly available at\nhttps://github.com/kuanchihhuang/VG-W3D.\n", "date": "2023-12-12", "categories": []} {"id": "2408.11207", "title": "Quantum Inverse Contextual Vision Transformers (Q-ICVT): A New Frontier\n in 3D Object Detection for AVs", "abstract": " The field of autonomous vehicles (AVs) predominantly leverages multi-modal\nintegration of LiDAR and camera data to achieve better performance compared to\nusing a single modality. However, the fusion process encounters challenges in\ndetecting distant objects due to the disparity between the high resolution of\ncameras and the sparse data from LiDAR. Insufficient integration of global\nperspectives with local-level details results in sub-optimal fusion\nperformance.To address this issue, we have developed an innovative two-stage\nfusion process called Quantum Inverse Contextual Vision Transformers (Q-ICVT).\nThis approach leverages adiabatic computing in quantum concepts to create a\nnovel reversible vision transformer known as the Global Adiabatic Transformer\n(GAT). GAT aggregates sparse LiDAR features with semantic features in dense\nimages for cross-modal integration in a global form. Additionally, the Sparse\nExpert of Local Fusion (SELF) module maps the sparse LiDAR 3D proposals and\nencodes position information of the raw point cloud onto the dense camera\nfeature space using a gating point fusion approach. Our experiments show that\nQ-ICVT achieves an mAPH of 82.54 for L2 difficulties on the Waymo dataset,\nimproving by 1.88% over current state-of-the-art fusion methods. We also\nanalyze GAT and SELF in ablation studies to highlight the impact of Q-ICVT. Our\ncode is available at https://github.com/sanjay-810/Qicvt Q-ICVT\n", "date": "2024-08-20", "categories": []} {"id": "2407.19156", "title": "Robust Multimodal 3D Object Detection via Modality-Agnostic Decoding and\n Proximity-based Modality Ensemble", "abstract": " Recent advancements in 3D object detection have benefited from multi-modal\ninformation from the multi-view cameras and LiDAR sensors. However, the\ninherent disparities between the modalities pose substantial challenges. We\nobserve that existing multi-modal 3D object detection methods heavily rely on\nthe LiDAR sensor, treating the camera as an auxiliary modality for augmenting\nsemantic details. This often leads to not only underutilization of camera data\nbut also significant performance degradation in scenarios where LiDAR data is\nunavailable. Additionally, existing fusion methods overlook the detrimental\nimpact of sensor noise induced by environmental changes, on detection\nperformance. In this paper, we propose MEFormer to address the LiDAR\nover-reliance problem by harnessing critical information for 3D object\ndetection from every available modality while concurrently safeguarding against\ncorrupted signals during the fusion process. Specifically, we introduce\nModality Agnostic Decoding (MOAD) that extracts geometric and semantic features\nwith a shared transformer decoder regardless of input modalities and provides\npromising improvement with a single modality as well as multi-modality.\nAdditionally, our Proximity-based Modality Ensemble (PME) module adaptively\nutilizes the strengths of each modality depending on the environment while\nmitigating the effects of a noisy sensor. Our MEFormer achieves\nstate-of-the-art performance of 73.9% NDS and 71.5% mAP in the nuScenes\nvalidation set. Extensive analyses validate that our MEFormer improves\nrobustness against challenging conditions such as sensor malfunctions or\nenvironmental changes. The source code is available at\nhttps://github.com/hanchaa/MEFormer\n", "date": "2024-07-27", "categories": []} {"id": "2408.09122", "title": "MaskBEV: Towards A Unified Framework for BEV Detection and Map\n Segmentation", "abstract": " Accurate and robust multimodal multi-task perception is crucial for modern\nautonomous driving systems. However, current multimodal perception research\nfollows independent paradigms designed for specific perception tasks, leading\nto a lack of complementary learning among tasks and decreased performance in\nmulti-task learning (MTL) due to joint training. In this paper, we propose\nMaskBEV, a masked attention-based MTL paradigm that unifies 3D object detection\nand bird's eye view (BEV) map segmentation. MaskBEV introduces a task-agnostic\nTransformer decoder to process these diverse tasks, enabling MTL to be\ncompleted in a unified decoder without requiring additional design of specific\ntask heads. To fully exploit the complementary information between BEV map\nsegmentation and 3D object detection tasks in BEV space, we propose spatial\nmodulation and scene-level context aggregation strategies. These strategies\nconsider the inherent dependencies between BEV segmentation and 3D detection,\nnaturally boosting MTL performance. Extensive experiments on nuScenes dataset\nshow that compared with previous state-of-the-art MTL methods, MaskBEV achieves\n1.3 NDS improvement in 3D object detection and 2.7 mIoU improvement in BEV map\nsegmentation, while also demonstrating slightly leading inference speed.\n", "date": "2024-08-17", "categories": []} {"id": "2308.05988", "title": "MS3D++: Ensemble of Experts for Multi-Source Unsupervised Domain\n Adaptation in 3D Object Detection", "abstract": " Deploying 3D detectors in unfamiliar domains has been demonstrated to result\nin a significant 70-90% drop in detection rate due to variations in lidar,\ngeography, or weather from their training dataset. This domain gap leads to\nmissing detections for densely observed objects, misaligned confidence scores,\nand increased high-confidence false positives, rendering the detector highly\nunreliable. To address this, we introduce MS3D++, a self-training framework for\nmulti-source unsupervised domain adaptation in 3D object detection. MS3D++\ngenerates high-quality pseudo-labels, allowing 3D detectors to achieve high\nperformance on a range of lidar types, regardless of their density. Our\napproach effectively fuses predictions of an ensemble of multi-frame\npre-trained detectors from different source domains to improve domain\ngeneralization. We subsequently refine predictions temporally to ensure\ntemporal consistency in box localization and object classification.\nFurthermore, we present an in-depth study into the performance and\nidiosyncrasies of various 3D detector components in a cross-domain context,\nproviding valuable insights for improved cross-domain detector ensembling.\nExperimental results on Waymo, nuScenes and Lyft demonstrate that detectors\ntrained with MS3D++ pseudo-labels achieve state-of-the-art performance,\ncomparable to training with human-annotated labels in Bird's Eye View (BEV)\nevaluation for both low and high density lidar. Code is available at\nhttps://github.com/darrenjkt/MS3D\n", "date": "2023-08-11", "categories": []} {"id": "2401.06542", "title": "Robustness-Aware 3D Object Detection in Autonomous Driving: A Review and\n Outlook", "abstract": " In the realm of modern autonomous driving, the perception system is\nindispensable for accurately assessing the state of the surrounding\nenvironment, thereby enabling informed prediction and planning. The key step to\nthis system is related to 3D object detection that utilizes vehicle-mounted\nsensors such as LiDAR and cameras to identify the size, the category, and the\nlocation of nearby objects. Despite the surge in 3D object detection methods\naimed at enhancing detection precision and efficiency, there is a gap in the\nliterature that systematically examines their resilience against environmental\nvariations, noise, and weather changes. This study emphasizes the importance of\nrobustness, alongside accuracy and latency, in evaluating perception systems\nunder practical scenarios. Our work presents an extensive survey of\ncamera-only, LiDAR-only, and multi-modal 3D object detection algorithms,\nthoroughly evaluating their trade-off between accuracy, latency, and\nrobustness, particularly on datasets like KITTI-C and nuScenes-C to ensure fair\ncomparisons. Among these, multi-modal 3D detection approaches exhibit superior\nrobustness, and a novel taxonomy is introduced to reorganize the literature for\nenhanced clarity. This survey aims to offer a more practical perspective on the\ncurrent capabilities and the constraints of 3D object detection algorithms in\nreal-world applications, thus steering future research towards\nrobustness-centric advancements.\n", "date": "2024-01-12", "categories": []} {"id": "2408.06604", "title": "MV-DETR: Multi-modality indoor object detection by Multi-View DEtecton\n TRansformers", "abstract": " We introduce a novel MV-DETR pipeline which is effective while efficient\ntransformer based detection method. Given input RGBD data, we notice that there\nare super strong pretraining weights for RGB data while less effective works\nfor depth related data. First and foremost , we argue that geometry and texture\ncues are both of vital importance while could be encoded separately. Secondly,\nwe find that visual texture feature is relatively hard to extract compared with\ngeometry feature in 3d space. Unfortunately, single RGBD dataset with thousands\nof data is not enough for training an discriminating filter for visual texture\nfeature extraction. Last but certainly not the least, we designed a lightweight\nVG module consists of a visual textual encoder, a geometry encoder and a VG\nconnector. Compared with previous state of the art works like V-DETR, gains\nfrom pretrained visual encoder could be seen. Extensive experiments on\nScanNetV2 dataset shows the effectiveness of our method. It is worth mentioned\nthat our method achieve 78\\% AP which create new state of the art on ScanNetv2\nbenchmark.\n", "date": "2024-08-13", "categories": []} {"id": "2408.06772", "title": "Exploring Domain Shift on Radar-Based 3D Object Detection Amidst Diverse\n Environmental Conditions", "abstract": " The rapid evolution of deep learning and its integration with autonomous\ndriving systems have led to substantial advancements in 3D perception using\nmultimodal sensors. Notably, radar sensors show greater robustness compared to\ncameras and lidar under adverse weather and varying illumination conditions.\nThis study delves into the often-overlooked yet crucial issue of domain shift\nin 4D radar-based object detection, examining how varying environmental\nconditions, such as different weather patterns and road types, impact 3D object\ndetection performance. Our findings highlight distinct domain shifts across\nvarious weather scenarios, revealing unique dataset sensitivities that\nunderscore the critical role of radar point cloud generation. Additionally, we\ndemonstrate that transitioning between different road types, especially from\nhighways to urban settings, introduces notable domain shifts, emphasizing the\nnecessity for diverse data collection across varied road environments. To the\nbest of our knowledge, this is the first comprehensive analysis of domain shift\neffects on 4D radar-based object detection. We believe this empirical study\ncontributes to understanding the complex nature of domain shifts in radar data\nand suggests paths forward for data collection strategy in the face of\nenvironmental variability.\n", "date": "2024-08-13", "categories": []} {"id": "2408.06901", "title": "Divide and Conquer: Improving Multi-Camera 3D Perception with 2D\n Semantic-Depth Priors and Input-Dependent Queries", "abstract": " 3D perception tasks, such as 3D object detection and Bird's-Eye-View (BEV)\nsegmentation using multi-camera images, have drawn significant attention\nrecently. Despite the fact that accurately estimating both semantic and 3D\nscene layouts are crucial for this task, existing techniques often neglect the\nsynergistic effects of semantic and depth cues, leading to the occurrence of\nclassification and position estimation errors. Additionally, the\ninput-independent nature of initial queries also limits the learning capacity\nof Transformer-based models. To tackle these challenges, we propose an\ninput-aware Transformer framework that leverages Semantics and Depth as priors\n(named SDTR). Our approach involves the use of an S-D Encoder that explicitly\nmodels semantic and depth priors, thereby disentangling the learning process of\nobject categorization and position estimation. Moreover, we introduce a\nPrior-guided Query Builder that incorporates the semantic prior into the\ninitial queries of the Transformer, resulting in more effective input-aware\nqueries. Extensive experiments on the nuScenes and Lyft benchmarks demonstrate\nthe state-of-the-art performance of our method in both 3D object detection and\nBEV segmentation tasks.\n", "date": "2024-08-13", "categories": []} {"id": "2308.16518", "title": "MS23D: A 3D Object Detection Method Using Multi-Scale Semantic Feature\n Points to Construct 3D Feature Layer", "abstract": " LiDAR point clouds can effectively depict the motion and posture of objects\nin three-dimensional space. Many studies accomplish the 3D object detection by\nvoxelizing point clouds. However, in autonomous driving scenarios, the sparsity\nand hollowness of point clouds create some difficulties for voxel-based\nmethods. The sparsity of point clouds makes it challenging to describe the\ngeometric features of objects. The hollowness of point clouds poses\ndifficulties for the aggregation of 3D features. We propose a two-stage 3D\nobject detection framework, called MS23D. (1) We propose a method using voxel\nfeature points from multi-branch to construct the 3D feature layer. Using voxel\nfeature points from different branches, we construct a relatively compact 3D\nfeature layer with rich semantic features. Additionally, we propose a\ndistance-weighted sampling method, reducing the loss of foreground points\ncaused by downsampling and allowing the 3D feature layer to retain more\nforeground points. (2) In response to the hollowness of point clouds, we\npredict the offsets between deep-level feature points and the object's\ncentroid, making them as close as possible to the object's centroid. This\nenables the aggregation of these feature points with abundant semantic\nfeatures. For feature points from shallow-level, we retain them on the object's\nsurface to describe the geometric features of the object. To validate our\napproach, we evaluated its effectiveness on both the KITTI and ONCE datasets.\n", "date": "2023-08-31", "categories": []} {"id": "2408.06137", "title": "MR3D-Net: Dynamic Multi-Resolution 3D Sparse Voxel Grid Fusion for\n LiDAR-Based Collective Perception", "abstract": " The safe operation of automated vehicles depends on their ability to perceive\nthe environment comprehensively. However, occlusion, sensor range, and\nenvironmental factors limit their perception capabilities. To overcome these\nlimitations, collective perception enables vehicles to exchange information.\nHowever, fusing this exchanged information is a challenging task. Early fusion\napproaches require large amounts of bandwidth, while intermediate fusion\napproaches face interchangeability issues. Late fusion of shared detections is\ncurrently the only feasible approach. However, it often results in inferior\nperformance due to information loss. To address this issue, we propose\nMR3D-Net, a dynamic multi-resolution 3D sparse voxel grid fusion backbone\narchitecture for LiDAR-based collective perception. We show that sparse voxel\ngrids at varying resolutions provide a meaningful and compact environment\nrepresentation that can adapt to the communication bandwidth. MR3D-Net achieves\nstate-of-the-art performance on the OPV2V 3D object detection benchmark while\nreducing the required bandwidth by up to 94% compared to early fusion. Code is\navailable at https://github.com/ekut-es/MR3D-Net\n", "date": "2024-08-12", "categories": []} {"id": "2405.01258", "title": "Towards Consistent Object Detection via LiDAR-Camera Synergy", "abstract": " As human-machine interaction continues to evolve, the capacity for\nenvironmental perception is becoming increasingly crucial. Integrating the two\nmost common types of sensory data, images, and point clouds, can enhance\ndetection accuracy. Currently, there is no existing model capable of detecting\nan object's position in both point clouds and images while also determining\ntheir corresponding relationship. This information is invaluable for\nhuman-machine interactions, offering new possibilities for their enhancement.\nIn light of this, this paper introduces an end-to-end Consistency Object\nDetection (COD) algorithm framework that requires only a single forward\ninference to simultaneously obtain an object's position in both point clouds\nand images and establish their correlation. Furthermore, to assess the accuracy\nof the object correlation between point clouds and images, this paper proposes\na new evaluation metric, Consistency Precision (CP). To verify the\neffectiveness of the proposed framework, an extensive set of experiments has\nbeen conducted on the KITTI and DAIR-V2X datasets. The study also explored how\nthe proposed consistency detection method performs on images when the\ncalibration parameters between images and point clouds are disturbed, compared\nto existing post-processing methods. The experimental results demonstrate that\nthe proposed method exhibits excellent detection performance and robustness,\nachieving end-to-end consistency detection. The source code will be made\npublicly available at https://github.com/xifen523/COD.\n", "date": "2024-05-02", "categories": []} {"id": "2408.05020", "title": "RadarPillars: Efficient Object Detection from 4D Radar Point Clouds", "abstract": " Automotive radar systems have evolved to provide not only range, azimuth and\nDoppler velocity, but also elevation data. This additional dimension allows for\nthe representation of 4D radar as a 3D point cloud. As a result, existing deep\nlearning methods for 3D object detection, which were initially developed for\nLiDAR data, are often applied to these radar point clouds. However, this\nneglects the special characteristics of 4D radar data, such as the extreme\nsparsity and the optimal utilization of velocity information. To address these\ngaps in the state-of-the-art, we present RadarPillars, a pillar-based object\ndetection network.\n By decomposing radial velocity data, introducing PillarAttention for\nefficient feature extraction, and studying layer scaling to accommodate radar\nsparsity, RadarPillars significantly outperform state-of-the-art detection\nresults on the View-of-Delft dataset. Importantly, this comes at a\nsignificantly reduced parameter count, surpassing existing methods in terms of\nefficiency and enabling real-time performance on edge devices.\n", "date": "2024-08-09", "categories": []} {"id": "2408.03790", "title": "Vision-Language Guidance for LiDAR-based Unsupervised 3D Object\n Detection", "abstract": " Accurate 3D object detection in LiDAR point clouds is crucial for autonomous\ndriving systems. To achieve state-of-the-art performance, the supervised\ntraining of detectors requires large amounts of human-annotated data, which is\nexpensive to obtain and restricted to predefined object categories. To mitigate\nmanual labeling efforts, recent unsupervised object detection approaches\ngenerate class-agnostic pseudo-labels for moving objects, subsequently serving\nas supervision signal to bootstrap a detector. Despite promising results, these\napproaches do not provide class labels or generalize well to static objects.\nFurthermore, they are mostly restricted to data containing multiple drives from\nthe same scene or images from a precisely calibrated and synchronized camera\nsetup. To overcome these limitations, we propose a vision-language-guided\nunsupervised 3D detection approach that operates exclusively on LiDAR point\nclouds. We transfer CLIP knowledge to classify point clusters of static and\nmoving objects, which we discover by exploiting the inherent spatio-temporal\ninformation of LiDAR point clouds for clustering, tracking, as well as box and\nlabel refinement. Our approach outperforms state-of-the-art unsupervised 3D\nobject detectors on the Waymo Open Dataset ($+23~\\text{AP}_{3D}$) and Argoverse\n2 ($+7.9~\\text{AP}_{3D}$) and provides class labels not solely based on object\nsize assumptions, marking a significant advancement in the field.\n", "date": "2024-08-07", "categories": []} {"id": "2403.15313", "title": "CR3DT: Camera-RADAR Fusion for 3D Detection and Tracking", "abstract": " To enable self-driving vehicles accurate detection and tracking of\nsurrounding objects is essential. While Light Detection and Ranging (LiDAR)\nsensors have set the benchmark for high-performance systems, the appeal of\ncamera-only solutions lies in their cost-effectiveness. Notably, despite the\nprevalent use of Radio Detection and Ranging (RADAR) sensors in automotive\nsystems, their potential in 3D detection and tracking has been largely\ndisregarded due to data sparsity and measurement noise. As a recent\ndevelopment, the combination of RADARs and cameras is emerging as a promising\nsolution. This paper presents Camera-RADAR 3D Detection and Tracking (CR3DT), a\ncamera-RADAR fusion model for 3D object detection, and Multi-Object Tracking\n(MOT). Building upon the foundations of the State-of-the-Art (SotA) camera-only\nBEVDet architecture, CR3DT demonstrates substantial improvements in both\ndetection and tracking capabilities, by incorporating the spatial and velocity\ninformation of the RADAR sensor. Experimental results demonstrate an absolute\nimprovement in detection performance of 5.3% in mean Average Precision (mAP)\nand a 14.9% increase in Average Multi-Object Tracking Accuracy (AMOTA) on the\nnuScenes dataset when leveraging both modalities. CR3DT bridges the gap between\nhigh-performance and cost-effective perception systems in autonomous driving,\nby capitalizing on the ubiquitous presence of RADAR in automotive applications.\nThe code is available at: https://github.com/ETH-PBL/CR3DT.\n", "date": "2024-03-22", "categories": []} {"id": "2408.01739", "title": "LAM3D: Leveraging Attention for Monocular 3D Object Detection", "abstract": " Since the introduction of the self-attention mechanism and the adoption of\nthe Transformer architecture for Computer Vision tasks, the Vision\nTransformer-based architectures gained a lot of popularity in the field, being\nused for tasks such as image classification, object detection and image\nsegmentation. However, efficiently leveraging the attention mechanism in vision\ntransformers for the Monocular 3D Object Detection task remains an open\nquestion. In this paper, we present LAM3D, a framework that Leverages\nself-Attention mechanism for Monocular 3D object Detection. To do so, the\nproposed method is built upon a Pyramid Vision Transformer v2 (PVTv2) as\nfeature extraction backbone and 2D/3D detection machinery. We evaluate the\nproposed method on the KITTI 3D Object Detection Benchmark, proving the\napplicability of the proposed solution in the autonomous driving domain and\noutperforming reference methods. Moreover, due to the usage of self-attention,\nLAM3D is able to systematically outperform the equivalent architecture that\ndoes not employ self-attention.\n", "date": "2024-08-03", "categories": []} {"id": "2408.01085", "title": "Effect of Fog Particle Size Distribution on 3D Object Detection Under\n Adverse Weather Conditions", "abstract": " LiDAR-based sensors employing optical spectrum signals play a vital role in\nproviding significant information about the target objects in autonomous\ndriving vehicle systems. However, the presence of fog in the atmosphere\nseverely degrades the overall system's performance. This manuscript analyzes\nthe role of fog particle size distributions in 3D object detection under\nadverse weather conditions. We utilise Mie theory and meteorological optical\nrange (MOR) to calculate the attenuation and backscattering coefficient values\nfor point cloud generation and analyze the overall system's accuracy in Car,\nCyclist, and Pedestrian case scenarios under easy, medium and hard detection\ndifficulties. Gamma and Junge (Power-Law) distributions are employed to\nmathematically model the fog particle size distribution under strong and\nmoderate advection fog environments. Subsequently, we modified the KITTI\ndataset based on the backscattering coefficient values and trained it on the\nPV-RCNN++ deep neural network model for Car, Cyclist, and Pedestrian cases\nunder different detection difficulties. The result analysis shows a significant\nvariation in the system's accuracy concerning the changes in target object\ndimensionality, the nature of the fog environment and increasing detection\ndifficulties, with the Car exhibiting the highest accuracy of around 99% and\nthe Pedestrian showing the lowest accuracy of around 73%.\n", "date": "2024-08-02", "categories": []} {"id": "2408.00286", "title": "Diff3DETR:Agent-based Diffusion Model for Semi-supervised 3D Object\n Detection", "abstract": " 3D object detection is essential for understanding 3D scenes. Contemporary\ntechniques often require extensive annotated training data, yet obtaining\npoint-wise annotations for point clouds is time-consuming and laborious. Recent\ndevelopments in semi-supervised methods seek to mitigate this problem by\nemploying a teacher-student framework to generate pseudo-labels for unlabeled\npoint clouds. However, these pseudo-labels frequently suffer from insufficient\ndiversity and inferior quality. To overcome these hurdles, we introduce an\nAgent-based Diffusion Model for Semi-supervised 3D Object Detection\n(Diff3DETR). Specifically, an agent-based object query generator is designed to\nproduce object queries that effectively adapt to dynamic scenes while striking\na balance between sampling locations and content embedding. Additionally, a\nbox-aware denoising module utilizes the DDIM denoising process and the\nlong-range attention in the transformer decoder to refine bounding boxes\nincrementally. Extensive experiments on ScanNet and SUN RGB-D datasets\ndemonstrate that Diff3DETR outperforms state-of-the-art semi-supervised 3D\nobject detection methods.\n", "date": "2024-08-01", "categories": []} {"id": "2408.00438", "title": "MonoMM: A Multi-scale Mamba-Enhanced Network for Real-time Monocular 3D\n Object Detection", "abstract": " Recent advancements in transformer-based monocular 3D object detection\ntechniques have exhibited exceptional performance in inferring 3D attributes\nfrom single 2D images. However, most existing methods rely on\nresource-intensive transformer architectures, which often lead to significant\ndrops in computational efficiency and performance when handling long sequence\ndata. To address these challenges and advance monocular 3D object detection\ntechnology, we propose an innovative network architecture, MonoMM, a\nMulti-scale \\textbf{M}amba-Enhanced network for real-time Monocular 3D object\ndetection. This well-designed architecture primarily includes the following two\ncore modules: Focused Multi-Scale Fusion (FMF) Module, which focuses on\neffectively preserving and fusing image information from different scales with\nlower computational resource consumption. By precisely regulating the\ninformation flow, the FMF module enhances the model adaptability and robustness\nto scale variations while maintaining image details. Depth-Aware Feature\nEnhancement Mamba (DMB) Module: It utilizes the fused features from image\ncharacteristics as input and employs a novel adaptive strategy to globally\nintegrate depth information and visual information. This depth fusion strategy\nnot only improves the accuracy of depth estimation but also enhances the model\nperformance under different viewing angles and environmental conditions.\nMoreover, the modular design of MonoMM provides high flexibility and\nscalability, facilitating adjustments and optimizations according to specific\napplication needs. Extensive experiments conducted on the KITTI dataset show\nthat our method outperforms previous monocular methods and achieves real-time\ndetection.\n", "date": "2024-08-01", "categories": []} {"id": "2408.00565", "title": "MUFASA: Multi-View Fusion and Adaptation Network with Spatial Awareness\n for Radar Object Detection", "abstract": " In recent years, approaches based on radar object detection have made\nsignificant progress in autonomous driving systems due to their robustness\nunder adverse weather compared to LiDAR. However, the sparsity of radar point\nclouds poses challenges in achieving precise object detection, highlighting the\nimportance of effective and comprehensive feature extraction technologies. To\naddress this challenge, this paper introduces a comprehensive feature\nextraction method for radar point clouds. This study first enhances the\ncapability of detection networks by using a plug-and-play module, GeoSPA. It\nleverages the Lalonde features to explore local geometric patterns.\nAdditionally, a distributed multi-view attention mechanism, DEMVA, is designed\nto integrate the shared information across the entire dataset with the global\ninformation of each individual frame. By employing the two modules, we present\nour method, MUFASA, which enhances object detection performance through\nimproved feature extraction. The approach is evaluated on the VoD and\nTJ4DRaDSet datasets to demonstrate its effectiveness. In particular, we achieve\nstate-of-the-art results among radar-based methods on the VoD dataset with the\nmAP of 50.24%.\n", "date": "2024-08-01", "categories": []} {"id": "2401.03836", "title": "WidthFormer: Toward Efficient Transformer-based BEV View Transformation", "abstract": " We present WidthFormer, a novel transformer-based module to compute\nBird's-Eye-View (BEV) representations from multi-view cameras for real-time\nautonomous-driving applications. WidthFormer is computationally efficient,\nrobust and does not require any special engineering effort to deploy. We first\nintroduce a novel 3D positional encoding mechanism capable of accurately\nencapsulating 3D geometric information, which enables our model to compute\nhigh-quality BEV representations with only a single transformer decoder layer.\nThis mechanism is also beneficial for existing sparse 3D object detectors.\nInspired by the recently proposed works, we further improve our model's\nefficiency by vertically compressing the image features when serving as\nattention keys and values, and then we develop two modules to compensate for\npotential information loss due to feature compression. Experimental evaluation\non the widely-used nuScenes 3D object detection benchmark demonstrates that our\nmethod outperforms previous approaches across different 3D detection\narchitectures. More importantly, our model is highly efficient. For example,\nwhen using $256\\times 704$ input images, it achieves 1.5 ms and 2.8 ms latency\non NVIDIA 3090 GPU and Horizon Journey-5 computation solutions. Furthermore,\nWidthFormer also exhibits strong robustness to different degrees of camera\nperturbations. Our study offers valuable insights into the deployment of BEV\ntransformation methods in real-world, complex road environments. Code is\navailable at https://github.com/ChenhongyiYang/WidthFormer .\n", "date": "2024-01-08", "categories": []} {"id": "2407.20818", "title": "WARM-3D: A Weakly-Supervised Sim2Real Domain Adaptation Framework for\n Roadside Monocular 3D Object Detection", "abstract": " Existing roadside perception systems are limited by the absence of publicly\navailable, large-scale, high-quality 3D datasets. Exploring the use of\ncost-effective, extensive synthetic datasets offers a viable solution to tackle\nthis challenge and enhance the performance of roadside monocular 3D detection.\nIn this study, we introduce the TUMTraf Synthetic Dataset, offering a diverse\nand substantial collection of high-quality 3D data to augment scarce real-world\ndatasets. Besides, we present WARM-3D, a concise yet effective framework to aid\nthe Sim2Real domain transfer for roadside monocular 3D detection. Our method\nleverages cheap synthetic datasets and 2D labels from an off-the-shelf 2D\ndetector for weak supervision. We show that WARM-3D significantly enhances\nperformance, achieving a +12.40% increase in mAP 3D over the baseline with only\npseudo-2D supervision. With 2D GT as weak labels, WARM-3D even reaches\nperformance close to the Oracle baseline. Moreover, WARM-3D improves the\nability of 3D detectors to unseen sample recognition across various real-world\nenvironments, highlighting its potential for practical applications.\n", "date": "2024-07-30", "categories": []} {"id": "2406.13891", "title": "DPO: Dual-Perturbation Optimization for Test-time Adaptation in 3D\n Object Detection", "abstract": " LiDAR-based 3D object detection has seen impressive advances in recent times.\nHowever, deploying trained 3D detectors in the real world often yields\nunsatisfactory performance when the distribution of the test data significantly\ndeviates from the training data due to different weather conditions, object\nsizes, \\textit{etc}. A key factor in this performance degradation is the\ndiminished generalizability of pre-trained models, which creates a sharp loss\nlandscape during training. Such sharpness, when encountered during testing, can\nprecipitate significant performance declines, even with minor data variations.\nTo address the aforementioned challenges, we propose \\textbf{dual-perturbation\noptimization (DPO)} for \\textbf{\\underline{T}est-\\underline{t}ime\n\\underline{A}daptation in \\underline{3}D \\underline{O}bject\n\\underline{D}etection (TTA-3OD)}. We minimize the sharpness to cultivate a flat\nloss landscape to ensure model resiliency to minor data variations, thereby\nenhancing the generalization of the adaptation process. To fully capture the\ninherent variability of the test point clouds, we further introduce adversarial\nperturbation to the input BEV features to better simulate the noisy test\nenvironment. As the dual perturbation strategy relies on trustworthy\nsupervision signals, we utilize a reliable Hungarian matcher to filter out\npseudo-labels sensitive to perturbations. Additionally, we introduce early\nHungarian cutoff to avoid error accumulation from incorrect pseudo-labels by\nhalting the adaptation process. Extensive experiments across three types of\ntransfer tasks demonstrate that the proposed DPO significantly surpasses\nprevious state-of-the-art approaches, specifically on Waymo $\\rightarrow$\nKITTI, outperforming the most competitive baseline by 57.72\\% in\n$\\text{AP}_\\text{3D}$ and reaching 91\\% of the fully supervised upper bound.\n", "date": "2024-06-19", "categories": []} {"id": "2407.16789", "title": "What Matters in Range View 3D Object Detection", "abstract": " Lidar-based perception pipelines rely on 3D object detection models to\ninterpret complex scenes. While multiple representations for lidar exist, the\nrange-view is enticing since it losslessly encodes the entire lidar sensor\noutput. In this work, we achieve state-of-the-art amongst range-view 3D object\ndetection models without using multiple techniques proposed in past range-view\nliterature. We explore range-view 3D object detection across two modern\ndatasets with substantially different properties: Argoverse 2 and Waymo Open.\nOur investigation reveals key insights: (1) input feature dimensionality\nsignificantly influences the overall performance, (2) surprisingly, employing a\nclassification loss grounded in 3D spatial proximity works as well or better\ncompared to more elaborate IoU-based losses, and (3) addressing non-uniform\nlidar density via a straightforward range subsampling technique outperforms\nexisting multi-resolution, range-conditioned networks. Our experiments reveal\nthat techniques proposed in recent range-view literature are not needed to\nachieve state-of-the-art performance. Combining the above findings, we\nestablish a new state-of-the-art model for range-view 3D object detection --\nimproving AP by 2.2% on the Waymo Open dataset while maintaining a runtime of\n10 Hz. We establish the first range-view model on the Argoverse 2 dataset and\noutperform strong voxel-based baselines. All models are multi-class and\nopen-source. Code is available at\nhttps://github.com/benjaminrwilson/range-view-3d-detection.\n", "date": "2024-07-23", "categories": []} {"id": "2310.05989", "title": "QE-BEV: Query Evolution for Bird's Eye View Object Detection in Varied\n Contexts", "abstract": " 3D object detection plays a pivotal role in autonomous driving and robotics,\ndemanding precise interpretation of Bird's Eye View (BEV) images. The dynamic\nnature of real-world environments necessitates the use of dynamic query\nmechanisms in 3D object detection to adaptively capture and process the complex\nspatio-temporal relationships present in these scenes. However, prior\nimplementations of dynamic queries have often faced difficulties in effectively\nleveraging these relationships, particularly when it comes to integrating\ntemporal information in a computationally efficient manner. Addressing this\nlimitation, we introduce a framework utilizing dynamic query evolution\nstrategy, harnesses K-means clustering and Top-K attention mechanisms for\nrefined spatio-temporal data processing. By dynamically segmenting the BEV\nspace and prioritizing key features through Top-K attention, our model achieves\na real-time, focused analysis of pertinent scene elements. Our extensive\nevaluation on the nuScenes and Waymo dataset showcases a marked improvement in\ndetection accuracy, setting a new benchmark in the domain of query-based BEV\nobject detection. Our dynamic query evolution strategy has the potential to\npush the boundaries of current BEV methods with enhanced adaptability and\ncomputational efficiency. Project page:\nhttps://github.com/Jiawei-Yao0812/QE-BEV\n", "date": "2023-10-07", "categories": []} {"id": "2407.18232", "title": "LION: Linear Group RNN for 3D Object Detection in Point Clouds", "abstract": " The benefit of transformers in large-scale 3D point cloud perception tasks,\nsuch as 3D object detection, is limited by their quadratic computation cost\nwhen modeling long-range relationships. In contrast, linear RNNs have low\ncomputational complexity and are suitable for long-range modeling. Toward this\ngoal, we propose a simple and effective window-based framework built on LInear\ngrOup RNN (i.e., perform linear RNN for grouped features) for accurate 3D\nobject detection, called LION. The key property is to allow sufficient feature\ninteraction in a much larger group than transformer-based methods. However,\neffectively applying linear group RNN to 3D object detection in highly sparse\npoint clouds is not trivial due to its limitation in handling spatial modeling.\nTo tackle this problem, we simply introduce a 3D spatial feature descriptor and\nintegrate it into the linear group RNN operators to enhance their spatial\nfeatures rather than blindly increasing the number of scanning orders for voxel\nfeatures. To further address the challenge in highly sparse point clouds, we\npropose a 3D voxel generation strategy to densify foreground features thanks to\nlinear group RNN as a natural property of auto-regressive models. Extensive\nexperiments verify the effectiveness of the proposed components and the\ngeneralization of our LION on different linear group RNN operators including\nMamba, RWKV, and RetNet. Furthermore, it is worth mentioning that our\nLION-Mamba achieves state-of-the-art on Waymo, nuScenes, Argoverse V2, and ONCE\ndataset. Last but not least, our method supports kinds of advanced linear RNN\noperators (e.g., RetNet, RWKV, Mamba, xLSTM and TTT) on small but popular KITTI\ndataset for a quick experience with our linear RNN-based framework.\n", "date": "2024-07-25", "categories": []} {"id": "2407.16955", "title": "DVPE: Divided View Position Embedding for Multi-View 3D Object Detection", "abstract": " Sparse query-based paradigms have achieved significant success in multi-view\n3D detection for autonomous vehicles. Current research faces challenges in\nbalancing between enlarging receptive fields and reducing interference when\naggregating multi-view features. Moreover, different poses of cameras present\nchallenges in training global attention models. To address these problems, this\npaper proposes a divided view method, in which features are modeled globally\nvia the visibility crossattention mechanism, but interact only with partial\nfeatures in a divided local virtual space. This effectively reduces\ninterference from other irrelevant features and alleviates the training\ndifficulties of the transformer by decoupling the position embedding from\ncamera poses. Additionally, 2D historical RoI features are incorporated into\nthe object-centric temporal modeling to utilize highlevel visual semantic\ninformation. The model is trained using a one-to-many assignment strategy to\nfacilitate stability. Our framework, named DVPE, achieves state-of-the-art\nperformance (57.2% mAP and 64.5% NDS) on the nuScenes test set. Codes will be\navailable at https://github.com/dop0/DVPE.\n", "date": "2024-07-24", "categories": []} {"id": "2403.19580", "title": "OV-Uni3DETR: Towards Unified Open-Vocabulary 3D Object Detection via\n Cycle-Modality Propagation", "abstract": " In the current state of 3D object detection research, the severe scarcity of\nannotated 3D data, substantial disparities across different data modalities,\nand the absence of a unified architecture, have impeded the progress towards\nthe goal of universality. In this paper, we propose \\textbf{OV-Uni3DETR}, a\nunified open-vocabulary 3D detector via cycle-modality propagation. Compared\nwith existing 3D detectors, OV-Uni3DETR offers distinct advantages: 1)\nOpen-vocabulary 3D detection: During training, it leverages various accessible\ndata, especially extensive 2D detection images, to boost training diversity.\nDuring inference, it can detect both seen and unseen classes. 2) Modality\nunifying: It seamlessly accommodates input data from any given modality,\neffectively addressing scenarios involving disparate modalities or missing\nsensor information, thereby supporting test-time modality switching. 3) Scene\nunifying: It provides a unified multi-modal model architecture for diverse\nscenes collected by distinct sensors. Specifically, we propose the\ncycle-modality propagation, aimed at propagating knowledge bridging 2D and 3D\nmodalities, to support the aforementioned functionalities. 2D semantic\nknowledge from large-vocabulary learning guides novel class discovery in the 3D\ndomain, and 3D geometric knowledge provides localization supervision for 2D\ndetection images. OV-Uni3DETR achieves the state-of-the-art performance on\nvarious scenarios, surpassing existing methods by more than 6\\% on average. Its\nperformance using only RGB images is on par with or even surpasses that of\nprevious point cloud based methods. Code and pre-trained models will be\nreleased later.\n", "date": "2024-03-28", "categories": []} {"id": "2407.16448", "title": "MonoWAD: Weather-Adaptive Diffusion Model for Robust Monocular 3D Object\n Detection", "abstract": " Monocular 3D object detection is an important challenging task in autonomous\ndriving. Existing methods mainly focus on performing 3D detection in ideal\nweather conditions, characterized by scenarios with clear and optimal\nvisibility. However, the challenge of autonomous driving requires the ability\nto handle changes in weather conditions, such as foggy weather, not just clear\nweather. We introduce MonoWAD, a novel weather-robust monocular 3D object\ndetector with a weather-adaptive diffusion model. It contains two components:\n(1) the weather codebook to memorize the knowledge of the clear weather and\ngenerate a weather-reference feature for any input, and (2) the\nweather-adaptive diffusion model to enhance the feature representation of the\ninput feature by incorporating a weather-reference feature. This serves an\nattention role in indicating how much improvement is needed for the input\nfeature according to the weather conditions. To achieve this goal, we introduce\na weather-adaptive enhancement loss to enhance the feature representation under\nboth clear and foggy weather conditions. Extensive experiments under various\nweather conditions demonstrate that MonoWAD achieves weather-robust monocular\n3D object detection. The code and dataset are released at\nhttps://github.com/VisualAIKHU/MonoWAD.\n", "date": "2024-07-23", "categories": []} {"id": "2407.10876", "title": "RepVF: A Unified Vector Fields Representation for Multi-task 3D\n Perception", "abstract": " Concurrent processing of multiple autonomous driving 3D perception tasks\nwithin the same spatiotemporal scene poses a significant challenge, in\nparticular due to the computational inefficiencies and feature competition\nbetween tasks when using traditional multi-task learning approaches. This paper\naddresses these issues by proposing a novel unified representation, RepVF,\nwhich harmonizes the representation of various perception tasks such as 3D\nobject detection and 3D lane detection within a single framework. RepVF\ncharacterizes the structure of different targets in the scene through a vector\nfield, enabling a single-head, multi-task learning model that significantly\nreduces computational redundancy and feature competition. Building upon RepVF,\nwe introduce RFTR, a network designed to exploit the inherent connections\nbetween different tasks by utilizing a hierarchical structure of queries that\nimplicitly model the relationships both between and within tasks. This approach\neliminates the need for task-specific heads and parameters, fundamentally\nreducing the conflicts inherent in traditional multi-task learning paradigms.\nWe validate our approach by combining labels from the OpenLane dataset with the\nWaymo Open dataset. Our work presents a significant advancement in the\nefficiency and effectiveness of multi-task perception in autonomous driving,\noffering a new perspective on handling multiple 3D perception tasks\nsynchronously and in parallel. The code will be available at:\nhttps://github.com/jbji/RepVF\n", "date": "2024-07-15", "categories": []} {"id": "2407.15334", "title": "Explore the LiDAR-Camera Dynamic Adjustment Fusion for 3D Object\n Detection", "abstract": " Camera and LiDAR serve as informative sensors for accurate and robust\nautonomous driving systems. However, these sensors often exhibit heterogeneous\nnatures, resulting in distributional modality gaps that present significant\nchallenges for fusion. To address this, a robust fusion technique is crucial,\nparticularly for enhancing 3D object detection. In this paper, we introduce a\ndynamic adjustment technology aimed at aligning modal distributions and\nlearning effective modality representations to enhance the fusion process.\nSpecifically, we propose a triphase domain aligning module. This module adjusts\nthe feature distributions from both the camera and LiDAR, bringing them closer\nto the ground truth domain and minimizing differences. Additionally, we explore\nimproved representation acquisition methods for dynamic fusion, which includes\nmodal interaction and specialty enhancement. Finally, an adaptive learning\ntechnique that merges the semantics and geometry information for dynamical\ninstance optimization. Extensive experiments in the nuScenes dataset present\ncompetitive performance with state-of-the-art approaches. Our code will be\nreleased in the future.\n", "date": "2024-07-22", "categories": []} {"id": "2407.15354", "title": "Learning High-resolution Vector Representation from Multi-Camera Images\n for 3D Object Detection", "abstract": " The Bird's-Eye-View (BEV) representation is a critical factor that directly\nimpacts the 3D object detection performance, but the traditional BEV grid\nrepresentation induces quadratic computational cost as the spatial resolution\ngrows. To address this limitation, we present a new camera-based 3D object\ndetector with high-resolution vector representation: VectorFormer. The\npresented high-resolution vector representation is combined with the\nlower-resolution BEV representation to efficiently exploit 3D geometry from\nmulti-camera images at a high resolution through our two novel modules: vector\nscattering and gathering. To this end, the learned vector representation with\nricher scene contexts can serve as the decoding query for final predictions. We\nconduct extensive experiments on the nuScenes dataset and demonstrate\nstate-of-the-art performance in NDS and inference time. Furthermore, we\ninvestigate query-BEV-based methods incorporated with our proposed vector\nrepresentation and observe a consistent performance improvement.\n", "date": "2024-07-22", "categories": []} {"id": "2401.00912", "title": "ScatterFormer: Efficient Voxel Transformer with Scattered Linear\n Attention", "abstract": " Window-based transformers excel in large-scale point cloud understanding by\ncapturing context-aware representations with affordable attention computation\nin a more localized manner. However, the sparse nature of point clouds leads to\na significant variance in the number of voxels per window. Existing methods\ngroup the voxels in each window into fixed-length sequences through extensive\nsorting and padding operations, resulting in a non-negligible computational and\nmemory overhead. In this paper, we introduce ScatterFormer, which to the best\nof our knowledge, is the first to directly apply attention to voxels across\ndifferent windows as a single sequence. The key of ScatterFormer is a Scattered\nLinear Attention (SLA) module, which leverages the pre-computation of key-value\npairs in linear attention to enable parallel computation on the variable-length\nvoxel sequences divided by windows. Leveraging the hierarchical structure of\nGPUs and shared memory, we propose a chunk-wise algorithm that reduces the SLA\nmodule's latency to less than 1 millisecond on moderate GPUs. Furthermore, we\ndevelop a cross-window interaction module that improves the locality and\nconnectivity of voxel features across different windows, eliminating the need\nfor extensive window shifting. Our proposed ScatterFormer demonstrates 73.8 mAP\n(L2) on the Waymo Open Dataset and 72.4 NDS on the NuScenes dataset, running at\nan outstanding detection rate of 23 FPS.The code is available at\n\\href{https://github.com/skyhehe123/ScatterFormer}{https://github.com/skyhehe123/ScatterFormer}.\n", "date": "2024-01-01", "categories": []} {"id": "2407.13748", "title": "General Geometry-aware Weakly Supervised 3D Object Detection", "abstract": " 3D object detection is an indispensable component for scene understanding.\nHowever, the annotation of large-scale 3D datasets requires significant human\neffort. To tackle this problem, many methods adopt weakly supervised 3D object\ndetection that estimates 3D boxes by leveraging 2D boxes and\nscene/class-specific priors. However, these approaches generally depend on\nsophisticated manual priors, which is hard to generalize to novel categories\nand scenes. In this paper, we are motivated to propose a general approach,\nwhich can be easily adapted to new scenes and/or classes. A unified framework\nis developed for learning 3D object detectors from RGB images and associated 2D\nboxes. In specific, we propose three general components: prior injection module\nto obtain general object geometric priors from LLM model, 2D space projection\nconstraint to minimize the discrepancy between the boundaries of projected 3D\nboxes and their corresponding 2D boxes on the image plane, and 3D space\ngeometry constraint to build a Point-to-Box alignment loss to further refine\nthe pose of estimated 3D boxes. Experiments on KITTI and SUN-RGBD datasets\ndemonstrate that our method yields surprisingly high-quality 3D bounding boxes\nwith only 2D annotation. The source code is available at\nhttps://github.com/gwenzhang/GGA.\n", "date": "2024-07-18", "categories": []} {"id": "2403.10353", "title": "SimPB: A Single Model for 2D and 3D Object Detection from Multiple\n Cameras", "abstract": " The field of autonomous driving has attracted considerable interest in\napproaches that directly infer 3D objects in the Bird's Eye View (BEV) from\nmultiple cameras. Some attempts have also explored utilizing 2D detectors from\nsingle images to enhance the performance of 3D detection. However, these\napproaches rely on a two-stage process with separate detectors, where the 2D\ndetection results are utilized only once for token selection or query\ninitialization. In this paper, we present a single model termed SimPB, which\nsimultaneously detects 2D objects in the perspective view and 3D objects in the\nBEV space from multiple cameras. To achieve this, we introduce a hybrid decoder\nconsisting of several multi-view 2D decoder layers and several 3D decoder\nlayers, specifically designed for their respective detection tasks. A Dynamic\nQuery Allocation module and an Adaptive Query Aggregation module are proposed\nto continuously update and refine the interaction between 2D and 3D results, in\na cyclic 3D-2D-3D manner. Additionally, Query-group Attention is utilized to\nstrengthen the interaction among 2D queries within each camera group. In the\nexperiments, we evaluate our method on the nuScenes dataset and demonstrate\npromising results for both 2D and 3D detection tasks. Our code is available at:\nhttps://github.com/nullmax-vision/SimPB.\n", "date": "2024-03-15", "categories": []} {"id": "2407.05256", "title": "Unlocking Textual and Visual Wisdom: Open-Vocabulary 3D Object Detection\n Enhanced by Comprehensive Guidance from Text and Image", "abstract": " Open-vocabulary 3D object detection (OV-3DDet) aims to localize and recognize\nboth seen and previously unseen object categories within any new 3D scene.\nWhile language and vision foundation models have achieved success in handling\nvarious open-vocabulary tasks with abundant training data, OV-3DDet faces a\nsignificant challenge due to the limited availability of training data.\nAlthough some pioneering efforts have integrated vision-language models (VLM)\nknowledge into OV-3DDet learning, the full potential of these foundational\nmodels has yet to be fully exploited. In this paper, we unlock the textual and\nvisual wisdom to tackle the open-vocabulary 3D detection task by leveraging the\nlanguage and vision foundation models. We leverage a vision foundation model to\nprovide image-wise guidance for discovering novel classes in 3D scenes.\nSpecifically, we utilize a object detection vision foundation model to enable\nthe zero-shot discovery of objects in images, which serves as the initial seeds\nand filtering guidance to identify novel 3D objects. Additionally, to align the\n3D space with the powerful vision-language space, we introduce a hierarchical\nalignment approach, where the 3D feature space is aligned with the\nvision-language feature space using a pre-trained VLM at the instance,\ncategory, and scene levels. Through extensive experimentation, we demonstrate\nsignificant improvements in accuracy and generalization, highlighting the\npotential of foundation models in advancing open-vocabulary 3D object detection\nin real-world scenarios.\n", "date": "2024-07-07", "categories": []} {"id": "2407.11382", "title": "Segment, Lift and Fit: Automatic 3D Shape Labeling from 2D Prompts", "abstract": " This paper proposes an algorithm for automatically labeling 3D objects from\n2D point or box prompts, especially focusing on applications in autonomous\ndriving. Unlike previous arts, our auto-labeler predicts 3D shapes instead of\nbounding boxes and does not require training on a specific dataset. We propose\na Segment, Lift, and Fit (SLF) paradigm to achieve this goal. Firstly, we\nsegment high-quality instance masks from the prompts using the Segment Anything\nModel (SAM) and transform the remaining problem into predicting 3D shapes from\ngiven 2D masks. Due to the ill-posed nature of this problem, it presents a\nsignificant challenge as multiple 3D shapes can project into an identical mask.\nTo tackle this issue, we then lift 2D masks to 3D forms and employ gradient\ndescent to adjust their poses and shapes until the projections fit the masks\nand the surfaces conform to surrounding LiDAR points. Notably, since we do not\ntrain on a specific dataset, the SLF auto-labeler does not overfit to biased\nannotation patterns in the training set as other methods do. Thus, the\ngeneralization ability across different datasets improves. Experimental results\non the KITTI dataset demonstrate that the SLF auto-labeler produces\nhigh-quality bounding box annotations, achieving an AP@0.5 IoU of nearly 90\\%.\nDetectors trained with the generated pseudo-labels perform nearly as well as\nthose trained with actual ground-truth annotations. Furthermore, the SLF\nauto-labeler shows promising results in detailed shape predictions, providing a\npotential alternative for the occupancy annotation of dynamic objects.\n", "date": "2024-07-16", "categories": []} {"id": "2307.13510", "title": "HeightFormer: Explicit Height Modeling without Extra Data for\n Camera-only 3D Object Detection in Bird's Eye View", "abstract": " Vision-based Bird's Eye View (BEV) representation is an emerging perception\nformulation for autonomous driving. The core challenge is to construct BEV\nspace with multi-camera features, which is a one-to-many ill-posed problem.\nDiving into all previous BEV representation generation methods, we found that\nmost of them fall into two types: modeling depths in image views or modeling\nheights in the BEV space, mostly in an implicit way. In this work, we propose\nto explicitly model heights in the BEV space, which needs no extra data like\nLiDAR and can fit arbitrary camera rigs and types compared to modeling depths.\nTheoretically, we give proof of the equivalence between height-based methods\nand depth-based methods. Considering the equivalence and some advantages of\nmodeling heights, we propose HeightFormer, which models heights and\nuncertainties in a self-recursive way. Without any extra data, the proposed\nHeightFormer could estimate heights in BEV accurately. Benchmark results show\nthat the performance of HeightFormer achieves SOTA compared with those\ncamera-only methods.\n", "date": "2023-07-25", "categories": []} {"id": "2312.12098", "title": "Rethinking LiDAR Domain Generalization: Single Source as Multiple\n Density Domains", "abstract": " In the realm of LiDAR-based perception, significant strides have been made,\nyet domain generalization remains a substantial challenge. The performance\noften deteriorates when models are applied to unfamiliar datasets with\ndifferent LiDAR sensors or deployed in new environments, primarily due to\nvariations in point cloud density distributions. To tackle this challenge, we\npropose a Density Discriminative Feature Embedding (DDFE) module, capitalizing\non the observation that a single source LiDAR point cloud encompasses a\nspectrum of densities. The DDFE module is meticulously designed to extract\ndensity-specific features within a single source domain, facilitating the\nrecognition of objects sharing similar density characteristics across different\nLiDAR sensors. In addition, we introduce a simple yet effective density\naugmentation technique aimed at expanding the spectrum of density in source\ndata, thereby enhancing the capabilities of the DDFE. Our DDFE stands out as a\nversatile and lightweight domain generalization module. It can be seamlessly\nintegrated into various 3D backbone networks, where it has demonstrated\nsuperior performance over current state-of-the-art domain generalization\nmethods. Code is available at https://github.com/dgist-cvlab/MultiDensityDG.\n", "date": "2023-12-19", "categories": []} {"id": "2404.05641", "title": "3D-COCO: extension of MS-COCO dataset for image detection and 3D\n reconstruction modules", "abstract": " We introduce 3D-COCO, an extension of the original MS-COCO dataset providing\n3D models and 2D-3D alignment annotations. 3D-COCO was designed to achieve\ncomputer vision tasks such as 3D reconstruction or image detection configurable\nwith textual, 2D image, and 3D CAD model queries. We complete the existing\nMS-COCO dataset with 28K 3D models collected on ShapeNet and Objaverse. By\nusing an IoU-based method, we match each MS-COCO annotation with the best 3D\nmodels to provide a 2D-3D alignment. The open-source nature of 3D-COCO is a\npremiere that should pave the way for new research on 3D-related topics. The\ndataset and its source codes is available at\nhttps://kalisteo.cea.fr/index.php/coco3d-object-detection-and-reconstruction/\n", "date": "2024-04-08", "categories": []} {"id": "2305.03716", "title": "3D Small Object Detection with Dynamic Spatial Pruning", "abstract": " In this paper, we propose an efficient feature pruning strategy for 3D small\nobject detection. Conventional 3D object detection methods struggle on small\nobjects due to the weak geometric information from a small number of points.\nAlthough increasing the spatial resolution of feature representations can\nimprove the detection performance on small objects, the additional\ncomputational overhead is unaffordable. With in-depth study, we observe the\ngrowth of computation mainly comes from the upsampling operation in the decoder\nof 3D detector. Motivated by this, we present a multi-level 3D detector named\nDSPDet3D which benefits from high spatial resolution to achieves high accuracy\non small object detection, while reducing redundant computation by only\nfocusing on small object areas. Specifically, we theoretically derive a dynamic\nspatial pruning (DSP) strategy to prune the redundant spatial representation of\n3D scene in a cascade manner according to the distribution of objects. Then we\ndesign DSP module following this strategy and construct DSPDet3D with this\nefficient module. On ScanNet and TO-SCENE dataset, our method achieves leading\nperformance on small object detection. Moreover, DSPDet3D trained with only\nScanNet rooms can generalize well to scenes in larger scale. It takes less than\n2s to directly process a whole building consisting of more than 4500k points\nwhile detecting out almost all objects, ranging from cups to beds, on a single\nRTX 3090 GPU. Project page: https://xuxw98.github.io/DSPDet3D/.\n", "date": "2023-05-05", "categories": []} {"id": "2407.09787", "title": "Semi-supervised 3D Object Detection with PatchTeacher and PillarMix", "abstract": " Semi-supervised learning aims to leverage numerous unlabeled data to improve\nthe model performance. Current semi-supervised 3D object detection methods\ntypically use a teacher to generate pseudo labels for a student, and the\nquality of the pseudo labels is essential for the final performance. In this\npaper, we propose PatchTeacher, which focuses on partial scene 3D object\ndetection to provide high-quality pseudo labels for the student. Specifically,\nwe divide a complete scene into a series of patches and feed them to our\nPatchTeacher sequentially. PatchTeacher leverages the low memory consumption\nadvantage of partial scene detection to process point clouds with a\nhigh-resolution voxelization, which can minimize the information loss of\nquantization and extract more fine-grained features. However, it is non-trivial\nto train a detector on fractions of the scene. Therefore, we introduce three\nkey techniques, i.e., Patch Normalizer, Quadrant Align, and Fovea Selection, to\nimprove the performance of PatchTeacher. Moreover, we devise PillarMix, a\nstrong data augmentation strategy that mixes truncated pillars from different\nLiDAR scans to generate diverse training samples and thus help the model learn\nmore general representation. Extensive experiments conducted on Waymo and ONCE\ndatasets verify the effectiveness and superiority of our method and we achieve\nnew state-of-the-art results, surpassing existing methods by a large margin.\nCodes are available at https://github.com/LittlePey/PTPM.\n", "date": "2024-07-13", "categories": []} {"id": "2407.10135", "title": "FSD-BEV: Foreground Self-Distillation for Multi-view 3D Object Detection", "abstract": " Although multi-view 3D object detection based on the Bird's-Eye-View (BEV)\nparadigm has garnered widespread attention as an economical and\ndeployment-friendly perception solution for autonomous driving, there is still\na performance gap compared to LiDAR-based methods. In recent years, several\ncross-modal distillation methods have been proposed to transfer beneficial\ninformation from teacher models to student models, with the aim of enhancing\nperformance. However, these methods face challenges due to discrepancies in\nfeature distribution originating from different data modalities and network\nstructures, making knowledge transfer exceptionally challenging. In this paper,\nwe propose a Foreground Self-Distillation (FSD) scheme that effectively avoids\nthe issue of distribution discrepancies, maintaining remarkable distillation\neffects without the need for pre-trained teacher models or cumbersome\ndistillation strategies. Additionally, we design two Point Cloud\nIntensification (PCI) strategies to compensate for the sparsity of point clouds\nby frame combination and pseudo point assignment. Finally, we develop a\nMulti-Scale Foreground Enhancement (MSFE) module to extract and fuse\nmulti-scale foreground features by predicted elliptical Gaussian heatmap,\nfurther improving the model's performance. We integrate all the above\ninnovations into a unified framework named FSD-BEV. Extensive experiments on\nthe nuScenes dataset exhibit that FSD-BEV achieves state-of-the-art\nperformance, highlighting its effectiveness. The code and models are available\nat: https://github.com/CocoBoom/fsd-bev.\n", "date": "2024-07-14", "categories": []} {"id": "2407.10164", "title": "LabelDistill: Label-guided Cross-modal Knowledge Distillation for\n Camera-based 3D Object Detection", "abstract": " Recent advancements in camera-based 3D object detection have introduced\ncross-modal knowledge distillation to bridge the performance gap with LiDAR 3D\ndetectors, leveraging the precise geometric information in LiDAR point clouds.\nHowever, existing cross-modal knowledge distillation methods tend to overlook\nthe inherent imperfections of LiDAR, such as the ambiguity of measurements on\ndistant or occluded objects, which should not be transferred to the image\ndetector. To mitigate these imperfections in LiDAR teacher, we propose a novel\nmethod that leverages aleatoric uncertainty-free features from ground truth\nlabels. In contrast to conventional label guidance approaches, we approximate\nthe inverse function of the teacher's head to effectively embed label inputs\ninto feature space. This approach provides additional accurate guidance\nalongside LiDAR teacher, thereby boosting the performance of the image\ndetector. Additionally, we introduce feature partitioning, which effectively\ntransfers knowledge from the teacher modality while preserving the distinctive\nfeatures of the student, thereby maximizing the potential of both modalities.\nExperimental results demonstrate that our approach improves mAP and NDS by 5.1\npoints and 4.9 points compared to the baseline model, proving the effectiveness\nof our approach. The code is available at\nhttps://github.com/sanmin0312/LabelDistill\n", "date": "2024-07-14", "categories": []} {"id": "2407.10749", "title": "SEED: A Simple and Effective 3D DETR in Point Clouds", "abstract": " Recently, detection transformers (DETRs) have gradually taken a dominant\nposition in 2D detection thanks to their elegant framework. However, DETR-based\ndetectors for 3D point clouds are still difficult to achieve satisfactory\nperformance. We argue that the main challenges are twofold: 1) How to obtain\nthe appropriate object queries is challenging due to the high sparsity and\nuneven distribution of point clouds; 2) How to implement an effective query\ninteraction by exploiting the rich geometric structure of point clouds is not\nfully explored. To this end, we propose a simple and effective 3D DETR method\n(SEED) for detecting 3D objects from point clouds, which involves a dual query\nselection (DQS) module and a deformable grid attention (DGA) module. More\nconcretely, to obtain appropriate queries, DQS first ensures a high recall to\nretain a large number of queries by the predicted confidence scores and then\nfurther picks out high-quality queries according to the estimated quality\nscores. DGA uniformly divides each reference box into grids as the reference\npoints and then utilizes the predicted offsets to achieve a flexible receptive\nfield, allowing the network to focus on relevant regions and capture more\ninformative features. Extensive ablation studies on DQS and DGA demonstrate its\neffectiveness. Furthermore, our SEED achieves state-of-the-art detection\nperformance on both the large-scale Waymo and nuScenes datasets, illustrating\nthe superiority of our proposed method. The code is available at\nhttps://github.com/happinesslz/SEED\n", "date": "2024-07-15", "categories": []} {"id": "2407.10753", "title": "OPEN: Object-wise Position Embedding for Multi-view 3D Object Detection", "abstract": " Accurate depth information is crucial for enhancing the performance of\nmulti-view 3D object detection. Despite the success of some existing multi-view\n3D detectors utilizing pixel-wise depth supervision, they overlook two\nsignificant phenomena: 1) the depth supervision obtained from LiDAR points is\nusually distributed on the surface of the object, which is not so friendly to\nexisting DETR-based 3D detectors due to the lack of the depth of 3D object\ncenter; 2) for distant objects, fine-grained depth estimation of the whole\nobject is more challenging. Therefore, we argue that the object-wise depth (or\n3D center of the object) is essential for accurate detection. In this paper, we\npropose a new multi-view 3D object detector named OPEN, whose main idea is to\neffectively inject object-wise depth information into the network through our\nproposed object-wise position embedding. Specifically, we first employ an\nobject-wise depth encoder, which takes the pixel-wise depth map as a prior, to\naccurately estimate the object-wise depth. Then, we utilize the proposed\nobject-wise position embedding to encode the object-wise depth information into\nthe transformer decoder, thereby producing 3D object-aware features for final\ndetection. Extensive experiments verify the effectiveness of our proposed\nmethod. Furthermore, OPEN achieves a new state-of-the-art performance with\n64.4% NDS and 56.7% mAP on the nuScenes test benchmark.\n", "date": "2024-07-15", "categories": []} {"id": "2403.13556", "title": "Find n' Propagate: Open-Vocabulary 3D Object Detection in Urban\n Environments", "abstract": " In this work, we tackle the limitations of current LiDAR-based 3D object\ndetection systems, which are hindered by a restricted class vocabulary and the\nhigh costs associated with annotating new object classes. Our exploration of\nopen-vocabulary (OV) learning in urban environments aims to capture novel\ninstances using pre-trained vision-language models (VLMs) with multi-sensor\ndata. We design and benchmark a set of four potential solutions as baselines,\ncategorizing them into either top-down or bottom-up approaches based on their\ninput data strategies. While effective, these methods exhibit certain\nlimitations, such as missing novel objects in 3D box estimation or applying\nrigorous priors, leading to biases towards objects near the camera or of\nrectangular geometries. To overcome these limitations, we introduce a universal\n\\textsc{Find n' Propagate} approach for 3D OV tasks, aimed at maximizing the\nrecall of novel objects and propagating this detection capability to more\ndistant areas thereby progressively capturing more. In particular, we utilize a\ngreedy box seeker to search against 3D novel boxes of varying orientations and\ndepth in each generated frustum and ensure the reliability of newly identified\nboxes by cross alignment and density ranker. Additionally, the inherent bias\ntowards camera-proximal objects is alleviated by the proposed remote simulator,\nwhich randomly diversifies pseudo-labeled novel instances in the self-training\nprocess, combined with the fusion of base samples in the memory bank. Extensive\nexperiments demonstrate a 53% improvement in novel recall across diverse OV\nsettings, VLMs, and 3D detectors. Notably, we achieve up to a 3.97-fold\nincrease in Average Precision (AP) for novel object classes. The source code is\nmade available at https://github.com/djamahl99/findnpropagate.\n", "date": "2024-03-20", "categories": []} {"id": "2407.04061", "title": "Detect Closer Surfaces that can be Seen: New Modeling and Evaluation in\n Cross-domain 3D Object Detection", "abstract": " The performance of domain adaptation technologies has not yet reached an\nideal level in the current 3D object detection field for autonomous driving,\nwhich is mainly due to significant differences in the size of vehicles, as well\nas the environments they operate in when applied across domains. These factors\ntogether hinder the effective transfer and application of knowledge learned\nfrom specific datasets. Since the existing evaluation metrics are initially\ndesigned for evaluation on a single domain by calculating the 2D or 3D overlap\nbetween the prediction and ground-truth bounding boxes, they often suffer from\nthe overfitting problem caused by the size differences among datasets. This\nraises a fundamental question related to the evaluation of the 3D object\ndetection models' cross-domain performance: Do we really need models to\nmaintain excellent performance in their original 3D bounding boxes after being\napplied across domains? From a practical application perspective, one of our\nmain focuses is actually on preventing collisions between vehicles and other\nobstacles, especially in cross-domain scenarios where correctly predicting the\nsize of vehicles is much more difficult. In other words, as long as a model can\naccurately identify the closest surfaces to the ego vehicle, it is sufficient\nto effectively avoid obstacles. In this paper, we propose two metrics to\nmeasure 3D object detection models' ability of detecting the closer surfaces to\nthe sensor on the ego vehicle, which can be used to evaluate their cross-domain\nperformance more comprehensively and reasonably. Furthermore, we propose a\nrefinement head, named EdgeHead, to guide models to focus more on the learnable\ncloser surfaces, which can greatly improve the cross-domain performance of\nexisting models not only under our new metrics, but even also under the\noriginal BEV/3D metrics.\n", "date": "2024-07-04", "categories": []} {"id": "2407.08931", "title": "Global-Local Collaborative Inference with LLM for Lidar-Based\n Open-Vocabulary Detection", "abstract": " Open-Vocabulary Detection (OVD) is the task of detecting all interesting\nobjects in a given scene without predefined object classes. Extensive work has\nbeen done to deal with the OVD for 2D RGB images, but the exploration of 3D OVD\nis still limited. Intuitively, lidar point clouds provide 3D information, both\nobject level and scene level, to generate trustful detection results. However,\nprevious lidar-based OVD methods only focus on the usage of object-level\nfeatures, ignoring the essence of scene-level information. In this paper, we\npropose a Global-Local Collaborative Scheme (GLIS) for the lidar-based OVD\ntask, which contains a local branch to generate object-level detection result\nand a global branch to obtain scene-level global feature. With the global-local\ninformation, a Large Language Model (LLM) is applied for chain-of-thought\ninference, and the detection result can be refined accordingly. We further\npropose Reflected Pseudo Labels Generation (RPLG) to generate high-quality\npseudo labels for supervision and Background-Aware Object Localization (BAOL)\nto select precise object proposals. Extensive experiments on ScanNetV2 and SUN\nRGB-D demonstrate the superiority of our methods. Code is released at\nhttps://github.com/GradiusTwinbee/GLIS.\n", "date": "2024-07-12", "categories": []} {"id": "2407.08569", "title": "Approaching Outside: Scaling Unsupervised 3D Object Detection from 2D\n Scene", "abstract": " The unsupervised 3D object detection is to accurately detect objects in\nunstructured environments with no explicit supervisory signals. This task,\ngiven sparse LiDAR point clouds, often results in compromised performance for\ndetecting distant or small objects due to the inherent sparsity and limited\nspatial resolution. In this paper, we are among the early attempts to integrate\nLiDAR data with 2D images for unsupervised 3D detection and introduce a new\nmethod, dubbed LiDAR-2D Self-paced Learning (LiSe). We argue that RGB images\nserve as a valuable complement to LiDAR data, offering precise 2D localization\ncues, particularly when scarce LiDAR points are available for certain objects.\nConsidering the unique characteristics of both modalities, our framework\ndevises a self-paced learning pipeline that incorporates adaptive sampling and\nweak model aggregation strategies. The adaptive sampling strategy dynamically\ntunes the distribution of pseudo labels during training, countering the\ntendency of models to overfit easily detected samples, such as nearby and\nlarge-sized objects. By doing so, it ensures a balanced learning trajectory\nacross varying object scales and distances. The weak model aggregation\ncomponent consolidates the strengths of models trained under different pseudo\nlabel distributions, culminating in a robust and powerful final model.\nExperimental evaluations validate the efficacy of our proposed LiSe method,\nmanifesting significant improvements of +7.1% AP$_{BEV}$ and +3.4% AP$_{3D}$ on\nnuScenes, and +8.3% AP$_{BEV}$ and +7.4% AP$_{3D}$ on Lyft compared to existing\ntechniques.\n", "date": "2024-07-11", "categories": []} {"id": "2403.07284", "title": "SparseLIF: High-Performance Sparse LiDAR-Camera Fusion for 3D Object\n Detection", "abstract": " Sparse 3D detectors have received significant attention since the query-based\nparadigm embraces low latency without explicit dense BEV feature construction.\nHowever, these detectors achieve worse performance than their dense\ncounterparts. In this paper, we find the key to bridging the performance gap is\nto enhance the awareness of rich representations in two modalities. Here, we\npresent a high-performance fully sparse detector for end-to-end multi-modality\n3D object detection. The detector, termed SparseLIF, contains three key\ndesigns, which are (1) Perspective-Aware Query Generation (PAQG) to generate\nhigh-quality 3D queries with perspective priors, (2) RoI-Aware Sampling (RIAS)\nto further refine prior queries by sampling RoI features from each modality,\n(3) Uncertainty-Aware Fusion (UAF) to precisely quantify the uncertainty of\neach sensor modality and adaptively conduct final multi-modality fusion, thus\nachieving great robustness against sensor noises. By the time of paper\nsubmission, SparseLIF achieves state-of-the-art performance on the nuScenes\ndataset, ranking 1st on both validation set and test benchmark, outperforming\nall state-of-the-art 3D object detectors by a notable margin.\n", "date": "2024-03-12", "categories": []} {"id": "2301.12511", "title": "Fast-BEV: A Fast and Strong Bird's-Eye View Perception Baseline", "abstract": " Recently, perception task based on Bird's-Eye View (BEV) representation has\ndrawn more and more attention, and BEV representation is promising as the\nfoundation for next-generation Autonomous Vehicle (AV) perception. However,\nmost existing BEV solutions either require considerable resources to execute\non-vehicle inference or suffer from modest performance. This paper proposes a\nsimple yet effective framework, termed Fast-BEV , which is capable of\nperforming faster BEV perception on the on-vehicle chips. Towards this goal, we\nfirst empirically find that the BEV representation can be sufficiently powerful\nwithout expensive transformer based transformation nor depth representation.\nOur Fast-BEV consists of five parts, We novelly propose (1) a lightweight\ndeployment-friendly view transformation which fast transfers 2D image feature\nto 3D voxel space, (2) an multi-scale image encoder which leverages multi-scale\ninformation for better performance, (3) an efficient BEV encoder which is\nparticularly designed to speed up on-vehicle inference. We further introduce\n(4) a strong data augmentation strategy for both image and BEV space to avoid\nover-fitting, (5) a multi-frame feature fusion mechanism to leverage the\ntemporal information. Through experiments, on 2080Ti platform, our R50 model\ncan run 52.6 FPS with 47.3% NDS on the nuScenes validation set, exceeding the\n41.3 FPS and 47.5% NDS of the BEVDepth-R50 model and 30.2 FPS and 45.7% NDS of\nthe BEVDet4D-R50 model. Our largest model (R101@900x1600) establishes a\ncompetitive 53.5% NDS on the nuScenes validation set. We further develop a\nbenchmark with considerable accuracy and efficiency on current popular\non-vehicle chips. The code is released at:\nhttps://github.com/Sense-GVT/Fast-BEV.\n", "date": "2023-01-29", "categories": []} {"id": "2207.02466", "title": "GLENet: Boosting 3D Object Detectors with Generative Label Uncertainty\n Estimation", "abstract": " The inherent ambiguity in ground-truth annotations of 3D bounding boxes,\ncaused by occlusions, signal missing, or manual annotation errors, can confuse\ndeep 3D object detectors during training, thus deteriorating detection\naccuracy. However, existing methods overlook such issues to some extent and\ntreat the labels as deterministic. In this paper, we formulate the label\nuncertainty problem as the diversity of potentially plausible bounding boxes of\nobjects. Then, we propose GLENet, a generative framework adapted from\nconditional variational autoencoders, to model the one-to-many relationship\nbetween a typical 3D object and its potential ground-truth bounding boxes with\nlatent variables. The label uncertainty generated by GLENet is a plug-and-play\nmodule and can be conveniently integrated into existing deep 3D detectors to\nbuild probabilistic detectors and supervise the learning of the localization\nuncertainty. Besides, we propose an uncertainty-aware quality estimator\narchitecture in probabilistic detectors to guide the training of the IoU-branch\nwith predicted localization uncertainty. We incorporate the proposed methods\ninto various popular base 3D detectors and demonstrate significant and\nconsistent performance gains on both KITTI and Waymo benchmark datasets.\nEspecially, the proposed GLENet-VR outperforms all published LiDAR-based\napproaches by a large margin and achieves the top rank among single-modal\nmethods on the challenging KITTI test set. The source code and pre-trained\nmodels are publicly available at \\url{https://github.com/Eaphan/GLENet}.\n", "date": "2022-07-06", "categories": []} {"id": "2304.01838", "title": "BugNIST -- a Large Volumetric Dataset for Object Detection under Domain\n Shift", "abstract": " Domain shift significantly influences the performance of deep learning\nalgorithms, particularly for object detection within volumetric 3D images.\nAnnotated training data is essential for deep learning-based object detection.\nHowever, annotating densely packed objects is time-consuming and costly.\nInstead, we suggest training models on individually scanned objects, causing a\ndomain shift between training and detection data. To address this challenge, we\nintroduce the BugNIST dataset, comprising 9154 micro-CT volumes of 12 bug types\nand 388 volumes of tightly packed bug mixtures. This dataset is characterized\nby having objects with the same appearance in the source and target domains,\nwhich is uncommon for other benchmark datasets for domain shift. During\ntraining, individual bug volumes labeled by class are utilized, while testing\nemploys mixtures with center point annotations and bug type labels. Together\nwith the dataset, we provide a baseline detection analysis, with the aim of\nadvancing the field of 3D object detection methods.\n", "date": "2023-04-04", "categories": []} {"id": "2403.06501", "title": "SeSame: Simple, Easy 3D Object Detection with Point-Wise Semantics", "abstract": " In autonomous driving, 3D object detection provides more precise information\nfor downstream tasks, including path planning and motion estimation, compared\nto 2D object detection. In this paper, we propose SeSame: a method aimed at\nenhancing semantic information in existing LiDAR-only based 3D object\ndetection. This addresses the limitation of existing 3D detectors, which\nprimarily focus on object presence and classification, thus lacking in\ncapturing relationships between elemental units that constitute the data, akin\nto semantic segmentation. Experiments demonstrate the effectiveness of our\nmethod with performance improvements on the KITTI object detection benchmark.\nOur code is available at https://github.com/HAMA-DL-dev/SeSame\n", "date": "2024-03-11", "categories": []} {"id": "2405.09883", "title": "RoScenes: A Large-scale Multi-view 3D Dataset for Roadside Perception", "abstract": " We introduce RoScenes, the largest multi-view roadside perception dataset,\nwhich aims to shed light on the development of vision-centric Bird's Eye View\n(BEV) approaches for more challenging traffic scenes. The highlights of\nRoScenes include significantly large perception area, full scene coverage and\ncrowded traffic. More specifically, our dataset achieves surprising 21.13M 3D\nannotations within 64,000 $m^2$. To relieve the expensive costs of roadside 3D\nlabeling, we present a novel BEV-to-3D joint annotation pipeline to efficiently\ncollect such a large volume of data. After that, we organize a comprehensive\nstudy for current BEV methods on RoScenes in terms of effectiveness and\nefficiency. Tested methods suffer from the vast perception area and variation\nof sensor layout across scenes, resulting in performance levels falling below\nexpectations. To this end, we propose RoBEV that incorporates feature-guided\nposition embedding for effective 2D-3D feature assignment. With its help, our\nmethod outperforms state-of-the-art by a large margin without extra\ncomputational overhead on validation set. Our dataset and devkit will be made\navailable at https://github.com/xiaosu-zhu/RoScenes.\n", "date": "2024-05-16", "categories": []} {"id": "2407.04305", "title": "Towards Stable 3D Object Detection", "abstract": " In autonomous driving, the temporal stability of 3D object detection greatly\nimpacts the driving safety. However, the detection stability cannot be accessed\nby existing metrics such as mAP and MOTA, and consequently is less explored by\nthe community. To bridge this gap, this work proposes Stability Index (SI), a\nnew metric that can comprehensively evaluate the stability of 3D detectors in\nterms of confidence, box localization, extent, and heading. By benchmarking\nstate-of-the-art object detectors on the Waymo Open Dataset, SI reveals\ninteresting properties of object stability that have not been previously\ndiscovered by other metrics. To help models improve their stability, we further\nintroduce a general and effective training strategy, called Prediction\nConsistency Learning (PCL). PCL essentially encourages the prediction\nconsistency of the same objects under different timestamps and augmentations,\nleading to enhanced detection stability. Furthermore, we examine the\neffectiveness of PCL with the widely-used CenterPoint, and achieve a remarkable\nSI of 86.00 for vehicle class, surpassing the baseline by 5.48. We hope our\nwork could serve as a reliable baseline and draw the community's attention to\nthis crucial issue in 3D object detection. Codes will be made publicly\navailable.\n", "date": "2024-07-05", "categories": []} {"id": "2407.03240", "title": "Cyclic Refiner: Object-Aware Temporal Representation Learning for\n Multi-View 3D Detection and Tracking", "abstract": " We propose a unified object-aware temporal learning framework for multi-view\n3D detection and tracking tasks. Having observed that the efficacy of the\ntemporal fusion strategy in recent multi-view perception methods may be\nweakened by distractors and background clutters in historical frames, we\npropose a cyclic learning mechanism to improve the robustness of multi-view\nrepresentation learning. The essence is constructing a backward bridge to\npropagate information from model predictions (e.g., object locations and sizes)\nto image and BEV features, which forms a circle with regular inference. After\nbackward refinement, the responses of target-irrelevant regions in historical\nframes would be suppressed, decreasing the risk of polluting future frames and\nimproving the object awareness ability of temporal fusion. We further tailor an\nobject-aware association strategy for tracking based on the cyclic learning\nmodel. The cyclic learning model not only provides refined features, but also\ndelivers finer clues (e.g., scale level) for tracklet association. The proposed\ncycle learning method and association module together contribute a novel and\nunified multi-task framework. Experiments on nuScenes show that the proposed\nmodel achieves consistent performance gains over baselines of different designs\n(i.e., dense query-based BEVFormer, sparse query-based SparseBEV and LSS-based\nBEVDet4D) on both detection and tracking evaluation.\n", "date": "2024-07-03", "categories": []} {"id": "2407.02098", "title": "DM3D: Distortion-Minimized Weight Pruning for Lossless 3D Object\n Detection", "abstract": " Applying deep neural networks to 3D point cloud processing has attracted\nincreasing attention due to its advanced performance in many areas, such as\nAR/VR, autonomous driving, and robotics. However, as neural network models and\n3D point clouds expand in size, it becomes a crucial challenge to reduce the\ncomputational and memory overhead to meet latency and energy constraints in\nreal-world applications. Although existing approaches have proposed to reduce\nboth computational cost and memory footprint, most of them only address the\nspatial redundancy in inputs, i.e. removing the redundancy of background points\nin 3D data. In this paper, we propose a novel post-training weight pruning\nscheme for 3D object detection that is (1) orthogonal to all existing point\ncloud sparsifying methods, which determines redundant parameters in the\npretrained model that lead to minimal distortion in both locality and\nconfidence (detection distortion); and (2) a universal plug-and-play pruning\nframework that works with arbitrary 3D detection model. This framework aims to\nminimize detection distortion of network output to maximally maintain detection\nprecision, by identifying layer-wise sparsity based on second-order Taylor\napproximation of the distortion. Albeit utilizing second-order information, we\nintroduced a lightweight scheme to efficiently acquire Hessian information, and\nsubsequently perform dynamic programming to solve the layer-wise sparsity.\nExtensive experiments on KITTI, Nuscenes and ONCE datasets demonstrate that our\napproach is able to maintain and even boost the detection precision on pruned\nmodel under noticeable computation reduction (FLOPs). Noticeably, we achieve\nover 3.89x, 3.72x FLOPs reduction on CenterPoint and PVRCNN model,\nrespectively, without mAP decrease, significantly improving the\nstate-of-the-art.\n", "date": "2024-07-02", "categories": []} {"id": "2407.01239", "title": "SGCCNet: Single-Stage 3D Object Detector With Saliency-Guided Data\n Augmentation and Confidence Correction Mechanism", "abstract": " The single-stage point-based 3D object detectors have attracted widespread\nresearch interest due to their advantages of lightweight and fast inference\nspeed. However, they still face challenges such as inadequate learning of\nlow-quality objects (ILQ) and misalignment between localization accuracy and\nclassification confidence (MLC). In this paper, we propose SGCCNet to alleviate\nthese two issues. For ILQ, SGCCNet adopts a Saliency-Guided Data Augmentation\n(SGDA) strategy to enhance the robustness of the model on low-quality objects\nby reducing its reliance on salient features. Specifically, We construct a\nclassification task and then approximate the saliency scores of points by\nmoving points towards the point cloud centroid in a differentiable process.\nDuring the training process, SGCCNet will be forced to learn from low saliency\nfeatures through dropping points. Meanwhile, to avoid internal covariate shift\nand contextual features forgetting caused by dropping points, we add a\ngeometric normalization module and skip connection block in each stage. For\nMLC, we design a Confidence Correction Mechanism (CCM) specifically for\npoint-based multi-class detectors. This mechanism corrects the confidence of\nthe current proposal by utilizing the predictions of other key points within\nthe local region in the post-processing stage. Extensive experiments on the\nKITTI dataset demonstrate the generality and effectiveness of our SGCCNet. On\nthe KITTI \\textit{test} set, SGCCNet achieves $80.82\\%$ for the metric of\n$AP_{3D}$ on the \\textit{Moderate} level, outperforming all other point-based\ndetectors, surpassing IA-SSD and Fast Point R-CNN by $2.35\\%$ and $3.42\\%$,\nrespectively. Additionally, SGCCNet demonstrates excellent portability for\nother point-based detectors\n", "date": "2024-07-01", "categories": []} {"id": "2406.19362", "title": "STAL3D: Unsupervised Domain Adaptation for 3D Object Detection via\n Collaborating Self-Training and Adversarial Learning", "abstract": " Existing 3D object detection suffers from expensive annotation costs and poor\ntransferability to unknown data due to the domain gap, Unsupervised Domain\nAdaptation (UDA) aims to generalize detection models trained in labeled source\ndomains to perform robustly on unexplored target domains, providing a promising\nsolution for cross-domain 3D object detection. Although Self-Training (ST)\nbased cross-domain 3D detection methods with the assistance of pseudo-labeling\ntechniques have achieved remarkable progress, they still face the issue of\nlow-quality pseudo-labels when there are significant domain disparities due to\nthe absence of a process for feature distribution alignment. While Adversarial\nLearning (AL) based methods can effectively align the feature distributions of\nthe source and target domains, the inability to obtain labels in the target\ndomain forces the adoption of asymmetric optimization losses, resulting in a\nchallenging issue of source domain bias. To overcome these limitations, we\npropose a novel unsupervised domain adaptation framework for 3D object\ndetection via collaborating ST and AL, dubbed as STAL3D, unleashing the\ncomplementary advantages of pseudo labels and feature distribution alignment.\nAdditionally, a Background Suppression Adversarial Learning (BS-AL) module and\na Scale Filtering Module (SFM) are designed tailored for 3D cross-domain\nscenes, effectively alleviating the issues of the large proportion of\nbackground interference and source domain size bias. Our STAL3D achieves\nstate-of-the-art performance on multiple cross-domain tasks and even surpasses\nthe Oracle results on Waymo $\\rightarrow$ KITTI and Waymo $\\rightarrow$\nKITTI-rain.\n", "date": "2024-06-27", "categories": []} {"id": "2404.16493", "title": "Commonsense Prototype for Outdoor Unsupervised 3D Object Detection", "abstract": " The prevalent approaches of unsupervised 3D object detection follow\ncluster-based pseudo-label generation and iterative self-training processes.\nHowever, the challenge arises due to the sparsity of LiDAR scans, which leads\nto pseudo-labels with erroneous size and position, resulting in subpar\ndetection performance. To tackle this problem, this paper introduces a\nCommonsense Prototype-based Detector, termed CPD, for unsupervised 3D object\ndetection. CPD first constructs Commonsense Prototype (CProto) characterized by\nhigh-quality bounding box and dense points, based on commonsense intuition.\nSubsequently, CPD refines the low-quality pseudo-labels by leveraging the size\nprior from CProto. Furthermore, CPD enhances the detection accuracy of sparsely\nscanned objects by the geometric knowledge from CProto. CPD outperforms\nstate-of-the-art unsupervised 3D detectors on Waymo Open Dataset (WOD),\nPandaSet, and KITTI datasets by a large margin. Besides, by training CPD on WOD\nand testing on KITTI, CPD attains 90.85% and 81.01% 3D Average Precision on\neasy and moderate car classes, respectively. These achievements position CPD in\nclose proximity to fully supervised detectors, highlighting the significance of\nour method. The code will be available at https://github.com/hailanyi/CPD.\n", "date": "2024-04-25", "categories": []} {"id": "2406.15805", "title": "Smart Feature is What You Need", "abstract": " Lack of shape guidance and label jitter caused by information deficiency of\nweak label are the main problems in 3D weakly-supervised object detection.\nCurrent weakly-supervised models often use heuristics or assumptions methods to\ninfer information from weak labels without taking advantage of the inherent\nclues of weakly-supervised and fully-supervised methods, thus it is difficult\nto explore a method that combines data utilization efficiency and model\naccuracy. In an attempt to address these issues, we propose a novel plug-and-in\npoint cloud feature representation network called Multi-scale Mixed Attention\n(MMA). MMA utilizes adjacency attention within neighborhoods and disparity\nattention at different density scales to build a feature representation\nnetwork. The smart feature representation obtained from MMA has shape tendency\nand object existence area inference, which can constrain the region of the\ndetection boxes, thereby alleviating the problems caused by the information\ndefault of weak labels. Extensive experiments show that in indoor weak label\nscenarios, the fully-supervised network can perform close to that of the\nweakly-supervised network merely through the improvement of point feature by\nMMA. At the same time, MMA can turn waste into treasure, reversing the label\njitter problem that originally interfered with weakly-supervised detection into\nthe source of data enhancement, strengthening the performance of existing weak\nsupervision detection methods. Our code is available at\nhttps://github.com/hzx-9894/MMA.\n", "date": "2024-06-22", "categories": []} {"id": "2302.05923", "title": "Uncertainty-Aware AB3DMOT by Variational 3D Object Detection", "abstract": " Autonomous driving needs to rely on high-quality 3D object detection to\nensure safe navigation in the world. Uncertainty estimation is an effective\ntool to provide statistically accurate predictions, while the associated\ndetection uncertainty can be used to implement a more safe navigation protocol\nor include the user in the loop. In this paper, we propose a Variational Neural\nNetwork-based TANet 3D object detector to generate 3D object detections with\nuncertainty and introduce these detections to an uncertainty-aware AB3DMOT\ntracker. This is done by applying a linear transformation to the estimated\nuncertainty matrix, which is subsequently used as a measurement noise for the\nadopted Kalman filter. We implement two ways to estimate output uncertainty,\ni.e., internally, by computing the variance of the CNN outputs and then\npropagating the uncertainty through the post-processing, and externally, by\nassociating the final predictions of different samples and computing the\ncovariance of each predicted box. In experiments, we show that the external\nuncertainty estimation leads to better results, outperforming both internal\nuncertainty estimation and classical tracking approaches. Furthermore, we\npropose a method to initialize the Variational 3D object detector with a\npretrained TANet model, which leads to the best performing models.\n", "date": "2023-02-12", "categories": []} {"id": "2404.12856", "title": "Language-Driven Active Learning for Diverse Open-Set 3D Object Detection", "abstract": " Object detection is crucial for ensuring safe autonomous driving. However,\ndata-driven approaches face challenges when encountering minority or novel\nobjects in the 3D driving scene. In this paper, we propose VisLED, a\nlanguage-driven active learning framework for diverse open-set 3D Object\nDetection. Our method leverages active learning techniques to query diverse and\ninformative data samples from an unlabeled pool, enhancing the model's ability\nto detect underrepresented or novel objects. Specifically, we introduce the\nVision-Language Embedding Diversity Querying (VisLED-Querying) algorithm, which\noperates in both open-world exploring and closed-world mining settings. In\nopen-world exploring, VisLED-Querying selects data points most novel relative\nto existing data, while in closed-world mining, it mines novel instances of\nknown classes. We evaluate our approach on the nuScenes dataset and demonstrate\nits efficiency compared to random sampling and entropy-querying methods. Our\nresults show that VisLED-Querying consistently outperforms random sampling and\noffers competitive performance compared to entropy-querying despite the\nlatter's model-optimality, highlighting the potential of VisLED for improving\nobject detection in autonomous driving scenarios. We make our code publicly\navailable at https://github.com/Bjork-crypto/VisLED-Querying\n", "date": "2024-04-19", "categories": []} {"id": "2406.10700", "title": "Voxel Mamba: Group-Free State Space Models for Point Cloud based 3D\n Object Detection", "abstract": " Serialization-based methods, which serialize the 3D voxels and group them\ninto multiple sequences before inputting to Transformers, have demonstrated\ntheir effectiveness in 3D object detection. However, serializing 3D voxels into\n1D sequences will inevitably sacrifice the voxel spatial proximity. Such an\nissue is hard to be addressed by enlarging the group size with existing\nserialization-based methods due to the quadratic complexity of Transformers\nwith feature sizes. Inspired by the recent advances of state space models\n(SSMs), we present a Voxel SSM, termed as Voxel Mamba, which employs a\ngroup-free strategy to serialize the whole space of voxels into a single\nsequence. The linear complexity of SSMs encourages our group-free design,\nalleviating the loss of spatial proximity of voxels. To further enhance the\nspatial proximity, we propose a Dual-scale SSM Block to establish a\nhierarchical structure, enabling a larger receptive field in the 1D\nserialization curve, as well as more complete local regions in 3D space.\nMoreover, we implicitly apply window partition under the group-free framework\nby positional encoding, which further enhances spatial proximity by encoding\nvoxel positional information. Our experiments on Waymo Open Dataset and\nnuScenes dataset show that Voxel Mamba not only achieves higher accuracy than\nstate-of-the-art methods, but also demonstrates significant advantages in\ncomputational efficiency.\n", "date": "2024-06-15", "categories": []} {"id": "2307.16121", "title": "Uncertainty-Encoded Multi-Modal Fusion for Robust Object Detection in\n Autonomous Driving", "abstract": " Multi-modal fusion has shown initial promising results for object detection\nof autonomous driving perception. However, many existing fusion schemes do not\nconsider the quality of each fusion input and may suffer from adverse\nconditions on one or more sensors. While predictive uncertainty has been\napplied to characterize single-modal object detection performance at run time,\nincorporating uncertainties into the multi-modal fusion still lacks effective\nsolutions due primarily to the uncertainty's cross-modal incomparability and\ndistinct sensitivities to various adverse conditions. To fill this gap, this\npaper proposes Uncertainty-Encoded Mixture-of-Experts (UMoE) that explicitly\nincorporates single-modal uncertainties into LiDAR-camera fusion. UMoE uses\nindividual expert network to process each sensor's detection result together\nwith encoded uncertainty. Then, the expert networks' outputs are analyzed by a\ngating network to determine the fusion weights. The proposed UMoE module can be\nintegrated into any proposal fusion pipeline. Evaluation shows that UMoE\nachieves a maximum of 10.67%, 3.17%, and 5.40% performance gain compared with\nthe state-of-the-art proposal-level multi-modal object detectors under extreme\nweather, adversarial, and blinding attack scenarios.\n", "date": "2023-07-30", "categories": []} {"id": "2406.10907", "title": "SparseDet: A Simple and Effective Framework for Fully Sparse LiDAR-based\n 3D Object Detection", "abstract": " LiDAR-based sparse 3D object detection plays a crucial role in autonomous\ndriving applications due to its computational efficiency advantages. Existing\nmethods either use the features of a single central voxel as an object proxy,\nor treat an aggregated cluster of foreground points as an object proxy.\nHowever, the former lacks the ability to aggregate contextual information,\nresulting in insufficient information expression in object proxies. The latter\nrelies on multi-stage pipelines and auxiliary tasks, which reduce the inference\nspeed. To maintain the efficiency of the sparse framework while fully\naggregating contextual information, in this work, we propose SparseDet which\ndesigns sparse queries as object proxies. It introduces two key modules, the\nLocal Multi-scale Feature Aggregation (LMFA) module and the Global Feature\nAggregation (GFA) module, aiming to fully capture the contextual information,\nthereby enhancing the ability of the proxies to represent objects. Where LMFA\nsub-module achieves feature fusion across different scales for sparse key\nvoxels %which does this through via coordinate transformations and using\nnearest neighbor relationships to capture object-level details and local\ncontextual information, GFA sub-module uses self-attention mechanisms to\nselectively aggregate the features of the key voxels across the entire scene\nfor capturing scene-level contextual information. Experiments on nuScenes and\nKITTI demonstrate the effectiveness of our method. Specifically, on nuScene,\nSparseDet surpasses the previous best sparse detector VoxelNeXt by 2.2\\% mAP\nwith 13.5 FPS, and on KITTI, it surpasses VoxelNeXt by 1.12\\%\n$\\mathbf{AP_{3D}}$ on hard level tasks with 17.9 FPS.\n", "date": "2024-06-16", "categories": []} {"id": "2406.11313", "title": "Semi-Supervised Domain Adaptation Using Target-Oriented Domain\n Augmentation for 3D Object Detection", "abstract": " 3D object detection is crucial for applications like autonomous driving and\nrobotics. However, in real-world environments, variations in sensor data\ndistribution due to sensor upgrades, weather changes, and geographic\ndifferences can adversely affect detection performance. Semi-Supervised Domain\nAdaptation (SSDA) aims to mitigate these challenges by transferring knowledge\nfrom a source domain, abundant in labeled data, to a target domain where labels\nare scarce. This paper presents a new SSDA method referred to as\nTarget-Oriented Domain Augmentation (TODA) specifically tailored for\nLiDAR-based 3D object detection. TODA efficiently utilizes all available data,\nincluding labeled data in the source domain, and both labeled data and\nunlabeled data in the target domain to enhance domain adaptation performance.\nTODA consists of two stages: TargetMix and AdvMix. TargetMix employs mixing\naugmentation accounting for LiDAR sensor characteristics to facilitate feature\nalignment between the source-domain and target-domain. AdvMix applies\npoint-wise adversarial augmentation with mixing augmentation, which perturbs\nthe unlabeled data to align the features within both labeled and unlabeled data\nin the target domain. Our experiments conducted on the challenging domain\nadaptation tasks demonstrate that TODA outperforms existing domain adaptation\ntechniques designed for 3D object detection by significant margins. The code is\navailable at: https://github.com/rasd3/TODA.\n", "date": "2024-06-17", "categories": []} {"id": "2406.10224", "title": "EFM3D: A Benchmark for Measuring Progress Towards 3D Egocentric\n Foundation Models", "abstract": " The advent of wearable computers enables a new source of context for AI that\nis embedded in egocentric sensor data. This new egocentric data comes equipped\nwith fine-grained 3D location information and thus presents the opportunity for\na novel class of spatial foundation models that are rooted in 3D space. To\nmeasure progress on what we term Egocentric Foundation Models (EFMs) we\nestablish EFM3D, a benchmark with two core 3D egocentric perception tasks.\nEFM3D is the first benchmark for 3D object detection and surface regression on\nhigh quality annotated egocentric data of Project Aria. We propose Egocentric\nVoxel Lifting (EVL), a baseline for 3D EFMs. EVL leverages all available\negocentric modalities and inherits foundational capabilities from 2D foundation\nmodels. This model, trained on a large simulated dataset, outperforms existing\nmethods on the EFM3D benchmark.\n", "date": "2024-06-14", "categories": []} {"id": "2406.08785", "title": "BEVSpread: Spread Voxel Pooling for Bird's-Eye-View Representation in\n Vision-based Roadside 3D Object Detection", "abstract": " Vision-based roadside 3D object detection has attracted rising attention in\nautonomous driving domain, since it encompasses inherent advantages in reducing\nblind spots and expanding perception range. While previous work mainly focuses\non accurately estimating depth or height for 2D-to-3D mapping, ignoring the\nposition approximation error in the voxel pooling process. Inspired by this\ninsight, we propose a novel voxel pooling strategy to reduce such error, dubbed\nBEVSpread. Specifically, instead of bringing the image features contained in a\nfrustum point to a single BEV grid, BEVSpread considers each frustum point as a\nsource and spreads the image features to the surrounding BEV grids with\nadaptive weights. To achieve superior propagation performance, a specific\nweight function is designed to dynamically control the decay speed of the\nweights according to distance and depth. Aided by customized CUDA parallel\nacceleration, BEVSpread achieves comparable inference time as the original\nvoxel pooling. Extensive experiments on two large-scale roadside benchmarks\ndemonstrate that, as a plug-in, BEVSpread can significantly improve the\nperformance of existing frustum-based BEV methods by a large margin of (1.12,\n5.26, 3.01) AP in vehicle, pedestrian and cyclist.\n", "date": "2024-06-13", "categories": []} {"id": "2305.03724", "title": "DualCross: Cross-Modality Cross-Domain Adaptation for Monocular BEV\n Perception", "abstract": " Closing the domain gap between training and deployment and incorporating\nmultiple sensor modalities are two challenging yet critical topics for\nself-driving. Existing work only focuses on single one of the above topics,\noverlooking the simultaneous domain and modality shift which pervasively exists\nin real-world scenarios. A model trained with multi-sensor data collected in\nEurope may need to run in Asia with a subset of input sensors available. In\nthis work, we propose DualCross, a cross-modality cross-domain adaptation\nframework to facilitate the learning of a more robust monocular bird's-eye-view\n(BEV) perception model, which transfers the point cloud knowledge from a LiDAR\nsensor in one domain during the training phase to the camera-only testing\nscenario in a different domain. This work results in the first open analysis of\ncross-domain cross-sensor perception and adaptation for monocular 3D tasks in\nthe wild. We benchmark our approach on large-scale datasets under a wide range\nof domain shifts and show state-of-the-art results against various baselines.\n", "date": "2023-05-05", "categories": []} {"id": "2406.07833", "title": "Sense Less, Generate More: Pre-training LiDAR Perception with Masked\n Autoencoders for Ultra-Efficient 3D Sensing", "abstract": " In this work, we propose a disruptively frugal LiDAR perception dataflow that\ngenerates rather than senses parts of the environment that are either\npredictable based on the extensive training of the environment or have limited\nconsequence to the overall prediction accuracy. Therefore, the proposed\nmethodology trades off sensing energy with training data for low-power robotics\nand autonomous navigation to operate frugally with sensors, extending their\nlifetime on a single battery charge. Our proposed generative pre-training\nstrategy for this purpose, called as radially masked autoencoding (R-MAE), can\nalso be readily implemented in a typical LiDAR system by selectively activating\nand controlling the laser power for randomly generated angular regions during\non-field operations. Our extensive evaluations show that pre-training with\nR-MAE enables focusing on the radial segments of the data, thereby capturing\nspatial relationships and distances between objects more effectively than\nconventional procedures. Therefore, the proposed methodology not only reduces\nsensing energy but also improves prediction accuracy. For example, our\nextensive evaluations on Waymo, nuScenes, and KITTI datasets show that the\napproach achieves over a 5% average precision improvement in detection tasks\nacross datasets and over a 4% accuracy improvement in transferring domains from\nWaymo and nuScenes to KITTI. In 3D object detection, it enhances small object\ndetection by up to 4.37% in AP at moderate difficulty levels in the KITTI\ndataset. Even with 90% radial masking, it surpasses baseline models by up to\n5.59% in mAP/mAPH across all object classes in the Waymo dataset. Additionally,\nour method achieves up to 3.17% and 2.31% improvements in mAP and NDS,\nrespectively, on the nuScenes dataset, demonstrating its effectiveness with\nboth single and fused LiDAR-camera modalities.\nhttps://github.com/sinatayebati/Radial_MAE.\n", "date": "2024-06-12", "categories": []} {"id": "2406.08152", "title": "CT3D++: Improving 3D Object Detection with Keypoint-induced Channel-wise\n Transformer", "abstract": " The field of 3D object detection from point clouds is rapidly advancing in\ncomputer vision, aiming to accurately and efficiently detect and localize\nobjects in three-dimensional space. Current 3D detectors commonly fall short in\nterms of flexibility and scalability, with ample room for advancements in\nperformance. In this paper, our objective is to address these limitations by\nintroducing two frameworks for 3D object detection with minimal hand-crafted\ndesign. Firstly, we propose CT3D, which sequentially performs raw-point-based\nembedding, a standard Transformer encoder, and a channel-wise decoder for point\nfeatures within each proposal. Secondly, we present an enhanced network called\nCT3D++, which incorporates geometric and semantic fusion-based embedding to\nextract more valuable and comprehensive proposal-aware information.\nAdditionally, CT3D ++ utilizes a point-to-key bidirectional encoder for more\nefficient feature encoding with reduced computational cost. By replacing the\ncorresponding components of CT3D with these novel modules, CT3D++ achieves\nstate-of-the-art performance on both the KITTI dataset and the large-scale\nWay\\-mo Open Dataset. The source code for our frameworks will be made\naccessible at https://github.com/hlsheng1/CT3D-plusplus.\n", "date": "2024-06-12", "categories": []} {"id": "2406.04648", "title": "UCDNet: Multi-UAV Collaborative 3D Object Detection Network by Reliable\n Feature Mapping", "abstract": " Multi-UAV collaborative 3D object detection can perceive and comprehend\ncomplex environments by integrating complementary information, with\napplications encompassing traffic monitoring, delivery services and\nagricultural management. However, the extremely broad observations in aerial\nremote sensing and significant perspective differences across multiple UAVs\nmake it challenging to achieve precise and consistent feature mapping from 2D\nimages to 3D space in multi-UAV collaborative 3D object detection paradigm. To\naddress the problem, we propose an unparalleled camera-based multi-UAV\ncollaborative 3D object detection paradigm called UCDNet. Specifically, the\ndepth information from the UAVs to the ground is explicitly utilized as a\nstrong prior to provide a reference for more accurate and generalizable feature\nmapping. Additionally, we design a homologous points geometric consistency loss\nas an auxiliary self-supervision, which directly influences the feature mapping\nmodule, thereby strengthening the global consistency of multi-view perception.\nExperiments on AeroCollab3D and CoPerception-UAVs datasets show our method\nincreases 4.7% and 10% mAP respectively compared to the baseline, which\ndemonstrates the superiority of UCDNet.\n", "date": "2024-06-07", "categories": []} {"id": "2406.03129", "title": "Enhanced Automotive Object Detection via RGB-D Fusion in a DiffusionDet\n Framework", "abstract": " Vision-based autonomous driving requires reliable and efficient object\ndetection. This work proposes a DiffusionDet-based framework that exploits data\nfusion from the monocular camera and depth sensor to provide the RGB and depth\n(RGB-D) data. Within this framework, ground truth bounding boxes are randomly\nreshaped as part of the training phase, allowing the model to learn the reverse\ndiffusion process of noise addition. The system methodically enhances a\nrandomly generated set of boxes at the inference stage, guiding them toward\naccurate final detections. By integrating the textural and color features from\nRGB images with the spatial depth information from the LiDAR sensors, the\nproposed framework employs a feature fusion that substantially enhances object\ndetection of automotive targets. The $2.3$ AP gain in detecting automotive\ntargets is achieved through comprehensive experiments using the KITTI dataset.\nSpecifically, the improved performance of the proposed approach in detecting\nsmall objects is demonstrated.\n", "date": "2024-06-05", "categories": []} {"id": "2406.02533", "title": "SatSplatYOLO: 3D Gaussian Splatting-based Virtual Object Detection\n Ensembles for Satellite Feature Recognition", "abstract": " On-orbit servicing (OOS), inspection of spacecraft, and active debris removal\n(ADR). Such missions require precise rendezvous and proximity operations in the\nvicinity of non-cooperative, possibly unknown, resident space objects. Safety\nconcerns with manned missions and lag times with ground-based control\nnecessitate complete autonomy. In this article, we present an approach for\nmapping geometries and high-confidence detection of components of unknown,\nnon-cooperative satellites on orbit. We implement accelerated 3D Gaussian\nsplatting to learn a 3D representation of the satellite, render virtual views\nof the target, and ensemble the YOLOv5 object detector over the virtual views,\nresulting in reliable, accurate, and precise satellite component detections.\nThe full pipeline capable of running on-board and stand to enable downstream\nmachine intelligence tasks necessary for autonomous guidance, navigation, and\ncontrol tasks.\n", "date": "2024-06-04", "categories": []} {"id": "2406.00714", "title": "A Survey of Deep Learning Based Radar and Vision Fusion for 3D Object\n Detection in Autonomous Driving", "abstract": " With the rapid advancement of autonomous driving technology, there is a\ngrowing need for enhanced safety and efficiency in the automatic environmental\nperception of vehicles during their operation. In modern vehicle setups,\ncameras and mmWave radar (radar), being the most extensively employed sensors,\ndemonstrate complementary characteristics, inherently rendering them conducive\nto fusion and facilitating the achievement of both robust performance and\ncost-effectiveness. This paper focuses on a comprehensive survey of\nradar-vision (RV) fusion based on deep learning methods for 3D object detection\nin autonomous driving. We offer a comprehensive overview of each RV fusion\ncategory, specifically those employing region of interest (ROI) fusion and\nend-to-end fusion strategies. As the most promising fusion strategy at present,\nwe provide a deeper classification of end-to-end fusion methods, including\nthose 3D bounding box prediction based and BEV based approaches. Moreover,\naligning with recent advancements, we delineate the latest information on 4D\nradar and its cutting-edge applications in autonomous vehicles (AVs). Finally,\nwe present the possible future trends of RV fusion and summarize this paper.\n", "date": "2024-06-02", "categories": []} {"id": "2405.20720", "title": "Power of Cooperative Supervision: Multiple Teachers Framework for\n Enhanced 3D Semi-Supervised Object Detection", "abstract": " To ensure safe urban driving for autonomous platforms, it is crucial not only\nto develop high-performance object detection techniques but also to establish a\ndiverse and representative dataset that captures various urban environments and\nobject characteristics. To address these two issues, we have constructed a\nmulti-class 3D LiDAR dataset reflecting diverse urban environments and object\ncharacteristics, and developed a robust 3D semi-supervised object detection\n(SSOD) based on a multiple teachers framework. This SSOD framework categorizes\nsimilar classes and assigns specialized teachers to each category. Through\ncollaborative supervision among these category-specialized teachers, the\nstudent network becomes increasingly proficient, leading to a highly effective\nobject detector. We propose a simple yet effective augmentation technique,\nPie-based Point Compensating Augmentation (PieAug), to enable the teacher\nnetwork to generate high-quality pseudo-labels. Extensive experiments on the\nWOD, KITTI, and our datasets validate the effectiveness of our proposed method\nand the quality of our dataset. Experimental results demonstrate that our\napproach consistently outperforms existing state-of-the-art 3D semi-supervised\nobject detection methods across all datasets. We plan to release our\nmulti-class LiDAR dataset and the source code available on our Github\nrepository in the near future.\n", "date": "2024-05-31", "categories": []} {"id": "2405.19682", "title": "Fully Test-Time Adaptation for Monocular 3D Object Detection", "abstract": " Monocular 3D object detection (Mono 3Det) aims to identify 3D objects from a\nsingle RGB image. However, existing methods often assume training and test data\nfollow the same distribution, which may not hold in real-world test scenarios.\nTo address the out-of-distribution (OOD) problems, we explore a new adaptation\nparadigm for Mono 3Det, termed Fully Test-time Adaptation. It aims to adapt a\nwell-trained model to unlabeled test data by handling potential data\ndistribution shifts at test time without access to training data and test\nlabels. However, applying this paradigm in Mono 3Det poses significant\nchallenges due to OOD test data causing a remarkable decline in object\ndetection scores. This decline conflicts with the pre-defined score thresholds\nof existing detection methods, leading to severe object omissions (i.e., rare\npositive detections and many false negatives). Consequently, the limited\npositive detection and plenty of noisy predictions cause test-time adaptation\nto fail in Mono 3Det. To handle this problem, we propose a novel Monocular\nTest-Time Adaptation (MonoTTA) method, based on two new strategies. 1)\nReliability-driven adaptation: we empirically find that high-score objects are\nstill reliable and the optimization of high-score objects can enhance\nconfidence across all detections. Thus, we devise a self-adaptive strategy to\nidentify reliable objects for model adaptation, which discovers potential\nobjects and alleviates omissions. 2) Noise-guard adaptation: since high-score\nobjects may be scarce, we develop a negative regularization term to exploit the\nnumerous low-score objects via negative learning, preventing overfitting to\nnoise and trivial solutions. Experimental results show that MonoTTA brings\nsignificant performance gains for Mono 3Det models in OOD test scenarios,\napproximately 190% gains by average on KITTI and 198% gains on nuScenes.\n", "date": "2024-05-30", "categories": []} {"id": "2309.04806", "title": "Timely Fusion of Surround Radar/Lidar for Object Detection in Autonomous\n Driving Systems", "abstract": " Fusing Radar and Lidar sensor data can fully utilize their complementary\nadvantages and provide more accurate reconstruction of the surrounding for\nautonomous driving systems. Surround Radar/Lidar can provide 360-degree view\nsampling with the minimal cost, which are promising sensing hardware solutions\nfor autonomous driving systems. However, due to the intrinsic physical\nconstraints, the rotating speed of surround Radar, and thus the frequency to\ngenerate Radar data frames, is much lower than surround Lidar. Existing\nRadar/Lidar fusion methods have to work at the low frequency of surround Radar,\nwhich cannot meet the high responsiveness requirement of autonomous driving\nsystems.This paper develops techniques to fuse surround Radar/Lidar with\nworking frequency only limited by the faster surround Lidar instead of the\nslower surround Radar, based on the state-of-the-art object detection model\nMVDNet. The basic idea of our approach is simple: we let MVDNet work with\ntemporally unaligned data from Radar/Lidar, so that fusion can take place at\nany time when a new Lidar data frame arrives, instead of waiting for the slow\nRadar data frame. However, directly applying MVDNet to temporally unaligned\nRadar/Lidar data greatly degrades its object detection accuracy. The key\ninformation revealed in this paper is that we can achieve high output frequency\nwith little accuracy loss by enhancing the training procedure to explore the\ntemporal redundancy in MVDNet so that it can tolerate the temporal unalignment\nof input data. We explore several different ways of training enhancement and\ncompare them quantitatively with experiments.\n", "date": "2023-09-09", "categories": []} {"id": "2405.15843", "title": "SpotNet: An Image Centric, Lidar Anchored Approach To Long Range\n Perception", "abstract": " In this paper, we propose SpotNet: a fast, single stage, image-centric but\nLiDAR anchored approach for long range 3D object detection. We demonstrate that\nour approach to LiDAR/image sensor fusion, combined with the joint learning of\n2D and 3D detection tasks, can lead to accurate 3D object detection with very\nsparse LiDAR support. Unlike more recent bird's-eye-view (BEV) sensor-fusion\nmethods which scale with range $r$ as $O(r^2)$, SpotNet scales as $O(1)$ with\nrange. We argue that such an architecture is ideally suited to leverage each\nsensor's strength, i.e. semantic understanding from images and accurate range\nfinding from LiDAR data. Finally we show that anchoring detections on LiDAR\npoints removes the need to regress distances, and so the architecture is able\nto transfer from 2MP to 8MP resolution images without re-training.\n", "date": "2024-05-24", "categories": []} {"id": "2405.17422", "title": "Hardness-Aware Scene Synthesis for Semi-Supervised 3D Object Detection", "abstract": " 3D object detection aims to recover the 3D information of concerning objects\nand serves as the fundamental task of autonomous driving perception. Its\nperformance greatly depends on the scale of labeled training data, yet it is\ncostly to obtain high-quality annotations for point cloud data. While\nconventional methods focus on generating pseudo-labels for unlabeled samples as\nsupplements for training, the structural nature of 3D point cloud data\nfacilitates the composition of objects and backgrounds to synthesize realistic\nscenes. Motivated by this, we propose a hardness-aware scene synthesis (HASS)\nmethod to generate adaptive synthetic scenes to improve the generalization of\nthe detection models. We obtain pseudo-labels for unlabeled objects and\ngenerate diverse scenes with different compositions of objects and backgrounds.\nAs the scene synthesis is sensitive to the quality of pseudo-labels, we further\npropose a hardness-aware strategy to reduce the effect of low-quality\npseudo-labels and maintain a dynamic pseudo-database to ensure the diversity\nand quality of synthetic scenes. Extensive experimental results on the widely\nused KITTI and Waymo datasets demonstrate the superiority of the proposed HASS\nmethod, which outperforms existing semi-supervised learning methods on 3D\nobject detection. Code: https://github.com/wzzheng/HASS.\n", "date": "2024-05-27", "categories": []} {"id": "2404.14780", "title": "ContextualFusion: Context-Based Multi-Sensor Fusion for 3D Object\n Detection in Adverse Operating Conditions", "abstract": " The fusion of multimodal sensor data streams such as camera images and lidar\npoint clouds plays an important role in the operation of autonomous vehicles\n(AVs). Robust perception across a range of adverse weather and lighting\nconditions is specifically required for AVs to be deployed widely. While\nmulti-sensor fusion networks have been previously developed for perception in\nsunny and clear weather conditions, these methods show a significant\ndegradation in performance under night-time and poor weather conditions. In\nthis paper, we propose a simple yet effective technique called ContextualFusion\nto incorporate the domain knowledge about cameras and lidars behaving\ndifferently across lighting and weather variations into 3D object detection\nmodels. Specifically, we design a Gated Convolutional Fusion (GatedConv)\napproach for the fusion of sensor streams based on the operational context. To\naid in our evaluation, we use the open-source simulator CARLA to create a\nmultimodal adverse-condition dataset called AdverseOp3D to address the\nshortcomings of existing datasets being biased towards daytime and good-weather\nconditions. Our ContextualFusion approach yields an mAP improvement of 6.2%\nover state-of-the-art methods on our context-balanced synthetic dataset.\nFinally, our method enhances state-of-the-art 3D objection performance at night\non the real-world NuScenes dataset with a significant mAP improvement of 11.7%.\n", "date": "2024-04-23", "categories": []} {"id": "2312.14474", "title": "MonoLSS: Learnable Sample Selection For Monocular 3D Detection", "abstract": " In the field of autonomous driving, monocular 3D detection is a critical task\nwhich estimates 3D properties (depth, dimension, and orientation) of objects in\na single RGB image. Previous works have used features in a heuristic way to\nlearn 3D properties, without considering that inappropriate features could have\nadverse effects. In this paper, sample selection is introduced that only\nsuitable samples should be trained to regress the 3D properties. To select\nsamples adaptively, we propose a Learnable Sample Selection (LSS) module, which\nis based on Gumbel-Softmax and a relative-distance sample divider. The LSS\nmodule works under a warm-up strategy leading to an improvement in training\nstability. Additionally, since the LSS module dedicated to 3D property sample\nselection relies on object-level features, we further develop a data\naugmentation method named MixUp3D to enrich 3D property samples which conforms\nto imaging principles without introducing ambiguity. As two orthogonal methods,\nthe LSS module and MixUp3D can be utilized independently or in conjunction.\nSufficient experiments have shown that their combined use can lead to\nsynergistic effects, yielding improvements that transcend the mere sum of their\nindividual applications. Leveraging the LSS module and the MixUp3D, without any\nextra data, our method named MonoLSS ranks 1st in all three categories (Car,\nCyclist, and Pedestrian) on KITTI 3D object detection benchmark, and achieves\ncompetitive results on both the Waymo dataset and KITTI-nuScenes cross-dataset\nevaluation. The code is included in the supplementary material and will be\nreleased to facilitate related academic and industrial studies.\n", "date": "2023-12-22", "categories": []} {"id": "2401.07240", "title": "DCDet: Dynamic Cross-based 3D Object Detector", "abstract": " Recently, significant progress has been made in the research of 3D object\ndetection. However, most prior studies have focused on the utilization of\ncenter-based or anchor-based label assignment schemes. Alternative label\nassignment strategies remain unexplored in 3D object detection. We find that\nthe center-based label assignment often fails to generate sufficient positive\nsamples for training, while the anchor-based label assignment tends to\nencounter an imbalanced issue when handling objects of varying scales. To solve\nthese issues, we introduce a dynamic cross label assignment (DCLA) scheme,\nwhich dynamically assigns positive samples for each object from a cross-shaped\nregion, thus providing sufficient and balanced positive samples for training.\nFurthermore, to address the challenge of accurately regressing objects with\nvarying scales, we put forth a rotation-weighted Intersection over Union\n(RWIoU) metric to replace the widely used L1 metric in regression loss.\nExtensive experiments demonstrate the generality and effectiveness of our DCLA\nand RWIoU-based regression loss. The Code will be available at\nhttps://github.com/Say2L/DCDet.git.\n", "date": "2024-01-14", "categories": []} {"id": "2405.13202", "title": "Empowering Urban Traffic Management: Elevated 3D LiDAR for Data\n Collection and Advanced Object Detection Analysis", "abstract": " The 3D object detection capabilities in urban environments have been\nenormously improved by recent developments in Light Detection and Range (LiDAR)\ntechnology. This paper presents a novel framework that transforms the detection\nand analysis of 3D objects in traffic scenarios by utilizing the power of\nelevated LiDAR sensors. We are presenting our methodology's remarkable capacity\nto collect complex 3D point cloud data, which allows us to accurately and in\ndetail capture the dynamics of urban traffic. Due to the limitation in\nobtaining real-world traffic datasets, we utilize the simulator to generate 3D\npoint cloud for specific scenarios. To support our experimental analysis, we\nfirstly simulate various 3D point cloud traffic-related objects. Then, we use\nthis dataset as a basis for training and evaluating our 3D object detection\nmodels, in identifying and monitoring both vehicles and pedestrians in\nsimulated urban traffic environments. Next, we fine tune the Point\nVoxel-Region-based Convolutional Neural Network (PV-RCNN) architecture, making\nit more suited to handle and understand the massive volumes of point cloud data\ngenerated by our urban traffic simulations. Our results show the effectiveness\nof the proposed solution in accurately detecting objects in traffic scenes and\nhighlight the role of LiDAR in improving urban safety and advancing intelligent\ntransportation systems.\n", "date": "2024-05-21", "categories": []} {"id": "2310.04800", "title": "Towards Long-Range 3D Object Detection for Autonomous Vehicles", "abstract": " 3D object detection at long range is crucial for ensuring the safety and\nefficiency of self driving vehicles, allowing them to accurately perceive and\nreact to objects, obstacles, and potential hazards from a distance. But most\ncurrent state of the art LiDAR based methods are range limited due to sparsity\nat long range, which generates a form of domain gap between points closer to\nand farther away from the ego vehicle. Another related problem is the label\nimbalance for faraway objects, which inhibits the performance of Deep Neural\nNetworks at long range. To address the above limitations, we investigate two\nways to improve long range performance of current LiDAR based 3D detectors.\nFirst, we combine two 3D detection networks, referred to as range experts, one\nspecializing at near to mid range objects, and one at long range 3D detection.\nTo train a detector at long range under a scarce label regime, we further weigh\nthe loss according to the labelled point's distance from ego vehicle. Second,\nwe augment LiDAR scans with virtual points generated using Multimodal Virtual\nPoints (MVP), a readily available image-based depth completion algorithm. Our\nexperiments on the long range Argoverse2 (AV2) dataset indicate that MVP is\nmore effective in improving long range performance, while maintaining a\nstraightforward implementation. On the other hand, the range experts offer a\ncomputationally efficient and simpler alternative, avoiding dependency on\nimage-based segmentation networks and perfect camera-LiDAR calibration.\n", "date": "2023-10-07", "categories": []} {"id": "2312.14919", "title": "Lift-Attend-Splat: Bird's-eye-view camera-lidar fusion using\n transformers", "abstract": " Combining complementary sensor modalities is crucial to providing robust\nperception for safety-critical robotics applications such as autonomous driving\n(AD). Recent state-of-the-art camera-lidar fusion methods for AD rely on\nmonocular depth estimation which is a notoriously difficult task compared to\nusing depth information from the lidar directly. Here, we find that this\napproach does not leverage depth as expected and show that naively improving\ndepth estimation does not lead to improvements in object detection performance.\nStrikingly, we also find that removing depth estimation altogether does not\ndegrade object detection performance substantially, suggesting that relying on\nmonocular depth could be an unnecessary architectural bottleneck during\ncamera-lidar fusion. In this work, we introduce a novel fusion method that\nbypasses monocular depth estimation altogether and instead selects and fuses\ncamera and lidar features in a bird's-eye-view grid using a simple attention\nmechanism. We show that our model can modulate its use of camera features based\non the availability of lidar features and that it yields better 3D object\ndetection on the nuScenes dataset than baselines relying on monocular depth\nestimation.\n", "date": "2023-12-22", "categories": []} {"id": "2403.02639", "title": "False Positive Sampling-based Data Augmentation for Enhanced 3D Object\n Detection Accuracy", "abstract": " Recent studies have focused on enhancing the performance of 3D object\ndetection models. Among various approaches, ground-truth sampling has been\nproposed as an augmentation technique to address the challenges posed by\nlimited ground-truth data. However, an inherent issue with ground-truth\nsampling is its tendency to increase false positives. Therefore, this study\naims to overcome the limitations of ground-truth sampling and improve the\nperformance of 3D object detection models by developing a new augmentation\ntechnique called false-positive sampling. False-positive sampling involves\nretraining the model using point clouds that are identified as false positives\nin the model's predictions. We propose an algorithm that utilizes both\nground-truth and false-positive sampling and an algorithm for building the\nfalse-positive sample database. Additionally, we analyze the principles behind\nthe performance enhancement due to false-positive sampling. Our experiments\ndemonstrate that models utilizing false-positive sampling show a reduction in\nfalse positives and exhibit improved object detection performance. On the KITTI\nand Waymo Open datasets, models with false-positive sampling surpass the\nbaseline models by a large margin.\n", "date": "2024-03-05", "categories": []} {"id": "2405.11682", "title": "FADet: A Multi-sensor 3D Object Detection Network based on Local\n Featured Attention", "abstract": " Camera, LiDAR and radar are common perception sensors for autonomous driving\ntasks. Robust prediction of 3D object detection is optimally based on the\nfusion of these sensors. To exploit their abilities wisely remains a challenge\nbecause each of these sensors has its own characteristics. In this paper, we\npropose FADet, a multi-sensor 3D detection network, which specifically studies\nthe characteristics of different sensors based on our local featured attention\nmodules. For camera images, we propose dual-attention-based sub-module. For\nLiDAR point clouds, triple-attention-based sub-module is utilized while\nmixed-attention-based sub-module is applied for features of radar points. With\nlocal featured attention sub-modules, our FADet has effective detection results\nin long-tail and complex scenes from camera, LiDAR and radar input. On NuScenes\nvalidation dataset, FADet achieves state-of-the-art performance on LiDAR-camera\nobject detection tasks with 71.8% NDS and 69.0% mAP, at the same time, on\nradar-camera object detection tasks with 51.7% NDS and 40.3% mAP. Code will be\nreleased at https://github.com/ZionGo6/FADet.\n", "date": "2024-05-19", "categories": []} {"id": "2405.12200", "title": "Multi-View Attentive Contextualization for Multi-View 3D Object\n Detection", "abstract": " We present Multi-View Attentive Contextualization (MvACon), a simple yet\neffective method for improving 2D-to-3D feature lifting in query-based\nmulti-view 3D (MV3D) object detection. Despite remarkable progress witnessed in\nthe field of query-based MV3D object detection, prior art often suffers from\neither the lack of exploiting high-resolution 2D features in dense\nattention-based lifting, due to high computational costs, or from\ninsufficiently dense grounding of 3D queries to multi-scale 2D features in\nsparse attention-based lifting. Our proposed MvACon hits the two birds with one\nstone using a representationally dense yet computationally sparse attentive\nfeature contextualization scheme that is agnostic to specific 2D-to-3D feature\nlifting approaches. In experiments, the proposed MvACon is thoroughly tested on\nthe nuScenes benchmark, using both the BEVFormer and its recent 3D deformable\nattention (DFA3D) variant, as well as the PETR, showing consistent detection\nperformance improvement, especially in enhancing performance in location,\norientation, and velocity prediction. It is also tested on the Waymo-mini\nbenchmark using BEVFormer with similar improvement. We qualitatively and\nquantitatively show that global cluster-based contexts effectively encode dense\nscene-level contexts for MV3D object detection. The promising results of our\nproposed MvACon reinforces the adage in computer vision -- ``(contextualized)\nfeature matters\".\n", "date": "2024-05-20", "categories": []} {"id": "2307.10249", "title": "RCM-Fusion: Radar-Camera Multi-Level Fusion for 3D Object Detection", "abstract": " While LiDAR sensors have been successfully applied to 3D object detection,\nthe affordability of radar and camera sensors has led to a growing interest in\nfusing radars and cameras for 3D object detection. However, previous\nradar-camera fusion models were unable to fully utilize the potential of radar\ninformation. In this paper, we propose Radar-Camera Multi-level fusion\n(RCM-Fusion), which attempts to fuse both modalities at both feature and\ninstance levels. For feature-level fusion, we propose a Radar Guided BEV\nEncoder which transforms camera features into precise BEV representations using\nthe guidance of radar Bird's-Eye-View (BEV) features and combines the radar and\ncamera BEV features. For instance-level fusion, we propose a Radar Grid Point\nRefinement module that reduces localization error by accounting for the\ncharacteristics of the radar point clouds. The experiments conducted on the\npublic nuScenes dataset demonstrate that our proposed RCM-Fusion achieves\nstate-of-the-art performances among single frame-based radar-camera fusion\nmethods in the nuScenes 3D object detection benchmark. Code will be made\npublicly available.\n", "date": "2023-07-17", "categories": []} {"id": "2405.06782", "title": "GraphRelate3D: Context-Dependent 3D Object Detection with Inter-Object\n Relationship Graphs", "abstract": " Accurate and effective 3D object detection is critical for ensuring the\ndriving safety of autonomous vehicles. Recently, state-of-the-art two-stage 3D\nobject detectors have exhibited promising performance. However, these methods\nrefine proposals individually, ignoring the rich contextual information in the\nobject relationships between the neighbor proposals. In this study, we\nintroduce an object relation module, consisting of a graph generator and a\ngraph neural network (GNN), to learn the spatial information from certain\npatterns to improve 3D object detection. Specifically, we create an\ninter-object relationship graph based on proposals in a frame via the graph\ngenerator to connect each proposal with its neighbor proposals. Afterward, the\nGNN module extracts edge features from the generated graph and iteratively\nrefines proposal features with the captured edge features. Ultimately, we\nleverage the refined features as input to the detection head to obtain\ndetection results. Our approach improves upon the baseline PV-RCNN on the KITTI\nvalidation set for the car class across easy, moderate, and hard difficulty\nlevels by 0.82%, 0.74%, and 0.58%, respectively. Additionally, our method\noutperforms the baseline by more than 1% under the moderate and hard levels BEV\nAP on the test server.\n", "date": "2024-05-10", "categories": []} {"id": "2301.02650", "title": "Hierarchical Point Attention for Indoor 3D Object Detection", "abstract": " 3D object detection is an essential vision technique for various robotic\nsystems, such as augmented reality and domestic robots. Transformers as\nversatile network architectures have recently seen great success in 3D point\ncloud object detection. However, the lack of hierarchy in a plain transformer\nrestrains its ability to learn features at different scales. Such limitation\nmakes transformer detectors perform worse on smaller objects and affects their\nreliability in indoor environments where small objects are the majority. This\nwork proposes two novel attention operations as generic hierarchical designs\nfor point-based transformer detectors. First, we propose Aggregated Multi-Scale\nAttention (MS-A) that builds multi-scale tokens from a single-scale input\nfeature to enable more fine-grained feature learning. Second, we propose\nSize-Adaptive Local Attention (Local-A) with adaptive attention regions for\nlocalized feature aggregation within bounding box proposals. Both attention\noperations are model-agnostic network modules that can be plugged into existing\npoint cloud transformers for end-to-end training. We evaluate our method on two\nwidely used indoor detection benchmarks. By plugging our proposed modules into\nthe state-of-the-art transformer-based 3D detectors, we improve the previous\nbest results on both benchmarks, with more significant improvements on smaller\nobjects.\n", "date": "2023-01-06", "categories": []} {"id": "2309.14516", "title": "UniBEV: Multi-modal 3D Object Detection with Uniform BEV Encoders for\n Robustness against Missing Sensor Modalities", "abstract": " Multi-sensor object detection is an active research topic in automated\ndriving, but the robustness of such detection models against missing sensor\ninput (modality missing), e.g., due to a sudden sensor failure, is a critical\nproblem which remains under-studied. In this work, we propose UniBEV, an\nend-to-end multi-modal 3D object detection framework designed for robustness\nagainst missing modalities: UniBEV can operate on LiDAR plus camera input, but\nalso on LiDAR-only or camera-only input without retraining. To facilitate its\ndetector head to handle different input combinations, UniBEV aims to create\nwell-aligned Bird's Eye View (BEV) feature maps from each available modality.\nUnlike prior BEV-based multi-modal detection methods, all sensor modalities\nfollow a uniform approach to resample features from the native sensor\ncoordinate systems to the BEV features. We furthermore investigate the\nrobustness of various fusion strategies w.r.t. missing modalities: the commonly\nused feature concatenation, but also channel-wise averaging, and a\ngeneralization to weighted averaging termed Channel Normalized Weights. To\nvalidate its effectiveness, we compare UniBEV to state-of-the-art BEVFusion and\nMetaBEV on nuScenes over all sensor input combinations. In this setting, UniBEV\nachieves $52.5 \\%$ mAP on average over all input combinations, significantly\nimproving over the baselines ($43.5 \\%$ mAP on average for BEVFusion, $48.7 \\%$\nmAP on average for MetaBEV). An ablation study shows the robustness benefits of\nfusing by weighted averaging over regular concatenation, and of sharing queries\nbetween the BEV encoders of each modality. Our code is available at\nhttps://github.com/tudelft-iv/UniBEV.\n", "date": "2023-09-25", "categories": []} {"id": "2405.03884", "title": "BadFusion: 2D-Oriented Backdoor Attacks against 3D Object Detection", "abstract": " 3D object detection plays an important role in autonomous driving; however,\nits vulnerability to backdoor attacks has become evident. By injecting\n''triggers'' to poison the training dataset, backdoor attacks manipulate the\ndetector's prediction for inputs containing these triggers. Existing backdoor\nattacks against 3D object detection primarily poison 3D LiDAR signals, where\nlarge-sized 3D triggers are injected to ensure their visibility within the\nsparse 3D space, rendering them easy to detect and impractical in real-world\nscenarios.\n In this paper, we delve into the robustness of 3D object detection, exploring\na new backdoor attack surface through 2D cameras. Given the prevalent adoption\nof camera and LiDAR signal fusion for high-fidelity 3D perception, we\ninvestigate the latent potential of camera signals to disrupt the process.\nAlthough the dense nature of camera signals enables the use of nearly\nimperceptible small-sized triggers to mislead 2D object detection, realizing\n2D-oriented backdoor attacks against 3D object detection is non-trivial. The\nprimary challenge emerges from the fusion process that transforms camera\nsignals into a 3D space, compromising the association with the 2D trigger to\nthe target output. To tackle this issue, we propose an innovative 2D-oriented\nbackdoor attack against LiDAR-camera fusion methods for 3D object detection,\nnamed BadFusion, for preserving trigger effectiveness throughout the entire\nfusion process. The evaluation demonstrates the effectiveness of BadFusion,\nachieving a significantly higher attack success rate compared to existing\n2D-oriented attacks.\n", "date": "2024-05-06", "categories": []} {"id": "2206.07705", "title": "LET-3D-AP: Longitudinal Error Tolerant 3D Average Precision for\n Camera-Only 3D Detection", "abstract": " The 3D Average Precision (3D AP) relies on the intersection over union\nbetween predictions and ground truth objects. However, camera-only detectors\nhave limited depth accuracy, which may cause otherwise reasonable predictions\nthat suffer from such longitudinal localization errors to be treated as false\npositives. We therefore propose variants of the 3D AP metric to be more\npermissive with respect to depth estimation errors. Specifically, our novel\nlongitudinal error tolerant metrics, LET-3D-AP and LET-3D-APL, allow\nlongitudinal localization errors of the prediction boxes up to a given\ntolerance. To evaluate the proposed metrics, we also construct a new test set\nfor the Waymo Open Dataset, tailored to camera-only 3D detection methods.\nSurprisingly, we find that state-of-the-art camera-based detectors can\noutperform popular LiDAR-based detectors with our new metrics past at 10% depth\nerror tolerance, suggesting that existing camera-based detectors already have\nthe potential to surpass LiDAR-based detectors in downstream applications. We\nbelieve the proposed metrics and the new benchmark dataset will facilitate\nadvances in the field of camera-only 3D detection by providing more informative\nsignals that can better indicate the system-level performance.\n", "date": "2022-06-15", "categories": []} {"id": "2405.02811", "title": "PVTransformer: Point-to-Voxel Transformer for Scalable 3D Object\n Detection", "abstract": " 3D object detectors for point clouds often rely on a pooling-based PointNet\nto encode sparse points into grid-like voxels or pillars. In this paper, we\nidentify that the common PointNet design introduces an information bottleneck\nthat limits 3D object detection accuracy and scalability. To address this\nlimitation, we propose PVTransformer: a transformer-based point-to-voxel\narchitecture for 3D detection. Our key idea is to replace the PointNet pooling\noperation with an attention module, leading to a better point-to-voxel\naggregation function. Our design respects the permutation invariance of sparse\n3D points while being more expressive than the pooling-based PointNet.\nExperimental results show our PVTransformer achieves much better performance\ncompared to the latest 3D object detectors. On the widely used Waymo Open\nDataset, our PVTransformer achieves state-of-the-art 76.5 mAPH L2,\noutperforming the prior art of SWFormer by +1.7 mAPH L2.\n", "date": "2024-05-05", "categories": []} {"id": "2404.19384", "title": "Pseudo Label Refinery for Unsupervised Domain Adaptation on\n Cross-dataset 3D Object Detection", "abstract": " Recent self-training techniques have shown notable improvements in\nunsupervised domain adaptation for 3D object detection (3D UDA). These\ntechniques typically select pseudo labels, i.e., 3D boxes, to supervise models\nfor the target domain. However, this selection process inevitably introduces\nunreliable 3D boxes, in which 3D points cannot be definitively assigned as\nforeground or background. Previous techniques mitigate this by reweighting\nthese boxes as pseudo labels, but these boxes can still poison the training\nprocess. To resolve this problem, in this paper, we propose a novel pseudo\nlabel refinery framework. Specifically, in the selection process, to improve\nthe reliability of pseudo boxes, we propose a complementary augmentation\nstrategy. This strategy involves either removing all points within an\nunreliable box or replacing it with a high-confidence box. Moreover, the point\nnumbers of instances in high-beam datasets are considerably higher than those\nin low-beam datasets, also degrading the quality of pseudo labels during the\ntraining process. We alleviate this issue by generating additional proposals\nand aligning RoI features across different domains. Experimental results\ndemonstrate that our method effectively enhances the quality of pseudo labels\nand consistently surpasses the state-of-the-art methods on six autonomous\ndriving benchmarks. Code will be available at\nhttps://github.com/Zhanwei-Z/PERE.\n", "date": "2024-04-30", "categories": []} {"id": "2304.12310", "title": "Fully Sparse Fusion for 3D Object Detection", "abstract": " Currently prevalent multimodal 3D detection methods are built upon\nLiDAR-based detectors that usually use dense Bird's-Eye-View (BEV) feature\nmaps. However, the cost of such BEV feature maps is quadratic to the detection\nrange, making it not suitable for long-range detection. Fully sparse\narchitecture is gaining attention as they are highly efficient in long-range\nperception. In this paper, we study how to effectively leverage image modality\nin the emerging fully sparse architecture. Particularly, utilizing instance\nqueries, our framework integrates the well-studied 2D instance segmentation\ninto the LiDAR side, which is parallel to the 3D instance segmentation part in\nthe fully sparse detector. This design achieves a uniform query-based fusion\nframework in both the 2D and 3D sides while maintaining the fully sparse\ncharacteristic. Extensive experiments showcase state-of-the-art results on the\nwidely used nuScenes dataset and the long-range Argoverse 2 dataset. Notably,\nthe inference speed of the proposed method under the long-range LiDAR\nperception setting is 2.7 $\\times$ faster than that of other state-of-the-art\nmultimodal 3D detection methods. Code will be released at\n\\url{https://github.com/BraveGroup/FullySparseFusion}.\n", "date": "2023-04-24", "categories": []} {"id": "2404.17910", "title": "Reliable Student: Addressing Noise in Semi-Supervised 3D Object\n Detection", "abstract": " Semi-supervised 3D object detection can benefit from the promising\npseudo-labeling technique when labeled data is limited. However, recent\napproaches have overlooked the impact of noisy pseudo-labels during training,\ndespite efforts to enhance pseudo-label quality through confidence-based\nfiltering. In this paper, we examine the impact of noisy pseudo-labels on\nIoU-based target assignment and propose the Reliable Student framework, which\nincorporates two complementary approaches to mitigate errors. First, it\ninvolves a class-aware target assignment strategy that reduces false negative\nassignments in difficult classes. Second, it includes a reliability weighting\nstrategy that suppresses false positive assignment errors while also addressing\nremaining false negatives from the first step. The reliability weights are\ndetermined by querying the teacher network for confidence scores of the\nstudent-generated proposals. Our work surpasses the previous state-of-the-art\non KITTI 3D object detection benchmark on point clouds in the semi-supervised\nsetting. On 1% labeled data, our approach achieves a 6.2% AP improvement for\nthe pedestrian class, despite having only 37 labeled samples available. The\nimprovements become significant for the 2% setting, achieving 6.0% AP and 5.7%\nAP improvements for the pedestrian and cyclist classes, respectively.\n", "date": "2024-04-27", "categories": []} {"id": "2404.16548", "title": "Cross-Domain Spatial Matching for Camera and Radar Sensor Data Fusion in\n Autonomous Vehicle Perception System", "abstract": " In this paper, we propose a novel approach to address the problem of camera\nand radar sensor fusion for 3D object detection in autonomous vehicle\nperception systems. Our approach builds on recent advances in deep learning and\nleverages the strengths of both sensors to improve object detection\nperformance. Precisely, we extract 2D features from camera images using a\nstate-of-the-art deep learning architecture and then apply a novel Cross-Domain\nSpatial Matching (CDSM) transformation method to convert these features into 3D\nspace. We then fuse them with extracted radar data using a complementary fusion\nstrategy to produce a final 3D object representation. To demonstrate the\neffectiveness of our approach, we evaluate it on the NuScenes dataset. We\ncompare our approach to both single-sensor performance and current\nstate-of-the-art fusion methods. Our results show that the proposed approach\nachieves superior performance over single-sensor solutions and could directly\ncompete with other top-level fusion methods.\n", "date": "2024-04-25", "categories": []} {"id": "2312.08371", "title": "PTT: Point-Trajectory Transformer for Efficient Temporal 3D Object\n Detection", "abstract": " Recent temporal LiDAR-based 3D object detectors achieve promising performance\nbased on the two-stage proposal-based approach. They generate 3D box candidates\nfrom the first-stage dense detector, followed by different temporal aggregation\nmethods. However, these approaches require per-frame objects or whole point\nclouds, posing challenges related to memory bank utilization. Moreover, point\nclouds and trajectory features are combined solely based on concatenation,\nwhich may neglect effective interactions between them. In this paper, we\npropose a point-trajectory transformer with long short-term memory for\nefficient temporal 3D object detection. To this end, we only utilize point\nclouds of current-frame objects and their historical trajectories as input to\nminimize the memory bank storage requirement. Furthermore, we introduce modules\nto encode trajectory features, focusing on long short-term and future-aware\nperspectives, and then effectively aggregate them with point cloud features. We\nconduct extensive experiments on the large-scale Waymo dataset to demonstrate\nthat our approach performs well against state-of-the-art methods. Code and\nmodels will be made publicly available at https://github.com/kuanchihhuang/PTT.\n", "date": "2023-12-13", "categories": []} {"id": "2404.15879", "title": "Revisiting Out-of-Distribution Detection in LiDAR-based 3D Object\n Detection", "abstract": " LiDAR-based 3D object detection has become an essential part of automated\ndriving due to its ability to localize and classify objects precisely in 3D.\nHowever, object detectors face a critical challenge when dealing with unknown\nforeground objects, particularly those that were not present in their original\ntraining data. These out-of-distribution (OOD) objects can lead to\nmisclassifications, posing a significant risk to the safety and reliability of\nautomated vehicles. Currently, LiDAR-based OOD object detection has not been\nwell studied. We address this problem by generating synthetic training data for\nOOD objects by perturbing known object categories. Our idea is that these\nsynthetic OOD objects produce different responses in the feature map of an\nobject detector compared to in-distribution (ID) objects. We then extract\nfeatures using a pre-trained and fixed object detector and train a simple\nmultilayer perceptron (MLP) to classify each detection as either ID or OOD. In\naddition, we propose a new evaluation protocol that allows the use of existing\ndatasets without modifying the point cloud, ensuring a more authentic\nevaluation of real-world scenarios. The effectiveness of our method is\nvalidated through experiments on the newly proposed nuScenes OOD benchmark. The\nsource code is available at https://github.com/uulm-mrm/mmood3d.\n", "date": "2024-04-24", "categories": []} {"id": "2401.03907", "title": "RoboFusion: Towards Robust Multi-Modal 3D Object Detection via SAM", "abstract": " Multi-modal 3D object detectors are dedicated to exploring secure and\nreliable perception systems for autonomous driving (AD).Although achieving\nstate-of-the-art (SOTA) performance on clean benchmark datasets, they tend to\noverlook the complexity and harsh conditions of real-world environments. With\nthe emergence of visual foundation models (VFMs), opportunities and challenges\nare presented for improving the robustness and generalization of multi-modal 3D\nobject detection in AD. Therefore, we propose RoboFusion, a robust framework\nthat leverages VFMs like SAM to tackle out-of-distribution (OOD) noise\nscenarios. We first adapt the original SAM for AD scenarios named SAM-AD. To\nalign SAM or SAM-AD with multi-modal methods, we then introduce AD-FPN for\nupsampling the image features extracted by SAM. We employ wavelet decomposition\nto denoise the depth-guided images for further noise reduction and weather\ninterference. At last, we employ self-attention mechanisms to adaptively\nreweight the fused features, enhancing informative features while suppressing\nexcess noise. In summary, RoboFusion significantly reduces noise by leveraging\nthe generalization and robustness of VFMs, thereby enhancing the resilience of\nmulti-modal 3D object detection. Consequently, RoboFusion achieves SOTA\nperformance in noisy scenarios, as demonstrated by the KITTI-C and nuScenes-C\nbenchmarks. Code is available at https://github.com/adept-thu/RoboFusion.\n", "date": "2024-01-08", "categories": []} {"id": "2402.11677", "title": "MultiCorrupt: A Multi-Modal Robustness Dataset and Benchmark of\n LiDAR-Camera Fusion for 3D Object Detection", "abstract": " Multi-modal 3D object detection models for automated driving have\ndemonstrated exceptional performance on computer vision benchmarks like\nnuScenes. However, their reliance on densely sampled LiDAR point clouds and\nmeticulously calibrated sensor arrays poses challenges for real-world\napplications. Issues such as sensor misalignment, miscalibration, and disparate\nsampling frequencies lead to spatial and temporal misalignment in data from\nLiDAR and cameras. Additionally, the integrity of LiDAR and camera data is\noften compromised by adverse environmental conditions such as inclement\nweather, leading to occlusions and noise interference. To address this\nchallenge, we introduce MultiCorrupt, a comprehensive benchmark designed to\nevaluate the robustness of multi-modal 3D object detectors against ten distinct\ntypes of corruptions. We evaluate five state-of-the-art multi-modal detectors\non MultiCorrupt and analyze their performance in terms of their resistance\nability. Our results show that existing methods exhibit varying degrees of\nrobustness depending on the type of corruption and their fusion strategy. We\nprovide insights into which multi-modal design choices make such models robust\nagainst certain perturbations. The dataset generation code and benchmark are\nopen-sourced at https://github.com/ika-rwth-aachen/MultiCorrupt.\n", "date": "2024-02-18", "categories": []} {"id": "2404.11737", "title": "Equivariant Spatio-Temporal Self-Supervision for LiDAR Object Detection", "abstract": " Popular representation learning methods encourage feature invariance under\ntransformations applied at the input. However, in 3D perception tasks like\nobject localization and segmentation, outputs are naturally equivariant to some\ntransformations, such as rotation. Using pre-training loss functions that\nencourage equivariance of features under certain transformations provides a\nstrong self-supervision signal while also retaining information of geometric\nrelationships between transformed feature representations. This can enable\nimproved performance in downstream tasks that are equivariant to such\ntransformations. In this paper, we propose a spatio-temporal equivariant\nlearning framework by considering both spatial and temporal augmentations\njointly. Our experiments show that the best performance arises with a\npre-training approach that encourages equivariance to translation, scaling, and\nflip, rotation and scene flow. For spatial augmentations, we find that\ndepending on the transformation, either a contrastive objective or an\nequivariance-by-classification objective yields best results. To leverage\nreal-world object deformations and motion, we consider sequential LiDAR scene\npairs and develop a novel 3D scene flow-based equivariance objective that leads\nto improved performance overall. We show our pre-training method for 3D object\ndetection which outperforms existing equivariant and invariant approaches in\nmany settings.\n", "date": "2024-04-17", "categories": []} {"id": "2404.11764", "title": "Multimodal 3D Object Detection on Unseen Domains", "abstract": " LiDAR datasets for autonomous driving exhibit biases in properties such as\npoint cloud density, range, and object dimensions. As a result, object\ndetection networks trained and evaluated in different environments often\nexperience performance degradation. Domain adaptation approaches assume access\nto unannotated samples from the test distribution to address this problem.\nHowever, in the real world, the exact conditions of deployment and access to\nsamples representative of the test dataset may be unavailable while training.\nWe argue that the more realistic and challenging formulation is to require\nrobustness in performance to unseen target domains. We propose to address this\nproblem in a two-pronged manner. First, we leverage paired LiDAR-image data\npresent in most autonomous driving datasets to perform multimodal object\ndetection. We suggest that working with multimodal features by leveraging both\nimages and LiDAR point clouds for scene understanding tasks results in object\ndetectors more robust to unseen domain shifts. Second, we train a 3D object\ndetector to learn multimodal object features across different distributions and\npromote feature invariance across these source domains to improve\ngeneralizability to unseen target domains. To this end, we propose\nCLIX$^\\text{3D}$, a multimodal fusion and supervised contrastive learning\nframework for 3D object detection that performs alignment of object features\nfrom same-class samples of different domains while pushing the features from\ndifferent classes apart. We show that CLIX$^\\text{3D}$ yields state-of-the-art\ndomain generalization performance under multiple dataset shifts.\n", "date": "2024-04-17", "categories": []} {"id": "2404.10978", "title": "Leveraging 3D LiDAR Sensors to Enable Enhanced Urban Safety and Public\n Health: Pedestrian Monitoring and Abnormal Activity Detection", "abstract": " The integration of Light Detection and Ranging (LiDAR) and Internet of Things\n(IoT) technologies offers transformative opportunities for public health\ninformatics in urban safety and pedestrian well-being. This paper proposes a\nnovel framework utilizing these technologies for enhanced 3D object detection\nand activity classification in urban traffic scenarios. By employing elevated\nLiDAR, we obtain detailed 3D point cloud data, enabling precise pedestrian\nactivity monitoring. To overcome urban data scarcity, we create a specialized\ndataset through simulated traffic environments in Blender, facilitating\ntargeted model training. Our approach employs a modified Point\nVoxel-Region-based Convolutional Neural Network (PV-RCNN) for robust 3D\ndetection and PointNet for classifying pedestrian activities, significantly\nbenefiting urban traffic management and public health by offering insights into\npedestrian behavior and promoting safer urban environments. Our dual-model\napproach not only enhances urban traffic management but also contributes\nsignificantly to public health by providing insights into pedestrian behavior\nand promoting safer urban environment.\n", "date": "2024-04-17", "categories": []} {"id": "2308.06791", "title": "PV-SSD: A Multi-Modal Point Cloud Feature Fusion Method for Projection\n Features and Variable Receptive Field Voxel Features", "abstract": " LiDAR-based 3D object detection and classification is crucial for autonomous\ndriving. However, real-time inference from extremely sparse 3D data is a\nformidable challenge. To address this problem, a typical class of approaches\ntransforms the point cloud cast into a regular data representation (voxels or\nprojection maps). Then, it performs feature extraction with convolutional\nneural networks. However, such methods often result in a certain degree of\ninformation loss due to down-sampling or over-compression of feature\ninformation. This paper proposes a multi-modal point cloud feature fusion\nmethod for projection features and variable receptive field voxel features\n(PV-SSD) based on projection and variable voxelization to solve the information\nloss problem. We design a two-branch feature extraction structure with a 2D\nconvolutional neural network to extract the point cloud's projection features\nin bird's-eye view to focus on the correlation between local features. A voxel\nfeature extraction branch is used to extract local fine-grained features.\nMeanwhile, we propose a voxel feature extraction method with variable sensory\nfields to reduce the information loss of voxel branches due to downsampling. It\navoids missing critical point information by selecting more useful feature\npoints based on feature point weights for the detection task. In addition, we\npropose a multi-modal feature fusion module for point clouds. To validate the\neffectiveness of our method, we tested it on the KITTI dataset and ONCE\ndataset.\n", "date": "2023-08-13", "categories": []} {"id": "2404.05139", "title": "Better Monocular 3D Detectors with LiDAR from the Past", "abstract": " Accurate 3D object detection is crucial to autonomous driving. Though\nLiDAR-based detectors have achieved impressive performance, the high cost of\nLiDAR sensors precludes their widespread adoption in affordable vehicles.\nCamera-based detectors are cheaper alternatives but often suffer inferior\nperformance compared to their LiDAR-based counterparts due to inherent depth\nambiguities in images. In this work, we seek to improve monocular 3D detectors\nby leveraging unlabeled historical LiDAR data. Specifically, at inference time,\nwe assume that the camera-based detectors have access to multiple unlabeled\nLiDAR scans from past traversals at locations of interest (potentially from\nother high-end vehicles equipped with LiDAR sensors). Under this setup, we\nproposed a novel, simple, and end-to-end trainable framework, termed\nAsyncDepth, to effectively extract relevant features from asynchronous LiDAR\ntraversals of the same location for monocular 3D detectors. We show consistent\nand significant performance gain (up to 9 AP) across multiple state-of-the-art\nmodels and datasets with a negligible additional latency of 9.66 ms and a small\nstorage cost.\n", "date": "2024-04-08", "categories": []} {"id": "2404.06700", "title": "Scaling Multi-Camera 3D Object Detection through Weak-to-Strong\n Eliciting", "abstract": " The emergence of Multi-Camera 3D Object Detection (MC3D-Det), facilitated by\nbird's-eye view (BEV) representation, signifies a notable progression in 3D\nobject detection. Scaling MC3D-Det training effectively accommodates varied\ncamera parameters and urban landscapes, paving the way for the MC3D-Det\nfoundation model. However, the multi-view fusion stage of the MC3D-Det method\nrelies on the ill-posed monocular perception during training rather than\nsurround refinement ability, leading to what we term \"surround refinement\ndegradation\". To this end, our study presents a weak-to-strong eliciting\nframework aimed at enhancing surround refinement while maintaining robust\nmonocular perception. Specifically, our framework employs weakly tuned experts\ntrained on distinct subsets, and each is inherently biased toward specific\ncamera configurations and scenarios. These biased experts can learn the\nperception of monocular degeneration, which can help the multi-view fusion\nstage to enhance surround refinement abilities. Moreover, a composite\ndistillation strategy is proposed to integrate the universal knowledge of 2D\nfoundation models and task-specific information. Finally, for MC3D-Det joint\ntraining, the elaborate dataset merge strategy is designed to solve the problem\nof inconsistent camera numbers and camera parameters. We set up a multiple\ndataset joint training benchmark for MC3D-Det and adequately evaluated existing\nmethods. Further, we demonstrate the proposed framework brings a generalized\nand significant boost over multiple baselines. Our code is at\n\\url{https://github.com/EnVision-Research/Scale-BEV}.\n", "date": "2024-04-10", "categories": []} {"id": "2404.06715", "title": "Sparse Points to Dense Clouds: Enhancing 3D Detection with Limited LiDAR\n Data", "abstract": " 3D detection is a critical task that enables machines to identify and locate\nobjects in three-dimensional space. It has a broad range of applications in\nseveral fields, including autonomous driving, robotics and augmented reality.\nMonocular 3D detection is attractive as it requires only a single camera,\nhowever, it lacks the accuracy and robustness required for real world\napplications. High resolution LiDAR on the other hand, can be expensive and\nlead to interference problems in heavy traffic given their active\ntransmissions. We propose a balanced approach that combines the advantages of\nmonocular and point cloud-based 3D detection. Our method requires only a small\nnumber of 3D points, that can be obtained from a low-cost, low-resolution\nsensor. Specifically, we use only 512 points, which is just 1% of a full LiDAR\nframe in the KITTI dataset. Our method reconstructs a complete 3D point cloud\nfrom this limited 3D information combined with a single image. The\nreconstructed 3D point cloud and corresponding image can be used by any\nmulti-modal off-the-shelf detector for 3D object detection. By using the\nproposed network architecture with an off-the-shelf multi-modal 3D detector,\nthe accuracy of 3D detection improves by 20% compared to the state-of-the-art\nmonocular detection methods and 6% to 9% compare to the baseline multi-modal\nmethods on KITTI and JackRabbot datasets.\n", "date": "2024-04-10", "categories": []} {"id": "2303.05970", "title": "Exploring Recurrent Long-term Temporal Fusion for Multi-view 3D\n Perception", "abstract": " Long-term temporal fusion is a crucial but often overlooked technique in\ncamera-based Bird's-Eye-View (BEV) 3D perception. Existing methods are mostly\nin a parallel manner. While parallel fusion can benefit from long-term\ninformation, it suffers from increasing computational and memory overheads as\nthe fusion window size grows. Alternatively, BEVFormer adopts a recurrent\nfusion pipeline so that history information can be efficiently integrated, yet\nit fails to benefit from longer temporal frames. In this paper, we explore an\nembarrassingly simple long-term recurrent fusion strategy built upon the\nLSS-based methods and find it already able to enjoy the merits from both sides,\ni.e., rich long-term information and efficient fusion pipeline. A temporal\nembedding module is further proposed to improve the model's robustness against\noccasionally missed frames in practical scenarios. We name this simple but\neffective fusing pipeline VideoBEV. Experimental results on the nuScenes\nbenchmark show that VideoBEV obtains strong performance on various camera-based\n3D perception tasks, including object detection (55.4\\% mAP and 62.9\\% NDS),\nsegmentation (48.6\\% vehicle mIoU), tracking (54.8\\% AMOTA), and motion\nprediction (0.80m minADE and 0.463 EPA).\n", "date": "2023-03-10", "categories": []} {"id": "2403.04198", "title": "CN-RMA: Combined Network with Ray Marching Aggregation for 3D Indoors\n Object Detection from Multi-view Images", "abstract": " This paper introduces CN-RMA, a novel approach for 3D indoor object detection\nfrom multi-view images. We observe the key challenge as the ambiguity of image\nand 3D correspondence without explicit geometry to provide occlusion\ninformation. To address this issue, CN-RMA leverages the synergy of 3D\nreconstruction networks and 3D object detection networks, where the\nreconstruction network provides a rough Truncated Signed Distance Function\n(TSDF) and guides image features to vote to 3D space correctly in an end-to-end\nmanner. Specifically, we associate weights to sampled points of each ray\nthrough ray marching, representing the contribution of a pixel in an image to\ncorresponding 3D locations. Such weights are determined by the predicted signed\ndistances so that image features vote only to regions near the reconstructed\nsurface. Our method achieves state-of-the-art performance in 3D object\ndetection from multi-view images, as measured by mAP@0.25 and mAP@0.5 on the\nScanNet and ARKitScenes datasets. The code and models are released at\nhttps://github.com/SerCharles/CN-RMA.\n", "date": "2024-03-07", "categories": []} {"id": "2404.06256", "title": "Label-Efficient 3D Object Detection For Road-Side Units", "abstract": " Occlusion presents a significant challenge for safety-critical applications\nsuch as autonomous driving. Collaborative perception has recently attracted a\nlarge research interest thanks to the ability to enhance the perception of\nautonomous vehicles via deep information fusion with intelligent roadside units\n(RSU), thus minimizing the impact of occlusion. While significant advancement\nhas been made, the data-hungry nature of these methods creates a major hurdle\nfor their real-world deployment, particularly due to the need for annotated RSU\ndata. Manually annotating the vast amount of RSU data required for training is\nprohibitively expensive, given the sheer number of intersections and the effort\ninvolved in annotating point clouds. We address this challenge by devising a\nlabel-efficient object detection method for RSU based on unsupervised object\ndiscovery. Our paper introduces two new modules: one for object discovery based\non a spatial-temporal aggregation of point clouds, and another for refinement.\nFurthermore, we demonstrate that fine-tuning on a small portion of annotated\ndata allows our object discovery models to narrow the performance gap with, or\neven surpass, fully supervised models. Extensive experiments are carried out in\nsimulated and real-world datasets to evaluate our method.\n", "date": "2024-04-09", "categories": []} {"id": "2402.15756", "title": "Detection Is Tracking: Point Cloud Multi-Sweep Deep Learning Models\n Revisited", "abstract": " Conventional tracking paradigm takes in instantaneous measurements such as\nrange and bearing, and produces object tracks across time. In applications such\nas autonomous driving, lidar measurements in the form of point clouds are\nusually passed through a \"virtual sensor\" realized by a deep learning model, to\nproduce \"measurements\" such as bounding boxes, which are in turn ingested by a\ntracking module to produce object tracks. Very often multiple lidar sweeps are\naccumulated in a buffer to merge and become the input to the virtual sensor. We\nargue in this paper that such an input already contains temporal information,\nand therefore the virtual sensor output should also contain temporal\ninformation, not just instantaneous values for the time corresponding to the\nend of the buffer. In particular, we present the deep learning model called\nMULti-Sweep PAired Detector (MULSPAD) that produces, for each detected object,\na pair of bounding boxes at both the end time and the beginning time of the\ninput buffer. This is achieved with fairly straightforward changes in commonly\nused lidar detection models, and with only marginal extra processing, but the\nresulting symmetry is satisfying. Such paired detections make it possible not\nonly to construct rudimentary trackers fairly easily, but also to construct\nmore sophisticated trackers that can exploit the extra information conveyed by\nthe pair and be robust to choices of motion models and object birth/death\nmodels. We have conducted preliminary training and experimentation using Waymo\nOpen Dataset, which shows the efficacy of our proposed method.\n", "date": "2024-02-24", "categories": []} {"id": "2404.05280", "title": "MOSE: Boosting Vision-based Roadside 3D Object Detection with Scene Cues", "abstract": " 3D object detection based on roadside cameras is an additional way for\nautonomous driving to alleviate the challenges of occlusion and short\nperception range from vehicle cameras. Previous methods for roadside 3D object\ndetection mainly focus on modeling the depth or height of objects, neglecting\nthe stationary of cameras and the characteristic of inter-frame consistency. In\nthis work, we propose a novel framework, namely MOSE, for MOnocular 3D object\ndetection with Scene cuEs. The scene cues are the frame-invariant\nscene-specific features, which are crucial for object localization and can be\nintuitively regarded as the height between the surface of the real road and the\nvirtual ground plane. In the proposed framework, a scene cue bank is designed\nto aggregate scene cues from multiple frames of the same scene with a carefully\ndesigned extrinsic augmentation strategy. Then, a transformer-based decoder\nlifts the aggregated scene cues as well as the 3D position embeddings for 3D\nobject location, which boosts generalization ability in heterologous scenes.\nThe extensive experiment results on two public benchmarks demonstrate the\nstate-of-the-art performance of the proposed method, which surpasses the\nexisting methods by a large margin.\n", "date": "2024-04-08", "categories": []} {"id": "2404.03181", "title": "MonoCD: Monocular 3D Object Detection with Complementary Depths", "abstract": " Monocular 3D object detection has attracted widespread attention due to its\npotential to accurately obtain object 3D localization from a single image at a\nlow cost. Depth estimation is an essential but challenging subtask of monocular\n3D object detection due to the ill-posedness of 2D to 3D mapping. Many methods\nexplore multiple local depth clues such as object heights and keypoints and\nthen formulate the object depth estimation as an ensemble of multiple depth\npredictions to mitigate the insufficiency of single-depth information. However,\nthe errors of existing multiple depths tend to have the same sign, which\nhinders them from neutralizing each other and limits the overall accuracy of\ncombined depth. To alleviate this problem, we propose to increase the\ncomplementarity of depths with two novel designs. First, we add a new depth\nprediction branch named complementary depth that utilizes global and efficient\ndepth clues from the entire image rather than the local clues to reduce the\ncorrelation of depth predictions. Second, we propose to fully exploit the\ngeometric relations between multiple depth clues to achieve complementarity in\nform. Benefiting from these designs, our method achieves higher\ncomplementarity. Experiments on the KITTI benchmark demonstrate that our method\nachieves state-of-the-art performance without introducing extra data. In\naddition, complementary depth can also be a lightweight and plug-and-play\nmodule to boost multiple existing monocular 3d object detectors. Code is\navailable at https://github.com/elvintanhust/MonoCD.\n", "date": "2024-04-04", "categories": []} {"id": "2404.01580", "title": "Learning Temporal Cues by Predicting Objects Move for Multi-camera 3D\n Object Detection", "abstract": " In autonomous driving and robotics, there is a growing interest in utilizing\nshort-term historical data to enhance multi-camera 3D object detection,\nleveraging the continuous and correlated nature of input video streams. Recent\nwork has focused on spatially aligning BEV-based features over timesteps.\nHowever, this is often limited as its gain does not scale well with long-term\npast observations. To address this, we advocate for supervising a model to\npredict objects' poses given past observations, thus explicitly guiding to\nlearn objects' temporal cues. To this end, we propose a model called DAP\n(Detection After Prediction), consisting of a two-branch network: (i) a branch\nresponsible for forecasting the current objects' poses given past observations\nand (ii) another branch that detects objects based on the current and past\nobservations. The features predicting the current objects from branch (i) is\nfused into branch (ii) to transfer predictive knowledge. We conduct extensive\nexperiments with the large-scale nuScenes datasets, and we observe that\nutilizing such predictive information significantly improves the overall\ndetection performance. Our model can be used plug-and-play, showing consistent\nperformance gain.\n", "date": "2024-04-02", "categories": []} {"id": "2404.00149", "title": "VSRD: Instance-Aware Volumetric Silhouette Rendering for Weakly\n Supervised 3D Object Detection", "abstract": " Monocular 3D object detection poses a significant challenge in 3D scene\nunderstanding due to its inherently ill-posed nature in monocular depth\nestimation. Existing methods heavily rely on supervised learning using abundant\n3D labels, typically obtained through expensive and labor-intensive annotation\non LiDAR point clouds. To tackle this problem, we propose a novel weakly\nsupervised 3D object detection framework named VSRD (Volumetric Silhouette\nRendering for Detection) to train 3D object detectors without any 3D\nsupervision but only weak 2D supervision. VSRD consists of multi-view 3D\nauto-labeling and subsequent training of monocular 3D object detectors using\nthe pseudo labels generated in the auto-labeling stage. In the auto-labeling\nstage, we represent the surface of each instance as a signed distance field\n(SDF) and render its silhouette as an instance mask through our proposed\ninstance-aware volumetric silhouette rendering. To directly optimize the 3D\nbounding boxes through rendering, we decompose the SDF of each instance into\nthe SDF of a cuboid and the residual distance field (RDF) that represents the\nresidual from the cuboid. This mechanism enables us to optimize the 3D bounding\nboxes in an end-to-end manner by comparing the rendered instance masks with the\nground truth instance masks. The optimized 3D bounding boxes serve as effective\ntraining data for 3D object detection. We conduct extensive experiments on the\nKITTI-360 dataset, demonstrating that our method outperforms the existing\nweakly supervised 3D object detection methods. The code is available at\nhttps://github.com/skmhrk1209/VSRD.\n", "date": "2024-03-29", "categories": []} {"id": "2404.00364", "title": "Accurate Cutting-point Estimation for Robotic Lychee Harvesting through\n Geometry-aware Learning", "abstract": " Accurately identifying lychee-picking points in unstructured orchard\nenvironments and obtaining their coordinate locations is critical to the\nsuccess of lychee-picking robots. However, traditional two-dimensional (2D)\nimage-based object detection methods often struggle due to the complex\ngeometric structures of branches, leaves and fruits, leading to incorrect\ndetermination of lychee picking points. In this study, we propose a\nFcaf3d-lychee network model specifically designed for the accurate localisation\nof lychee picking points. Point cloud data of lychee picking points in natural\nenvironments are acquired using Microsoft's Azure Kinect DK time-of-flight\n(TOF) camera through multi-view stitching. We augment the Fully Convolutional\nAnchor-Free 3D Object Detection (Fcaf3d) model with a\nsqueeze-and-excitation(SE) module, which exploits human visual attention\nmechanisms for improved feature extraction of lychee picking points. The\ntrained network model is evaluated on a test set of lychee-picking locations\nand achieves an impressive F1 score of 88.57%, significantly outperforming\nexisting models. Subsequent three-dimensional (3D) position detection of\npicking points in real lychee orchard environments yields high accuracy, even\nunder varying degrees of occlusion. Localisation errors of lychee picking\npoints are within 1.5 cm in all directions, demonstrating the robustness and\ngenerality of the model.\n", "date": "2024-03-30", "categories": []} {"id": "2404.00679", "title": "Weak-to-Strong 3D Object Detection with X-Ray Distillation", "abstract": " This paper addresses the critical challenges of sparsity and occlusion in\nLiDAR-based 3D object detection. Current methods often rely on supplementary\nmodules or specific architectural designs, potentially limiting their\napplicability to new and evolving architectures. To our knowledge, we are the\nfirst to propose a versatile technique that seamlessly integrates into any\nexisting framework for 3D Object Detection, marking the first instance of\nWeak-to-Strong generalization in 3D computer vision. We introduce a novel\nframework, X-Ray Distillation with Object-Complete Frames, suitable for both\nsupervised and semi-supervised settings, that leverages the temporal aspect of\npoint cloud sequences. This method extracts crucial information from both\nprevious and subsequent LiDAR frames, creating Object-Complete frames that\nrepresent objects from multiple viewpoints, thus addressing occlusion and\nsparsity. Given the limitation of not being able to generate Object-Complete\nframes during online inference, we utilize Knowledge Distillation within a\nTeacher-Student framework. This technique encourages the strong Student model\nto emulate the behavior of the weaker Teacher, which processes simple and\ninformative Object-Complete frames, effectively offering a comprehensive view\nof objects as if seen through X-ray vision. Our proposed methods surpass\nstate-of-the-art in semi-supervised learning by 1-1.5 mAP and enhance the\nperformance of five established supervised models by 1-2 mAP on standard\nautonomous driving datasets, even with default hyperparameters. Code for\nObject-Complete frames is available here:\nhttps://github.com/sakharok13/X-Ray-Teacher-Patching-Tools.\n", "date": "2024-03-31", "categories": []} {"id": "2403.20318", "title": "SeaBird: Segmentation in Bird's View with Dice Loss Improves Monocular\n 3D Detection of Large Objects", "abstract": " Monocular 3D detectors achieve remarkable performance on cars and smaller\nobjects. However, their performance drops on larger objects, leading to fatal\naccidents. Some attribute the failures to training data scarcity or their\nreceptive field requirements of large objects. In this paper, we highlight this\nunderstudied problem of generalization to large objects. We find that modern\nfrontal detectors struggle to generalize to large objects even on nearly\nbalanced datasets. We argue that the cause of failure is the sensitivity of\ndepth regression losses to noise of larger objects. To bridge this gap, we\ncomprehensively investigate regression and dice losses, examining their\nrobustness under varying error levels and object sizes. We mathematically prove\nthat the dice loss leads to superior noise-robustness and model convergence for\nlarge objects compared to regression losses for a simplified case. Leveraging\nour theoretical insights, we propose SeaBird (Segmentation in Bird's View) as\nthe first step towards generalizing to large objects. SeaBird effectively\nintegrates BEV segmentation on foreground objects for 3D detection, with the\nsegmentation head trained with the dice loss. SeaBird achieves SoTA results on\nthe KITTI-360 leaderboard and improves existing detectors on the nuScenes\nleaderboard, particularly for large objects. Code and models at\nhttps://github.com/abhi1kumar/SeaBird\n", "date": "2024-03-29", "categories": []} {"id": "2403.19104", "title": "CRKD: Enhanced Camera-Radar Object Detection with Cross-modality\n Knowledge Distillation", "abstract": " In the field of 3D object detection for autonomous driving, LiDAR-Camera (LC)\nfusion is the top-performing sensor configuration. Still, LiDAR is relatively\nhigh cost, which hinders adoption of this technology for consumer automobiles.\nAlternatively, camera and radar are commonly deployed on vehicles already on\nthe road today, but performance of Camera-Radar (CR) fusion falls behind LC\nfusion. In this work, we propose Camera-Radar Knowledge Distillation (CRKD) to\nbridge the performance gap between LC and CR detectors with a novel\ncross-modality KD framework. We use the Bird's-Eye-View (BEV) representation as\nthe shared feature space to enable effective knowledge distillation. To\naccommodate the unique cross-modality KD path, we propose four distillation\nlosses to help the student learn crucial features from the teacher model. We\npresent extensive evaluations on the nuScenes dataset to demonstrate the\neffectiveness of the proposed CRKD framework. The project page for CRKD is\nhttps://song-jingyu.github.io/CRKD.\n", "date": "2024-03-28", "categories": []} {"id": "2211.17126", "title": "BEVUDA: Multi-geometric Space Alignments for Domain Adaptive BEV 3D\n Object Detection", "abstract": " Vision-centric bird-eye-view (BEV) perception has shown promising potential\nin autonomous driving. Recent works mainly focus on improving efficiency or\naccuracy but neglect the challenges when facing environment changing, resulting\nin severe degradation of transfer performance. For BEV perception, we figure\nout the significant domain gaps existing in typical real-world cross-domain\nscenarios and comprehensively solve the Domain Adaption (DA) problem for\nmulti-view 3D object detection. Since BEV perception approaches are complicated\nand contain several components, the domain shift accumulation on multiple\ngeometric spaces (i.e., 2D, 3D Voxel, BEV) makes BEV DA even challenging. In\nthis paper, we propose a Multi-space Alignment Teacher-Student (MATS) framework\nto ease the domain shift accumulation, which consists of a Depth-Aware Teacher\n(DAT) and a Geometric-space Aligned Student (GAS) model. DAT tactfully combines\ntarget lidar and reliable depth prediction to construct depth-aware\ninformation, extracting target domain-specific knowledge in Voxel and BEV\nfeature spaces. It then transfers the sufficient domain knowledge of multiple\nspaces to the student model. In order to jointly alleviate the domain shift,\nGAS projects multi-geometric space features to a shared geometric embedding\nspace and decreases data distribution distance between two domains. To verify\nthe effectiveness of our method, we conduct BEV 3D object detection experiments\non three cross-domain scenarios and achieve state-of-the-art performance.\n", "date": "2022-11-30", "categories": []} {"id": "2403.02640", "title": "HoloVIC: Large-scale Dataset and Benchmark for Multi-Sensor Holographic\n Intersection and Vehicle-Infrastructure Cooperative", "abstract": " Vehicle-to-everything (V2X) is a popular topic in the field of Autonomous\nDriving in recent years. Vehicle-infrastructure cooperation (VIC) becomes one\nof the important research area. Due to the complexity of traffic conditions\nsuch as blind spots and occlusion, it greatly limits the perception\ncapabilities of single-view roadside sensing systems. To further enhance the\naccuracy of roadside perception and provide better information to the vehicle\nside, in this paper, we constructed holographic intersections with various\nlayouts to build a large-scale multi-sensor holographic vehicle-infrastructure\ncooperation dataset, called HoloVIC. Our dataset includes 3 different types of\nsensors (Camera, Lidar, Fisheye) and employs 4 sensor-layouts based on the\ndifferent intersections. Each intersection is equipped with 6-18 sensors to\ncapture synchronous data. While autonomous vehicles pass through these\nintersections for collecting VIC data. HoloVIC contains in total on 100k+\nsynchronous frames from different sensors. Additionally, we annotated 3D\nbounding boxes based on Camera, Fisheye, and Lidar. We also associate the IDs\nof the same objects across different devices and consecutive frames in\nsequence. Based on HoloVIC, we formulated four tasks to facilitate the\ndevelopment of related research. We also provide benchmarks for these tasks.\n", "date": "2024-03-05", "categories": []} {"id": "2403.17390", "title": "SSF3D: Strict Semi-Supervised 3D Object Detection with Switching Filter", "abstract": " SSF3D modified the semi-supervised 3D object detection (SS3DOD) framework,\nwhich designed specifically for point cloud data. Leveraging the\ncharacteristics of non-coincidence and weak correlation of target objects in\npoint cloud, we adopt a strategy of retaining only the truth-determining pseudo\nlabels and trimming the other fuzzy labels with points, instead of pursuing a\nbalance between the quantity and quality of pseudo labels. Besides, we notice\nthat changing the filter will make the model meet different distributed\ntargets, which is beneficial to break the training bottleneck. Two mechanism\nare introduced to achieve above ideas: strict threshold and filter switching.\nThe experiments are conducted to analyze the effectiveness of above approaches\nand their impact on the overall performance of the system. Evaluating on the\nKITTI dataset, SSF3D exhibits superior performance compared to the current\nstate-of-the-art methods. The code will be released here.\n", "date": "2024-03-26", "categories": []} {"id": "2312.01696", "title": "BEVNeXt: Reviving Dense BEV Frameworks for 3D Object Detection", "abstract": " Recently, the rise of query-based Transformer decoders is reshaping\ncamera-based 3D object detection. These query-based decoders are surpassing the\ntraditional dense BEV (Bird's Eye View)-based methods. However, we argue that\ndense BEV frameworks remain important due to their outstanding abilities in\ndepth estimation and object localization, depicting 3D scenes accurately and\ncomprehensively. This paper aims to address the drawbacks of the existing dense\nBEV-based 3D object detectors by introducing our proposed enhanced components,\nincluding a CRF-modulated depth estimation module enforcing object-level\nconsistencies, a long-term temporal aggregation module with extended receptive\nfields, and a two-stage object decoder combining perspective techniques with\nCRF-modulated depth embedding. These enhancements lead to a \"modernized\" dense\nBEV framework dubbed BEVNeXt. On the nuScenes benchmark, BEVNeXt outperforms\nboth BEV-based and query-based frameworks under various settings, achieving a\nstate-of-the-art result of 64.2 NDS on the nuScenes test set. Code will be\navailable at \\url{https://github.com/woxihuanjiangguo/BEVNeXt}.\n", "date": "2023-12-04", "categories": []} {"id": "2402.19463", "title": "SeMoLi: What Moves Together Belongs Together", "abstract": " We tackle semi-supervised object detection based on motion cues. Recent\nresults suggest that heuristic-based clustering methods in conjunction with\nobject trackers can be used to pseudo-label instances of moving objects and use\nthese as supervisory signals to train 3D object detectors in Lidar data without\nmanual supervision. We re-think this approach and suggest that both, object\ndetection, as well as motion-inspired pseudo-labeling, can be tackled in a\ndata-driven manner. We leverage recent advances in scene flow estimation to\nobtain point trajectories from which we extract long-term, class-agnostic\nmotion patterns. Revisiting correlation clustering in the context of message\npassing networks, we learn to group those motion patterns to cluster points to\nobject instances. By estimating the full extent of the objects, we obtain\nper-scan 3D bounding boxes that we use to supervise a Lidar object detection\nnetwork. Our method not only outperforms prior heuristic-based approaches (57.5\nAP, +14 improvement over prior work), more importantly, we show we can\npseudo-label and train object detectors across datasets.\n", "date": "2024-02-29", "categories": []} {"id": "2403.15317", "title": "Point-DETR3D: Leveraging Imagery Data with Spatial Point Prior for\n Weakly Semi-supervised 3D Object Detection", "abstract": " Training high-accuracy 3D detectors necessitates massive labeled 3D\nannotations with 7 degree-of-freedom, which is laborious and time-consuming.\nTherefore, the form of point annotations is proposed to offer significant\nprospects for practical applications in 3D detection, which is not only more\naccessible and less expensive but also provides strong spatial information for\nobject localization. In this paper, we empirically discover that it is\nnon-trivial to merely adapt Point-DETR to its 3D form, encountering two main\nbottlenecks: 1) it fails to encode strong 3D prior into the model, and 2) it\ngenerates low-quality pseudo labels in distant regions due to the extreme\nsparsity of LiDAR points. To overcome these challenges, we introduce\nPoint-DETR3D, a teacher-student framework for weakly semi-supervised 3D\ndetection, designed to fully capitalize on point-wise supervision within a\nconstrained instance-wise annotation budget.Different from Point-DETR which\nencodes 3D positional information solely through a point encoder, we propose an\nexplicit positional query initialization strategy to enhance the positional\nprior. Considering the low quality of pseudo labels at distant regions produced\nby the teacher model, we enhance the detector's perception by incorporating\ndense imagery data through a novel Cross-Modal Deformable RoI Fusion\n(D-RoI).Moreover, an innovative point-guided self-supervised learning technique\nis proposed to allow for fully exploiting point priors, even in student\nmodels.Extensive experiments on representative nuScenes dataset demonstrate our\nPoint-DETR3D obtains significant improvements compared to previous works.\nNotably, with only 5% of labeled data, Point-DETR3D achieves over 90%\nperformance of its fully supervised counterpart.\n", "date": "2024-03-22", "categories": []} {"id": "2403.16440", "title": "RCBEVDet: Radar-camera Fusion in Bird's Eye View for 3D Object Detection", "abstract": " Three-dimensional object detection is one of the key tasks in autonomous\ndriving. To reduce costs in practice, low-cost multi-view cameras for 3D object\ndetection are proposed to replace the expansive LiDAR sensors. However, relying\nsolely on cameras is difficult to achieve highly accurate and robust 3D object\ndetection. An effective solution to this issue is combining multi-view cameras\nwith the economical millimeter-wave radar sensor to achieve more reliable\nmulti-modal 3D object detection. In this paper, we introduce RCBEVDet, a\nradar-camera fusion 3D object detection method in the bird's eye view (BEV).\nSpecifically, we first design RadarBEVNet for radar BEV feature extraction.\nRadarBEVNet consists of a dual-stream radar backbone and a Radar Cross-Section\n(RCS) aware BEV encoder. In the dual-stream radar backbone, a point-based\nencoder and a transformer-based encoder are proposed to extract radar features,\nwith an injection and extraction module to facilitate communication between the\ntwo encoders. The RCS-aware BEV encoder takes RCS as the object size prior to\nscattering the point feature in BEV. Besides, we present the Cross-Attention\nMulti-layer Fusion module to automatically align the multi-modal BEV feature\nfrom radar and camera with the deformable attention mechanism, and then fuse\nthe feature with channel and spatial fusion layers. Experimental results show\nthat RCBEVDet achieves new state-of-the-art radar-camera fusion results on\nnuScenes and view-of-delft (VoD) 3D object detection benchmarks. Furthermore,\nRCBEVDet achieves better 3D detection results than all real-time camera-only\nand radar-camera 3D object detectors with a faster inference speed at 21~28\nFPS. The source code will be released at https://github.com/VDIGPKU/RCBEVDet.\n", "date": "2024-03-25", "categories": []} {"id": "2403.15241", "title": "IS-Fusion: Instance-Scene Collaborative Fusion for Multimodal 3D Object\n Detection", "abstract": " Bird's eye view (BEV) representation has emerged as a dominant solution for\ndescribing 3D space in autonomous driving scenarios. However, objects in the\nBEV representation typically exhibit small sizes, and the associated point\ncloud context is inherently sparse, which leads to great challenges for\nreliable 3D perception. In this paper, we propose IS-Fusion, an innovative\nmultimodal fusion framework that jointly captures the Instance- and Scene-level\ncontextual information. IS-Fusion essentially differs from existing approaches\nthat only focus on the BEV scene-level fusion by explicitly incorporating\ninstance-level multimodal information, thus facilitating the instance-centric\ntasks like 3D object detection. It comprises a Hierarchical Scene Fusion (HSF)\nmodule and an Instance-Guided Fusion (IGF) module. HSF applies Point-to-Grid\nand Grid-to-Region transformers to capture the multimodal scene context at\ndifferent granularities. IGF mines instance candidates, explores their\nrelationships, and aggregates the local multimodal context for each instance.\nThese instances then serve as guidance to enhance the scene feature and yield\nan instance-aware BEV representation. On the challenging nuScenes benchmark,\nIS-Fusion outperforms all the published multimodal works to date. Code is\navailable at: https://github.com/yinjunbo/IS-Fusion.\n", "date": "2024-03-22", "categories": []} {"id": "2403.14133", "title": "3D Object Detection from Point Cloud via Voting Step Diffusion", "abstract": " 3D object detection is a fundamental task in scene understanding. Numerous\nresearch efforts have been dedicated to better incorporate Hough voting into\nthe 3D object detection pipeline. However, due to the noisy, cluttered, and\npartial nature of real 3D scans, existing voting-based methods often receive\nvotes from the partial surfaces of individual objects together with severe\nnoises, leading to sub-optimal detection performance. In this work, we focus on\nthe distributional properties of point clouds and formulate the voting process\nas generating new points in the high-density region of the distribution of\nobject centers. To achieve this, we propose a new method to move random 3D\npoints toward the high-density region of the distribution by estimating the\nscore function of the distribution with a noise conditioned score network.\nSpecifically, we first generate a set of object center proposals to coarsely\nidentify the high-density region of the object center distribution. To estimate\nthe score function, we perturb the generated object center proposals by adding\nnormalized Gaussian noise, and then jointly estimate the score function of all\nperturbed distributions. Finally, we generate new votes by moving random 3D\npoints to the high-density region of the object center distribution according\nto the estimated score function. Extensive experiments on two large scale\nindoor 3D scene datasets, SUN RGB-D and ScanNet V2, demonstrate the superiority\nof our proposed method. The code will be released at\nhttps://github.com/HHrEtvP/DiffVote.\n", "date": "2024-03-21", "categories": []} {"id": "2403.11573", "title": "Just Add $100 More: Augmenting NeRF-based Pseudo-LiDAR Point Cloud for\n Resolving Class-imbalance Problem", "abstract": " Typical LiDAR-based 3D object detection models are trained in a supervised\nmanner with real-world data collection, which is often imbalanced over classes\n(or long-tailed). To deal with it, augmenting minority-class examples by\nsampling ground truth (GT) LiDAR points from a database and pasting them into a\nscene of interest is often used, but challenges still remain: inflexibility in\nlocating GT samples and limited sample diversity. In this work, we propose to\nleverage pseudo-LiDAR point clouds generated (at a low cost) from videos\ncapturing a surround view of miniatures or real-world objects of minor classes.\nOur method, called Pseudo Ground Truth Augmentation (PGT-Aug), consists of\nthree main steps: (i) volumetric 3D instance reconstruction using a 2D-to-3D\nview synthesis model, (ii) object-level domain alignment with LiDAR intensity\nestimation and (iii) a hybrid context-aware placement method from ground and\nmap information. We demonstrate the superiority and generality of our method\nthrough performance improvements in extensive experiments conducted on three\npopular benchmarks, i.e., nuScenes, KITTI, and Lyft, especially for the\ndatasets with large domain gaps captured by different LiDAR configurations. Our\ncode and data will be publicly available upon publication.\n", "date": "2024-03-18", "categories": []} {"id": "2403.12310", "title": "Prototipo de un Contador Bidireccional Autom\\'atico de Personas basado\n en sensores de visi\\'on 3D", "abstract": " 3D sensors, also known as RGB-D sensors, utilize depth images where each\npixel measures the distance from the camera to objects, using principles like\nstructured light or time-of-flight. Advances in artificial vision have led to\naffordable 3D cameras capable of real-time object detection without object\nmovement, surpassing 2D cameras in information depth. These cameras can\nidentify objects of varying colors and reflectivities and are less affected by\nlighting changes. The described prototype uses RGB-D sensors for bidirectional\npeople counting in venues, aiding security and surveillance in spaces like\nstadiums or airports. It determines real-time occupancy and checks against\nmaximum capacity, crucial during emergencies. The system includes a RealSense\nD415 depth camera and a mini-computer running object detection algorithms to\ncount people and a 2D camera for identity verification. The system supports\nstatistical analysis and uses C++, Python, and PHP with OpenCV for image\nprocessing, demonstrating a comprehensive approach to monitoring venue\noccupancy.\n", "date": "2024-03-18", "categories": []} {"id": "2403.12573", "title": "Lifting Multi-View Detection and Tracking to the Bird's Eye View", "abstract": " Taking advantage of multi-view aggregation presents a promising solution to\ntackle challenges such as occlusion and missed detection in multi-object\ntracking and detection. Recent advancements in multi-view detection and 3D\nobject recognition have significantly improved performance by strategically\nprojecting all views onto the ground plane and conducting detection analysis\nfrom a Bird's Eye View. In this paper, we compare modern lifting methods, both\nparameter-free and parameterized, to multi-view aggregation. Additionally, we\npresent an architecture that aggregates the features of multiple times steps to\nlearn robust detection and combines appearance- and motion-based cues for\ntracking. Most current tracking approaches either focus on pedestrians or\nvehicles. In our work, we combine both branches and add new challenges to\nmulti-view detection with cross-scene setups. Our method generalizes to three\npublic datasets across two domains: (1) pedestrian: Wildtrack and MultiviewX,\nand (2) roadside perception: Synthehicle, achieving state-of-the-art\nperformance in detection and tracking. https://github.com/tteepe/TrackTacular\n", "date": "2024-03-19", "categories": []} {"id": "2208.08780", "title": "GraVoS: Voxel Selection for 3D Point-Cloud Detection", "abstract": " 3D object detection within large 3D scenes is challenging not only due to the\nsparsity and irregularity of 3D point clouds, but also due to both the extreme\nforeground-background scene imbalance and class imbalance. A common approach is\nto add ground-truth objects from other scenes. Differently, we propose to\nmodify the scenes by removing elements (voxels), rather than adding ones. Our\napproach selects the \"meaningful\" voxels, in a manner that addresses both types\nof dataset imbalance. The approach is general and can be applied to any\nvoxel-based detector, yet the meaningfulness of a voxel is network-dependent.\nOur voxel selection is shown to improve the performance of several prominent 3D\ndetection methods.\n", "date": "2022-08-18", "categories": []} {"id": "2403.02769", "title": "HUNTER: Unsupervised Human-centric 3D Detection via Transferring\n Knowledge from Synthetic Instances to Real Scenes", "abstract": " Human-centric 3D scene understanding has recently drawn increasing attention,\ndriven by its critical impact on robotics. However, human-centric real-life\nscenarios are extremely diverse and complicated, and humans have intricate\nmotions and interactions. With limited labeled data, supervised methods are\ndifficult to generalize to general scenarios, hindering real-life applications.\nMimicking human intelligence, we propose an unsupervised 3D detection method\nfor human-centric scenarios by transferring the knowledge from synthetic human\ninstances to real scenes. To bridge the gap between the distinct data\nrepresentations and feature distributions of synthetic models and real point\nclouds, we introduce novel modules for effective instance-to-scene\nrepresentation transfer and synthetic-to-real feature alignment. Remarkably,\nour method exhibits superior performance compared to current state-of-the-art\ntechniques, achieving 87.8% improvement in mAP and closely approaching the\nperformance of fully supervised methods (62.15 mAP vs. 69.02 mAP) on HuCenLife\nDataset.\n", "date": "2024-03-05", "categories": []} {"id": "2403.10036", "title": "SparseFusion: Efficient Sparse Multi-Modal Fusion Framework for\n Long-Range 3D Perception", "abstract": " Multi-modal 3D object detection has exhibited significant progress in recent\nyears. However, most existing methods can hardly scale to long-range scenarios\ndue to their reliance on dense 3D features, which substantially escalate\ncomputational demands and memory usage. In this paper, we introduce\nSparseFusion, a novel multi-modal fusion framework fully built upon sparse 3D\nfeatures to facilitate efficient long-range perception. The core of our method\nis the Sparse View Transformer module, which selectively lifts regions of\ninterest in 2D image space into the unified 3D space. The proposed module\nintroduces sparsity from both semantic and geometric aspects which only fill\ngrids that foreground objects potentially reside in. Comprehensive experiments\nhave verified the efficiency and effectiveness of our framework in long-range\n3D perception. Remarkably, on the long-range Argoverse2 dataset, SparseFusion\nreduces memory footprint and accelerates the inference by about two times\ncompared to dense detectors. It also achieves state-of-the-art performance with\nmAP of 41.2% and CDS of 32.1%. The versatility of SparseFusion is also\nvalidated in the temporal object detection task and 3D lane detection task.\nCodes will be released upon acceptance.\n", "date": "2024-03-15", "categories": []} {"id": "2403.09230", "title": "Improving Distant 3D Object Detection Using 2D Box Supervision", "abstract": " Improving the detection of distant 3d objects is an important yet challenging\ntask. For camera-based 3D perception, the annotation of 3d bounding relies\nheavily on LiDAR for accurate depth information. As such, the distance of\nannotation is often limited due to the sparsity of LiDAR points on distant\nobjects, which hampers the capability of existing detectors for long-range\nscenarios. We address this challenge by considering only 2D box supervision for\ndistant objects since they are easy to annotate. We propose LR3D, a framework\nthat learns to recover the missing depth of distant objects. LR3D adopts an\nimplicit projection head to learn the generation of mapping between 2D boxes\nand depth using the 3D supervision on close objects. This mapping allows the\ndepth estimation of distant objects conditioned on their 2D boxes, making\nlong-range 3D detection with 2D supervision feasible. Experiments show that\nwithout distant 3D annotations, LR3D allows camera-based methods to detect\ndistant objects (over 200m) with comparable accuracy to full 3D supervision.\nOur framework is general, and could widely benefit 3D detection methods to a\nlarge extent.\n", "date": "2024-03-14", "categories": []} {"id": "2309.17336", "title": "Robust 3D Object Detection from LiDAR-Radar Point Clouds via Cross-Modal\n Feature Augmentation", "abstract": " This paper presents a novel framework for robust 3D object detection from\npoint clouds via cross-modal hallucination. Our proposed approach is agnostic\nto either hallucination direction between LiDAR and 4D radar. We introduce\nmultiple alignments on both spatial and feature levels to achieve simultaneous\nbackbone refinement and hallucination generation. Specifically, spatial\nalignment is proposed to deal with the geometry discrepancy for better instance\nmatching between LiDAR and radar. The feature alignment step further bridges\nthe intrinsic attribute gap between the sensing modalities and stabilizes the\ntraining. The trained object detection models can deal with difficult detection\ncases better, even though only single-modal data is used as the input during\nthe inference stage. Extensive experiments on the View-of-Delft (VoD) dataset\nshow that our proposed method outperforms the state-of-the-art (SOTA) methods\nfor both radar and LiDAR object detection while maintaining competitive\nefficiency in runtime. Code is available at\nhttps://github.com/DJNing/See_beyond_seeing.\n", "date": "2023-09-29", "categories": []} {"id": "2402.03634", "title": "Ray Denoising: Depth-aware Hard Negative Sampling for Multi-view 3D\n Object Detection", "abstract": " Multi-view 3D object detection systems often struggle with generating precise\npredictions due to the challenges in estimating depth from images, increasing\nredundant and incorrect detections. Our paper presents Ray Denoising, an\ninnovative method that enhances detection accuracy by strategically sampling\nalong camera rays to construct hard negative examples. These examples, visually\nchallenging to differentiate from true positives, compel the model to learn\ndepth-aware features, thereby improving its capacity to distinguish between\ntrue and false positives. Ray Denoising is designed as a plug-and-play module,\ncompatible with any DETR-style multi-view 3D detectors, and it only minimally\nincreases training computational costs without affecting inference speed. Our\ncomprehensive experiments, including detailed ablation studies, consistently\ndemonstrate that Ray Denoising outperforms strong baselines across multiple\ndatasets. It achieves a 1.9\\% improvement in mean Average Precision (mAP) over\nthe state-of-the-art StreamPETR method on the NuScenes dataset. It shows\nsignificant performance gains on the Argoverse 2 dataset, highlighting its\ngeneralization capability. The code will be available at\nhttps://github.com/LiewFeng/RayDN.\n", "date": "2024-02-06", "categories": []} {"id": "2403.07071", "title": "LISO: Lidar-only Self-Supervised 3D Object Detection", "abstract": " 3D object detection is one of the most important components in any\nSelf-Driving stack, but current state-of-the-art (SOTA) lidar object detectors\nrequire costly & slow manual annotation of 3D bounding boxes to perform well.\nRecently, several methods emerged to generate pseudo ground truth without human\nsupervision, however, all of these methods have various drawbacks: Some methods\nrequire sensor rigs with full camera coverage and accurate calibration, partly\nsupplemented by an auxiliary optical flow engine. Others require expensive\nhigh-precision localization to find objects that disappeared over multiple\ndrives. We introduce a novel self-supervised method to train SOTA lidar object\ndetection networks which works on unlabeled sequences of lidar point clouds\nonly, which we call trajectory-regularized self-training. It utilizes a SOTA\nself-supervised lidar scene flow network under the hood to generate, track, and\niteratively refine pseudo ground truth. We demonstrate the effectiveness of our\napproach for multiple SOTA object detection networks across multiple real-world\ndatasets. Code will be released.\n", "date": "2024-03-11", "categories": []} {"id": "2403.07372", "title": "Eliminating Cross-modal Conflicts in BEV Space for LiDAR-Camera 3D\n Object Detection", "abstract": " Recent 3D object detectors typically utilize multi-sensor data and unify\nmulti-modal features in the shared bird's-eye view (BEV) representation space.\nHowever, our empirical findings indicate that previous methods have limitations\nin generating fusion BEV features free from cross-modal conflicts. These\nconflicts encompass extrinsic conflicts caused by BEV feature construction and\ninherent conflicts stemming from heterogeneous sensor signals. Therefore, we\npropose a novel Eliminating Conflicts Fusion (ECFusion) method to explicitly\neliminate the extrinsic/inherent conflicts in BEV space and produce improved\nmulti-modal BEV features. Specifically, we devise a Semantic-guided Flow-based\nAlignment (SFA) module to resolve extrinsic conflicts via unifying spatial\ndistribution in BEV space before fusion. Moreover, we design a Dissolved Query\nRecovering (DQR) mechanism to remedy inherent conflicts by preserving\nobjectness clues that are lost in the fusion BEV feature. In general, our\nmethod maximizes the effective information utilization of each modality and\nleverages inter-modal complementarity. Our method achieves state-of-the-art\nperformance in the highly competitive nuScenes 3D object detection dataset. The\ncode is released at https://github.com/fjhzhixi/ECFusion.\n", "date": "2024-03-12", "categories": []} {"id": "2401.11913", "title": "Large receptive field strategy and important feature extraction strategy\n in 3D object detection", "abstract": " The enhancement of 3D object detection is pivotal for precise environmental\nperception and improved task execution capabilities in autonomous driving.\nLiDAR point clouds, offering accurate depth information, serve as a crucial\ninformation for this purpose. Our study focuses on key challenges in 3D target\ndetection. To tackle the challenge of expanding the receptive field of a 3D\nconvolutional kernel, we introduce the Dynamic Feature Fusion Module (DFFM).\nThis module achieves adaptive expansion of the 3D convolutional kernel's\nreceptive field, balancing the expansion with acceptable computational loads.\nThis innovation reduces operations, expands the receptive field, and allows the\nmodel to dynamically adjust to different object requirements. Simultaneously,\nwe identify redundant information in 3D features. Employing the Feature\nSelection Module (FSM) quantitatively evaluates and eliminates non-important\nfeatures, achieving the separation of output box fitting and feature\nextraction. This innovation enables the detector to focus on critical features,\nresulting in model compression, reduced computational burden, and minimized\ncandidate frame interference. Extensive experiments confirm that both DFFM and\nFSM not only enhance current benchmarks, particularly in small target\ndetection, but also accelerate network performance. Importantly, these modules\nexhibit effective complementarity.\n", "date": "2024-01-22", "categories": []} {"id": "2403.06093", "title": "Enhancing 3D Object Detection with 2D Detection-Guided Query Anchors", "abstract": " Multi-camera-based 3D object detection has made notable progress in the past\nseveral years. However, we observe that there are cases (e.g. faraway regions)\nin which popular 2D object detectors are more reliable than state-of-the-art 3D\ndetectors. In this paper, to improve the performance of query-based 3D object\ndetectors, we present a novel query generating approach termed QAF2D, which\ninfers 3D query anchors from 2D detection results. A 2D bounding box of an\nobject in an image is lifted to a set of 3D anchors by associating each sampled\npoint within the box with depth, yaw angle, and size candidates. Then, the\nvalidity of each 3D anchor is verified by comparing its projection in the image\nwith its corresponding 2D box, and only valid anchors are kept and used to\nconstruct queries. The class information of the 2D bounding box associated with\neach query is also utilized to match the predicted boxes with ground truth for\nthe set-based loss. The image feature extraction backbone is shared between the\n3D detector and 2D detector by adding a small number of prompt parameters. We\nintegrate QAF2D into three popular query-based 3D object detectors and carry\nout comprehensive evaluations on the nuScenes dataset. The largest improvement\nthat QAF2D can bring about on the nuScenes validation subset is $2.3\\%$ NDS and\n$2.7\\%$ mAP. Code is available at https://github.com/nullmax-vision/QAF2D.\n", "date": "2024-03-10", "categories": []} {"id": "2403.06166", "title": "Cross-Cluster Shifting for Efficient and Effective 3D Object Detection\n in Autonomous Driving", "abstract": " We present a new 3D point-based detector model, named Shift-SSD, for precise\n3D object detection in autonomous driving. Traditional point-based 3D object\ndetectors often employ architectures that rely on a progressive downsampling of\npoints. While this method effectively reduces computational demands and\nincreases receptive fields, it will compromise the preservation of crucial\nnon-local information for accurate 3D object detection, especially in the\ncomplex driving scenarios. To address this, we introduce an intriguing\nCross-Cluster Shifting operation to unleash the representation capacity of the\npoint-based detector by efficiently modeling longer-range inter-dependency\nwhile including only a negligible overhead. Concretely, the Cross-Cluster\nShifting operation enhances the conventional design by shifting partial\nchannels from neighboring clusters, which enables richer interaction with\nnon-local regions and thus enlarges the receptive field of clusters. We conduct\nextensive experiments on the KITTI, Waymo, and nuScenes datasets, and the\nresults demonstrate the state-of-the-art performance of Shift-SSD in both\ndetection accuracy and runtime efficiency.\n", "date": "2024-03-10", "categories": []} {"id": "2403.06433", "title": "Fine-Grained Pillar Feature Encoding Via Spatio-Temporal Virtual Grid\n for 3D Object Detection", "abstract": " Developing high-performance, real-time architectures for LiDAR-based 3D\nobject detectors is essential for the successful commercialization of\nautonomous vehicles. Pillar-based methods stand out as a practical choice for\nonboard deployment due to their computational efficiency. However, despite\ntheir efficiency, these methods can sometimes underperform compared to\nalternative point encoding techniques such as Voxel-encoding or PointNet++. We\nargue that current pillar-based methods have not sufficiently captured the\nfine-grained distributions of LiDAR points within each pillar structure.\nConsequently, there exists considerable room for improvement in pillar feature\nencoding. In this paper, we introduce a novel pillar encoding architecture\nreferred to as Fine-Grained Pillar Feature Encoding (FG-PFE). FG-PFE utilizes\nSpatio-Temporal Virtual (STV) grids to capture the distribution of point clouds\nwithin each pillar across vertical, temporal, and horizontal dimensions.\nThrough STV grids, points within each pillar are individually encoded using\nVertical PFE (V-PFE), Temporal PFE (T-PFE), and Horizontal PFE (H-PFE). These\nencoded features are then aggregated through an Attentive Pillar Aggregation\nmethod. Our experiments conducted on the nuScenes dataset demonstrate that\nFG-PFE achieves significant performance improvements over baseline models such\nas PointPillar, CenterPoint-Pillar, and PillarNet, with only a minor increase\nin computational overhead.\n", "date": "2024-03-11", "categories": []} {"id": "2403.03721", "title": "CMDA: Cross-Modal and Domain Adversarial Adaptation for LiDAR-Based 3D\n Object Detection", "abstract": " Recent LiDAR-based 3D Object Detection (3DOD) methods show promising results,\nbut they often do not generalize well to target domains outside the source (or\ntraining) data distribution. To reduce such domain gaps and thus to make 3DOD\nmodels more generalizable, we introduce a novel unsupervised domain adaptation\n(UDA) method, called CMDA, which (i) leverages visual semantic cues from an\nimage modality (i.e., camera images) as an effective semantic bridge to close\nthe domain gap in the cross-modal Bird's Eye View (BEV) representations.\nFurther, (ii) we also introduce a self-training-based learning strategy,\nwherein a model is adversarially trained to generate domain-invariant features,\nwhich disrupt the discrimination of whether a feature instance comes from a\nsource or an unseen target domain. Overall, our CMDA framework guides the 3DOD\nmodel to generate highly informative and domain-adaptive features for novel\ndata distributions. In our extensive experiments with large-scale benchmarks,\nsuch as nuScenes, Waymo, and KITTI, those mentioned above provide significant\nperformance gains for UDA tasks, achieving state-of-the-art performance.\n", "date": "2024-03-06", "categories": []} {"id": "2301.12515", "title": "LiDAR-CS Dataset: LiDAR Point Cloud Dataset with Cross-Sensors for 3D\n Object Detection", "abstract": " Over the past few years, there has been remarkable progress in research on 3D\npoint clouds and their use in autonomous driving scenarios has become\nwidespread. However, deep learning methods heavily rely on annotated data and\noften face domain generalization issues. Unlike 2D images whose domains usually\npertain to the texture information present in them, the features derived from a\n3D point cloud are affected by the distribution of the points. The lack of a 3D\ndomain adaptation benchmark leads to the common practice of training a model on\none benchmark (e.g. Waymo) and then assessing it on another dataset (e.g.\nKITTI). This setting results in two distinct domain gaps: scenarios and\nsensors, making it difficult to analyze and evaluate the method accurately. To\ntackle this problem, this paper presents LiDAR Dataset with Cross Sensors\n(LiDAR-CS Dataset), which contains large-scale annotated LiDAR point cloud\nunder six groups of different sensors but with the same corresponding\nscenarios, captured from hybrid realistic LiDAR simulator. To our knowledge,\nLiDAR-CS Dataset is the first dataset that addresses the sensor-related gaps in\nthe domain of 3D object detection in real traffic. Furthermore, we evaluate and\nanalyze the performance using various baseline detectors and demonstrated its\npotential applications. Project page: https://opendriving.github.io/lidar-cs.\n", "date": "2023-01-29", "categories": []} {"id": "2312.03085", "title": "ScAR: Scaling Adversarial Robustness for LiDAR Object Detection", "abstract": " The adversarial robustness of a model is its ability to resist adversarial\nattacks in the form of small perturbations to input data. Universal adversarial\nattack methods such as Fast Sign Gradient Method (FSGM) and Projected Gradient\nDescend (PGD) are popular for LiDAR object detection, but they are often\ndeficient compared to task-specific adversarial attacks. Additionally, these\nuniversal methods typically require unrestricted access to the model's\ninformation, which is difficult to obtain in real-world applications. To\naddress these limitations, we present a black-box Scaling Adversarial\nRobustness (ScAR) method for LiDAR object detection. By analyzing the\nstatistical characteristics of 3D object detection datasets such as KITTI,\nWaymo, and nuScenes, we have found that the model's prediction is sensitive to\nscaling of 3D instances. We propose three black-box scaling adversarial attack\nmethods based on the available information: model-aware attack,\ndistribution-aware attack, and blind attack. We also introduce a strategy for\ngenerating scaling adversarial examples to improve the model's robustness\nagainst these three scaling adversarial attacks. Comparison with other methods\non public datasets under different 3D object detection architectures\ndemonstrates the effectiveness of our proposed method. Our code is available at\nhttps://github.com/xiaohulugo/ScAR-IROS2023.\n", "date": "2023-12-05", "categories": []} {"id": "2304.02867", "title": "Voxel or Pillar: Exploring Efficient Point Cloud Representation for 3D\n Object Detection", "abstract": " Efficient representation of point clouds is fundamental for LiDAR-based 3D\nobject detection. While recent grid-based detectors often encode point clouds\ninto either voxels or pillars, the distinctions between these approaches remain\nunderexplored. In this paper, we quantify the differences between the current\nencoding paradigms and highlight the limited vertical learning within. To\ntackle these limitations, we introduce a hybrid Voxel-Pillar Fusion network\n(VPF), which synergistically combines the unique strengths of both voxels and\npillars. Specifically, we first develop a sparse voxel-pillar encoder that\nencodes point clouds into voxel and pillar features through 3D and 2D sparse\nconvolutions respectively, and then introduce the Sparse Fusion Layer (SFL),\nfacilitating bidirectional interaction between sparse voxel and pillar\nfeatures. Our efficient, fully sparse method can be seamlessly integrated into\nboth dense and sparse detectors. Leveraging this powerful yet straightforward\nframework, VPF delivers competitive performance, achieving real-time inference\nspeeds on the nuScenes and Waymo Open Dataset. The code will be available.\n", "date": "2023-04-06", "categories": []} {"id": "2304.14614", "title": "Fusion is Not Enough: Single Modal Attacks on Fusion Models for 3D\n Object Detection", "abstract": " Multi-sensor fusion (MSF) is widely used in autonomous vehicles (AVs) for\nperception, particularly for 3D object detection with camera and LiDAR sensors.\nThe purpose of fusion is to capitalize on the advantages of each modality while\nminimizing its weaknesses. Advanced deep neural network (DNN)-based fusion\ntechniques have demonstrated the exceptional and industry-leading performance.\nDue to the redundant information in multiple modalities, MSF is also recognized\nas a general defence strategy against adversarial attacks. In this paper, we\nattack fusion models from the camera modality that is considered to be of\nlesser importance in fusion but is more affordable for attackers. We argue that\nthe weakest link of fusion models depends on their most vulnerable modality,\nand propose an attack framework that targets advanced camera-LiDAR fusion-based\n3D object detection models through camera-only adversarial attacks. Our\napproach employs a two-stage optimization-based strategy that first thoroughly\nevaluates vulnerable image areas under adversarial attacks, and then applies\ndedicated attack strategies for different fusion models to generate deployable\npatches. The evaluations with six advanced camera-LiDAR fusion models and one\ncamera-only model indicate that our attacks successfully compromise all of\nthem. Our approach can either decrease the mean average precision (mAP) of\ndetection performance from 0.824 to 0.353, or degrade the detection score of a\ntarget object from 0.728 to 0.156, demonstrating the efficacy of our proposed\nattack framework. Code is available.\n", "date": "2023-04-28", "categories": []} {"id": "2403.01978", "title": "Leveraging Anchor-based LiDAR 3D Object Detection via Point Assisted\n Sample Selection", "abstract": " 3D object detection based on LiDAR point cloud and prior anchor boxes is a\ncritical technology for autonomous driving environment perception and\nunderstanding. Nevertheless, an overlooked practical issue in existing methods\nis the ambiguity in training sample allocation based on box Intersection over\nUnion (IoU_box). This problem impedes further enhancements in the performance\nof anchor-based LiDAR 3D object detectors. To tackle this challenge, this paper\nintroduces a new training sample selection method that utilizes point cloud\ndistribution for anchor sample quality measurement, named Point Assisted Sample\nSelection (PASS). This method has undergone rigorous evaluation on two widely\nutilized datasets. Experimental results demonstrate that the application of\nPASS elevates the average precision of anchor-based LiDAR 3D object detectors\nto a novel state-of-the-art, thereby proving the effectiveness of the proposed\napproach. The codes will be made available at\nhttps://github.com/XJTU-Haolin/Point_Assisted_Sample_Selection.\n", "date": "2024-03-04", "categories": []} {"id": "2403.02037", "title": "Scalable Vision-Based 3D Object Detection and Monocular Depth Estimation\n for Autonomous Driving", "abstract": " This dissertation is a multifaceted contribution to the advancement of\nvision-based 3D perception technologies. In the first segment, the thesis\nintroduces structural enhancements to both monocular and stereo 3D object\ndetection algorithms. By integrating ground-referenced geometric priors into\nmonocular detection models, this research achieves unparalleled accuracy in\nbenchmark evaluations for monocular 3D detection. Concurrently, the work\nrefines stereo 3D detection paradigms by incorporating insights and inferential\nstructures gleaned from monocular networks, thereby augmenting the operational\nefficiency of stereo detection systems. The second segment is devoted to\ndata-driven strategies and their real-world applications in 3D vision\ndetection. A novel training regimen is introduced that amalgamates datasets\nannotated with either 2D or 3D labels. This approach not only augments the\ndetection models through the utilization of a substantially expanded dataset\nbut also facilitates economical model deployment in real-world scenarios where\nonly 2D annotations are readily available. Lastly, the dissertation presents an\ninnovative pipeline tailored for unsupervised depth estimation in autonomous\ndriving contexts. Extensive empirical analyses affirm the robustness and\nefficacy of this newly proposed pipeline. Collectively, these contributions lay\na robust foundation for the widespread adoption of vision-based 3D perception\ntechnologies in autonomous driving applications.\n", "date": "2024-03-04", "categories": []} {"id": "2403.00325", "title": "Small, Versatile and Mighty: A Range-View Perception Framework", "abstract": " Despite its compactness and information integrity, the range view\nrepresentation of LiDAR data rarely occurs as the first choice for 3D\nperception tasks. In this work, we further push the envelop of the range-view\nrepresentation with a novel multi-task framework, achieving unprecedented 3D\ndetection performances. Our proposed Small, Versatile, and Mighty (SVM) network\nutilizes a pure convolutional architecture to fully unleash the efficiency and\nmulti-tasking potentials of the range view representation. To boost detection\nperformances, we first propose a range-view specific Perspective Centric Label\nAssignment (PCLA) strategy, and a novel View Adaptive Regression (VAR) module\nto further refine hard-to-predict box properties. In addition, our framework\nseamlessly integrates semantic segmentation and panoptic segmentation tasks for\nthe LiDAR point cloud, without extra modules. Among range-view-based methods,\nour model achieves new state-of-the-art detection performances on the Waymo\nOpen Dataset. Especially, over 10 mAP improvement over convolutional\ncounterparts can be obtained on the vehicle class. Our presented results for\nother tasks further reveal the multi-task capabilities of the proposed small\nbut mighty framework.\n", "date": "2024-03-01", "categories": []} {"id": "2402.19144", "title": "Weakly Supervised Monocular 3D Detection with a Single-View Image", "abstract": " Monocular 3D detection (M3D) aims for precise 3D object localization from a\nsingle-view image which usually involves labor-intensive annotation of 3D\ndetection boxes. Weakly supervised M3D has recently been studied to obviate the\n3D annotation process by leveraging many existing 2D annotations, but it often\nrequires extra training data such as LiDAR point clouds or multi-view images\nwhich greatly degrades its applicability and usability in various applications.\nWe propose SKD-WM3D, a weakly supervised monocular 3D detection framework that\nexploits depth information to achieve M3D with a single-view image exclusively\nwithout any 3D annotations or other training data. One key design in SKD-WM3D\nis a self-knowledge distillation framework, which transforms image features\ninto 3D-like representations by fusing depth information and effectively\nmitigates the inherent depth ambiguity in monocular scenarios with little\ncomputational overhead in inference. In addition, we design an\nuncertainty-aware distillation loss and a gradient-targeted transfer modulation\nstrategy which facilitate knowledge acquisition and knowledge transfer,\nrespectively. Extensive experiments show that SKD-WM3D surpasses the\nstate-of-the-art clearly and is even on par with many fully supervised methods.\n", "date": "2024-02-29", "categories": []} {"id": "2311.10845", "title": "Domain Generalization of 3D Object Detection by Density-Resampling", "abstract": " Point-cloud-based 3D object detection suffers from performance degradation\nwhen encountering data with novel domain gaps. To tackle it, the single-domain\ngeneralization (SDG) aims to generalize the detection model trained in a\nlimited single source domain to perform robustly on unexplored domains. In this\npaper, we propose an SDG method to improve the generalizability of 3D object\ndetection to unseen target domains. Unlike prior SDG works for 3D object\ndetection solely focusing on data augmentation, our work introduces a novel\ndata augmentation method and contributes a new multi-task learning strategy in\nthe methodology. Specifically, from the perspective of data augmentation, we\ndesign a universal physical-aware density-based data augmentation (PDDA) method\nto mitigate the performance loss stemming from diverse point densities. From\nthe learning methodology viewpoint, we develop a multi-task learning for 3D\nobject detection: during source training, besides the main standard detection\ntask, we leverage an auxiliary self-supervised 3D scene restoration task to\nenhance the comprehension of the encoder on background and foreground details\nfor better recognition and detection of objects. Furthermore, based on the\nauxiliary self-supervised task, we propose the first test-time adaptation\nmethod for domain generalization of 3D object detection, which efficiently\nadjusts the encoder's parameters to adapt to unseen target domains during\ntesting time, to further bridge domain gaps. Extensive cross-dataset\nexperiments covering \"Car\", \"Pedestrian\", and \"Cyclist\" detections, demonstrate\nour method outperforms state-of-the-art SDG methods and even overpass\nunsupervised domain adaptation methods under some circumstances.\n", "date": "2023-11-17", "categories": []} {"id": "2402.18493", "title": "Sunshine to Rainstorm: Cross-Weather Knowledge Distillation for Robust\n 3D Object Detection", "abstract": " LiDAR-based 3D object detection models have traditionally struggled under\nrainy conditions due to the degraded and noisy scanning signals. Previous\nresearch has attempted to address this by simulating the noise from rain to\nimprove the robustness of detection models. However, significant disparities\nexist between simulated and actual rain-impacted data points. In this work, we\npropose a novel rain simulation method, termed DRET, that unifies Dynamics and\nRainy Environment Theory to provide a cost-effective means of expanding the\navailable realistic rain data for 3D detection training. Furthermore, we\npresent a Sunny-to-Rainy Knowledge Distillation (SRKD) approach to enhance 3D\ndetection under rainy conditions. Extensive experiments on the WaymoOpenDataset\nlarge-scale dataset show that, when combined with the state-of-the-art DSVT\nmodel and other classical 3D detectors, our proposed framework demonstrates\nsignificant detection accuracy improvements, without losing efficiency.\nRemarkably, our framework also improves detection capabilities under sunny\nconditions, therefore offering a robust solution for 3D detection regardless of\nwhether the weather is rainy or sunny\n", "date": "2024-02-28", "categories": []} {"id": "2402.17562", "title": "An Empirical Study of the Generalization Ability of Lidar 3D Object\n Detectors to Unseen Domains", "abstract": " 3D Object Detectors (3D-OD) are crucial for understanding the environment in\nmany robotic tasks, especially autonomous driving. Including 3D information via\nLidar sensors improves accuracy greatly. However, such detectors perform poorly\non domains they were not trained on, i.e. different locations, sensors,\nweather, etc., limiting their reliability in safety-critical applications.\nThere exist methods to adapt 3D-ODs to these domains; however, these methods\ntreat 3D-ODs as a black box, neglecting underlying architectural decisions and\nsource-domain training strategies. Instead, we dive deep into the details of\n3D-ODs, focusing our efforts on fundamental factors that influence robustness\nprior to domain adaptation.\n We systematically investigate four design choices (and the interplay between\nthem) often overlooked in 3D-OD robustness and domain adaptation: architecture,\nvoxel encoding, data augmentations, and anchor strategies. We assess their\nimpact on the robustness of nine state-of-the-art 3D-ODs across six benchmarks\nencompassing three types of domain gaps - sensor type, weather, and location.\n Our main findings are: (1) transformer backbones with local point features\nare more robust than 3D CNNs, (2) test-time anchor size adjustment is crucial\nfor adaptation across geographical locations, significantly boosting scores\nwithout retraining, (3) source-domain augmentations allow the model to\ngeneralize to low-resolution sensors, and (4) surprisingly, robustness to bad\nweather is improved when training directly on more clean weather data than on\ntraining with bad weather data. We outline our main conclusions and findings to\nprovide practical guidance on developing more robust 3D-ODs.\n", "date": "2024-02-27", "categories": []} {"id": "2310.14592", "title": "Pre-Training LiDAR-Based 3D Object Detectors Through Colorization", "abstract": " Accurate 3D object detection and understanding for self-driving cars heavily\nrelies on LiDAR point clouds, necessitating large amounts of labeled data to\ntrain. In this work, we introduce an innovative pre-training approach, Grounded\nPoint Colorization (GPC), to bridge the gap between data and labels by teaching\nthe model to colorize LiDAR point clouds, equipping it with valuable semantic\ncues. To tackle challenges arising from color variations and selection bias, we\nincorporate color as \"context\" by providing ground-truth colors as hints during\ncolorization. Experimental results on the KITTI and Waymo datasets demonstrate\nGPC's remarkable effectiveness. Even with limited labeled data, GPC\nsignificantly improves fine-tuning performance; notably, on just 20% of the\nKITTI dataset, GPC outperforms training from scratch with the entire dataset.\nIn sum, we introduce a fresh perspective on pre-training for 3D object\ndetection, aligning the objective with the model's intended role and ultimately\nadvancing the accuracy and efficiency of 3D object detection for autonomous\nvehicles.\n", "date": "2023-10-23", "categories": []} {"id": "2307.01864", "title": "MaskBEV: Joint Object Detection and Footprint Completion for Bird's-eye\n View 3D Point Clouds", "abstract": " Recent works in object detection in LiDAR point clouds mostly focus on\npredicting bounding boxes around objects. This prediction is commonly achieved\nusing anchor-based or anchor-free detectors that predict bounding boxes,\nrequiring significant explicit prior knowledge about the objects to work\nproperly. To remedy these limitations, we propose MaskBEV, a bird's-eye view\n(BEV) mask-based object detector neural architecture. MaskBEV predicts a set of\nBEV instance masks that represent the footprints of detected objects. Moreover,\nour approach allows object detection and footprint completion in a single pass.\nMaskBEV also reformulates the detection problem purely in terms of\nclassification, doing away with regression usually done to predict bounding\nboxes. We evaluate the performance of MaskBEV on both SemanticKITTI and KITTI\ndatasets while analyzing the architecture advantages and limitations.\n", "date": "2023-07-04", "categories": []} {"id": "2402.15272", "title": "EMIFF: Enhanced Multi-scale Image Feature Fusion for\n Vehicle-Infrastructure Cooperative 3D Object Detection", "abstract": " In autonomous driving, cooperative perception makes use of multi-view cameras\nfrom both vehicles and infrastructure, providing a global vantage point with\nrich semantic context of road conditions beyond a single vehicle viewpoint.\nCurrently, two major challenges persist in vehicle-infrastructure cooperative\n3D (VIC3D) object detection: $1)$ inherent pose errors when fusing multi-view\nimages, caused by time asynchrony across cameras; $2)$ information loss in\ntransmission process resulted from limited communication bandwidth. To address\nthese issues, we propose a novel camera-based 3D detection framework for VIC3D\ntask, Enhanced Multi-scale Image Feature Fusion (EMIFF). To fully exploit\nholistic perspectives from both vehicles and infrastructure, we propose\nMulti-scale Cross Attention (MCA) and Camera-aware Channel Masking (CCM)\nmodules to enhance infrastructure and vehicle features at scale, spatial, and\nchannel levels to correct the pose error introduced by camera asynchrony. We\nalso introduce a Feature Compression (FC) module with channel and spatial\ncompression blocks for transmission efficiency. Experiments show that EMIFF\nachieves SOTA on DAIR-V2X-C datasets, significantly outperforming previous\nearly-fusion and late-fusion methods with comparable transmission costs.\n", "date": "2024-02-23", "categories": []} {"id": "2402.14464", "title": "NeRF-Det++: Incorporating Semantic Cues and Perspective-aware Depth\n Supervision for Indoor Multi-View 3D Detection", "abstract": " NeRF-Det has achieved impressive performance in indoor multi-view 3D\ndetection by innovatively utilizing NeRF to enhance representation learning.\nDespite its notable performance, we uncover three decisive shortcomings in its\ncurrent design, including semantic ambiguity, inappropriate sampling, and\ninsufficient utilization of depth supervision. To combat the aforementioned\nproblems, we present three corresponding solutions: 1) Semantic Enhancement. We\nproject the freely available 3D segmentation annotations onto the 2D plane and\nleverage the corresponding 2D semantic maps as the supervision signal,\nsignificantly enhancing the semantic awareness of multi-view detectors. 2)\nPerspective-aware Sampling. Instead of employing the uniform sampling strategy,\nwe put forward the perspective-aware sampling policy that samples densely near\nthe camera while sparsely in the distance, more effectively collecting the\nvaluable geometric clues. 3)Ordinal Residual Depth Supervision. As opposed to\ndirectly regressing the depth values that are difficult to optimize, we divide\nthe depth range of each scene into a fixed number of ordinal bins and\nreformulate the depth prediction as the combination of the classification of\ndepth bins as well as the regression of the residual depth values, thereby\nbenefiting the depth learning process. The resulting algorithm, NeRF-Det++, has\nexhibited appealing performance in the ScanNetV2 and ARKITScenes datasets.\nNotably, in ScanNetV2, NeRF-Det++ outperforms the competitive NeRF-Det by +1.9%\nin mAP@0.25 and +3.5% in mAP@0.50$. The code will be publicly at\nhttps://github.com/mrsempress/NeRF-Detplusplus.\n", "date": "2024-02-22", "categories": []} {"id": "2212.10049", "title": "OBMO: One Bounding Box Multiple Objects for Monocular 3D Object\n Detection", "abstract": " Compared to typical multi-sensor systems, monocular 3D object detection has\nattracted much attention due to its simple configuration. However, there is\nstill a significant gap between LiDAR-based and monocular-based methods. In\nthis paper, we find that the ill-posed nature of monocular imagery can lead to\ndepth ambiguity. Specifically, objects with different depths can appear with\nthe same bounding boxes and similar visual features in the 2D image.\nUnfortunately, the network cannot accurately distinguish different depths from\nsuch non-discriminative visual features, resulting in unstable depth training.\nTo facilitate depth learning, we propose a simple yet effective plug-and-play\nmodule, \\underline{O}ne \\underline{B}ounding Box \\underline{M}ultiple\n\\underline{O}bjects (OBMO). Concretely, we add a set of suitable pseudo labels\nby shifting the 3D bounding box along the viewing frustum. To constrain the\npseudo-3D labels to be reasonable, we carefully design two label scoring\nstrategies to represent their quality. In contrast to the original hard depth\nlabels, such soft pseudo labels with quality scores allow the network to learn\na reasonable depth range, boosting training stability and thus improving final\nperformance. Extensive experiments on KITTI and Waymo benchmarks show that our\nmethod significantly improves state-of-the-art monocular 3D detectors by a\nsignificant margin (The improvements under the moderate setting on KITTI\nvalidation set are $\\mathbf{1.82\\sim 10.91\\%}$ \\textbf{mAP in BEV} and\n$\\mathbf{1.18\\sim 9.36\\%}$ \\textbf{mAP in 3D}). Codes have been released at\n\\url{https://github.com/mrsempress/OBMO}.\n", "date": "2022-12-20", "categories": []} {"id": "2402.11735", "title": "LiRaFusion: Deep Adaptive LiDAR-Radar Fusion for 3D Object Detection", "abstract": " We propose LiRaFusion to tackle LiDAR-radar fusion for 3D object detection to\nfill the performance gap of existing LiDAR-radar detectors. To improve the\nfeature extraction capabilities from these two modalities, we design an early\nfusion module for joint voxel feature encoding, and a middle fusion module to\nadaptively fuse feature maps via a gated network. We perform extensive\nevaluation on nuScenes to demonstrate that LiRaFusion leverages the\ncomplementary information of LiDAR and radar effectively and achieves notable\nimprovement over existing methods.\n", "date": "2024-02-18", "categories": []} {"id": "2309.09599", "title": "MEDL-U: Uncertainty-aware 3D Automatic Annotation based on Evidential\n Deep Learning", "abstract": " Advancements in deep learning-based 3D object detection necessitate the\navailability of large-scale datasets. However, this requirement introduces the\nchallenge of manual annotation, which is often both burdensome and\ntime-consuming. To tackle this issue, the literature has seen the emergence of\nseveral weakly supervised frameworks for 3D object detection which can\nautomatically generate pseudo labels for unlabeled data. Nevertheless, these\ngenerated pseudo labels contain noise and are not as accurate as those labeled\nby humans. In this paper, we present the first approach that addresses the\ninherent ambiguities present in pseudo labels by introducing an Evidential Deep\nLearning (EDL) based uncertainty estimation framework. Specifically, we propose\nMEDL-U, an EDL framework based on MTrans, which not only generates pseudo\nlabels but also quantifies the associated uncertainties. However, applying EDL\nto 3D object detection presents three primary challenges: (1) relatively lower\npseudolabel quality in comparison to other autolabelers; (2) excessively high\nevidential uncertainty estimates; and (3) lack of clear interpretability and\neffective utilization of uncertainties for downstream tasks. We tackle these\nissues through the introduction of an uncertainty-aware IoU-based loss, an\nevidence-aware multi-task loss function, and the implementation of a\npost-processing stage for uncertainty refinement. Our experimental results\ndemonstrate that probabilistic detectors trained using the outputs of MEDL-U\nsurpass deterministic detectors trained using outputs from previous 3D\nannotators on the KITTI val set for all difficulty levels. Moreover, MEDL-U\nachieves state-of-the-art results on the KITTI official test set compared to\nexisting 3D automatic annotators.\n", "date": "2023-09-18", "categories": []} {"id": "2305.17927", "title": "VCVW-3D: A Virtual Construction Vehicles and Workers Dataset with 3D\n Annotations", "abstract": " Currently, object detection applications in construction are almost based on\npure 2D data (both image and annotation are 2D-based), resulting in the\ndeveloped artificial intelligence (AI) applications only applicable to some\nscenarios that only require 2D information. However, most advanced applications\nusually require AI agents to perceive 3D spatial information, which limits the\nfurther development of the current computer vision (CV) in construction. The\nlack of 3D annotated datasets for construction object detection worsens the\nsituation. Therefore, this study creates and releases a virtual dataset with 3D\nannotations named VCVW-3D, which covers 15 construction scenes and involves ten\ncategories of construction vehicles and workers. The VCVW-3D dataset is\ncharacterized by multi-scene, multi-category, multi-randomness,\nmulti-viewpoint, multi-annotation, and binocular vision. Several typical 2D and\nmonocular 3D object detection models are then trained and evaluated on the\nVCVW-3D dataset to provide a benchmark for subsequent research. The VCVW-3D is\nexpected to bring considerable economic benefits and practical significance by\nreducing the costs of data construction, prototype development, and exploration\nof space-awareness applications, thus promoting the development of CV in\nconstruction, especially those of 3D applications.\n", "date": "2023-05-29", "categories": []} {"id": "2310.03821", "title": "WLST: Weak Labels Guided Self-training for Weakly-supervised Domain\n Adaptation on 3D Object Detection", "abstract": " In the field of domain adaptation (DA) on 3D object detection, most of the\nwork is dedicated to unsupervised domain adaptation (UDA). Yet, without any\ntarget annotations, the performance gap between the UDA approaches and the\nfully-supervised approach is still noticeable, which is impractical for\nreal-world applications. On the other hand, weakly-supervised domain adaptation\n(WDA) is an underexplored yet practical task that only requires few labeling\neffort on the target domain. To improve the DA performance in a cost-effective\nway, we propose a general weak labels guided self-training framework, WLST,\ndesigned for WDA on 3D object detection. By incorporating autolabeler, which\ncan generate 3D pseudo labels from 2D bounding boxes, into the existing\nself-training pipeline, our method is able to generate more robust and\nconsistent pseudo labels that would benefit the training process on the target\ndomain. Extensive experiments demonstrate the effectiveness, robustness, and\ndetector-agnosticism of our WLST framework. Notably, it outperforms previous\nstate-of-the-art methods on all evaluation tasks.\n", "date": "2023-10-05", "categories": []} {"id": "2402.04883", "title": "Toward Accurate Camera-based 3D Object Detection via Cascade Depth\n Estimation and Calibration", "abstract": " Recent camera-based 3D object detection is limited by the precision of\ntransforming from image to 3D feature spaces, as well as the accuracy of object\nlocalization within the 3D space. This paper aims to address such a fundamental\nproblem of camera-based 3D object detection: How to effectively learn depth\ninformation for accurate feature lifting and object localization. Different\nfrom previous methods which directly predict depth distributions by using a\nsupervised estimation model, we propose a cascade framework consisting of two\ndepth-aware learning paradigms. First, a depth estimation (DE) scheme leverages\nrelative depth information to realize the effective feature lifting from 2D to\n3D spaces. Furthermore, a depth calibration (DC) scheme introduces depth\nreconstruction to further adjust the 3D object localization perturbation along\nthe depth axis. In practice, the DE is explicitly realized by using both the\nabsolute and relative depth optimization loss to promote the precision of depth\nprediction, while the capability of DC is implicitly embedded into the\ndetection Transformer through a depth denoising mechanism in the training\nphase. The entire model training is accomplished through an end-to-end manner.\nWe propose a baseline detector and evaluate the effectiveness of our proposal\nwith +2.2%/+2.7% NDS/mAP improvements on NuScenes benchmark, and gain a\ncomparable performance with 55.9%/45.7% NDS/mAP. Furthermore, we conduct\nextensive experiments to demonstrate its generality based on various detectors\nwith about +2% NDS improvements.\n", "date": "2024-02-07", "categories": []} {"id": "2211.06108", "title": "RaLiBEV: Radar and LiDAR BEV Fusion Learning for Anchor Box Free Object\n Detection Systems", "abstract": " In autonomous driving, LiDAR and radar are crucial for environmental\nperception. LiDAR offers precise 3D spatial sensing information but struggles\nin adverse weather like fog. Conversely, radar signals can penetrate rain or\nmist due to their specific wavelength but are prone to noise disturbances.\nRecent state-of-the-art works reveal that the fusion of radar and LiDAR can\nlead to robust detection in adverse weather. The existing works adopt\nconvolutional neural network architecture to extract features from each sensor\ndata, then align and aggregate the two branch features to predict object\ndetection results. However, these methods have low accuracy of predicted\nbounding boxes due to a simple design of label assignment and fusion\nstrategies. In this paper, we propose a bird's-eye view fusion learning-based\nanchor box-free object detection system, which fuses the feature derived from\nthe radar range-azimuth heatmap and the LiDAR point cloud to estimate possible\nobjects. Different label assignment strategies have been designed to facilitate\nthe consistency between the classification of foreground or background anchor\npoints and the corresponding bounding box regressions. Furthermore, the\nperformance of the proposed object detector is further enhanced by employing a\nnovel interactive transformer module. The superior performance of the methods\nproposed in this paper has been demonstrated using the recently published\nOxford Radar RobotCar dataset. Our system's average precision significantly\noutperforms the state-of-the-art method by 13.1% and 19.0% at Intersection of\nUnion (IoU) of 0.8 under 'Clear+Foggy' training conditions for 'Clear' and\n'Foggy' testing, respectively.\n", "date": "2022-11-11", "categories": []} {"id": "2306.05418", "title": "Weakly Supervised 3D Object Detection with Multi-Stage Generalization", "abstract": " With the rapid development of large models, the need for data has become\nincreasingly crucial. Especially in 3D object detection, costly manual\nannotations have hindered further advancements. To reduce the burden of\nannotation, we study the problem of achieving 3D object detection solely based\non 2D annotations. Thanks to advanced 3D reconstruction techniques, it is now\nfeasible to reconstruct the overall static 3D scene. However, extracting\nprecise object-level annotations from the entire scene and generalizing these\nlimited annotations to the entire scene remain challenges. In this paper, we\nintroduce a novel paradigm called BA$^2$-Det, encompassing pseudo label\ngeneration and multi-stage generalization. We devise the DoubleClustering\nalgorithm to obtain object clusters from reconstructed scene-level points, and\nfurther enhance the model's detection capabilities by developing three stages\nof generalization: progressing from complete to partial, static to dynamic, and\nclose to distant. Experiments conducted on the large-scale Waymo Open Dataset\nshow that the performance of BA$^2$-Det is on par with the fully-supervised\nmethods using 10% annotations. Additionally, using large raw videos for\npretraining,BA$^2$-Det can achieve a 20% relative improvement on the KITTI\ndataset. The method also has great potential for detecting open-set 3D objects\nin complex scenes. Project page: https://ba2det.site.\n", "date": "2023-06-08", "categories": []} {"id": "2306.02245", "title": "SAM3D: Zero-Shot 3D Object Detection via Segment Anything Model", "abstract": " With the development of large language models, many remarkable linguistic\nsystems like ChatGPT have thrived and achieved astonishing success on many\ntasks, showing the incredible power of foundation models. In the spirit of\nunleashing the capability of foundation models on vision tasks, the Segment\nAnything Model (SAM), a vision foundation model for image segmentation, has\nbeen proposed recently and presents strong zero-shot ability on many downstream\n2D tasks. However, whether SAM can be adapted to 3D vision tasks has yet to be\nexplored, especially 3D object detection. With this inspiration, we explore\nadapting the zero-shot ability of SAM to 3D object detection in this paper. We\npropose a SAM-powered BEV processing pipeline to detect objects and get\npromising results on the large-scale Waymo open dataset. As an early attempt,\nour method takes a step toward 3D object detection with vision foundation\nmodels and presents the opportunity to unleash their power on 3D vision tasks.\nThe code is released at https://github.com/DYZhang09/SAM3D.\n", "date": "2023-06-04", "categories": []} {"id": "2401.15319", "title": "You Only Look Bottom-Up for Monocular 3D Object Detection", "abstract": " Monocular 3D Object Detection is an essential task for autonomous driving.\nMeanwhile, accurate 3D object detection from pure images is very challenging\ndue to the loss of depth information. Most existing image-based methods infer\nobjects' location in 3D space based on their 2D sizes on the image plane, which\nusually ignores the intrinsic position clues from images, leading to\nunsatisfactory performances. Motivated by the fact that humans could leverage\nthe bottom-up positional clues to locate objects in 3D space from a single\nimage, in this paper, we explore the position modeling from the image feature\ncolumn and propose a new method named You Only Look Bottum-Up (YOLOBU).\nSpecifically, our YOLOBU leverages Column-based Cross Attention to determine\nhow much a pixel contributes to pixels above it. Next, the Row-based Reverse\nCumulative Sum (RRCS) is introduced to build the connections of pixels in the\nbottom-up direction. Our YOLOBU fully explores the position clues for monocular\n3D detection via building the relationship of pixels from the bottom-up way.\nExtensive experiments on the KITTI dataset demonstrate the effectiveness and\nsuperiority of our method.\n", "date": "2024-01-27", "categories": []} {"id": "2401.15865", "title": "LiDAR-PTQ: Post-Training Quantization for Point Cloud 3D Object\n Detection", "abstract": " Due to highly constrained computing power and memory, deploying 3D\nlidar-based detectors on edge devices equipped in autonomous vehicles and\nrobots poses a crucial challenge. Being a convenient and straightforward model\ncompression approach, Post-Training Quantization (PTQ) has been widely adopted\nin 2D vision tasks. However, applying it directly to 3D lidar-based tasks\ninevitably leads to performance degradation. As a remedy, we propose an\neffective PTQ method called LiDAR-PTQ, which is particularly curated for 3D\nlidar detection (both SPConv-based and SPConv-free). Our LiDAR-PTQ features\nthree main components, \\textbf{(1)} a sparsity-based calibration method to\ndetermine the initialization of quantization parameters, \\textbf{(2)} a\nTask-guided Global Positive Loss (TGPL) to reduce the disparity between the\nfinal predictions before and after quantization, \\textbf{(3)} an adaptive\nrounding-to-nearest operation to minimize the layerwise reconstruction error.\nExtensive experiments demonstrate that our LiDAR-PTQ can achieve\nstate-of-the-art quantization performance when applied to CenterPoint (both\nPillar-based and Voxel-based). To our knowledge, for the very first time in\nlidar-based 3D detection tasks, the PTQ INT8 model's accuracy is almost the\nsame as the FP32 model while enjoying $3\\times$ inference speedup. Moreover,\nour LiDAR-PTQ is cost-effective being $30\\times$ faster than the\nquantization-aware training method. Code will be released at\n\\url{https://github.com/StiphyJay/LiDAR-PTQ}.\n", "date": "2024-01-29", "categories": []} {"id": "2401.16305", "title": "MixSup: Mixed-grained Supervision for Label-efficient LiDAR-based 3D\n Object Detection", "abstract": " Label-efficient LiDAR-based 3D object detection is currently dominated by\nweakly/semi-supervised methods. Instead of exclusively following one of them,\nwe propose MixSup, a more practical paradigm simultaneously utilizing massive\ncheap coarse labels and a limited number of accurate labels for Mixed-grained\nSupervision. We start by observing that point clouds are usually textureless,\nmaking it hard to learn semantics. However, point clouds are geometrically rich\nand scale-invariant to the distances from sensors, making it relatively easy to\nlearn the geometry of objects, such as poses and shapes. Thus, MixSup leverages\nmassive coarse cluster-level labels to learn semantics and a few expensive\nbox-level labels to learn accurate poses and shapes. We redesign the label\nassignment in mainstream detectors, which allows them seamlessly integrated\ninto MixSup, enabling practicality and universality. We validate its\neffectiveness in nuScenes, Waymo Open Dataset, and KITTI, employing various\ndetectors. MixSup achieves up to 97.31% of fully supervised performance, using\ncheap cluster annotations and only 10% box annotations. Furthermore, we propose\nPointSAM based on the Segment Anything Model for automated coarse labeling,\nfurther reducing the annotation burden. The code is available at\nhttps://github.com/BraveGroup/PointSAM-for-MixSup.\n", "date": "2024-01-29", "categories": []} {"id": "2401.12447", "title": "NIV-SSD: Neighbor IoU-Voting Single-Stage Object Detector From Point\n Cloud", "abstract": " Previous single-stage detectors typically suffer the misalignment between\nlocalization accuracy and classification confidence. To solve the misalignment\nproblem, we introduce a novel rectification method named neighbor IoU-voting\n(NIV) strategy. Typically, classification and regression are treated as\nseparate branches, making it challenging to establish a connection between\nthem. Consequently, the classification confidence cannot accurately reflect the\nregression quality. NIV strategy can serve as a bridge between classification\nand regression branches by calculating two types of statistical data from the\nregression output to correct the classification confidence. Furthermore, to\nalleviate the imbalance of detection accuracy for complete objects with dense\npoints (easy objects) and incomplete objects with sparse points (difficult\nobjects), we propose a new data augmentation scheme named object resampling. It\nundersamples easy objects and oversamples difficult objects by randomly\ntransforming part of easy objects into difficult objects. Finally, combining\nthe NIV strategy and object resampling augmentation, we design an efficient\nsingle-stage detector termed NIV-SSD. Extensive experiments on several datasets\nindicate the effectiveness of the NIV strategy and the competitive performance\nof the NIV-SSD detector. The code will be available at\nhttps://github.com/Say2L/NIV-SSD.\n", "date": "2024-01-23", "categories": []} {"id": "2212.05758", "title": "BEV-MAE: Bird's Eye View Masked Autoencoders for Point Cloud\n Pre-training in Autonomous Driving Scenarios", "abstract": " Existing LiDAR-based 3D object detection methods for autonomous driving\nscenarios mainly adopt the training-from-scratch paradigm. Unfortunately, this\nparadigm heavily relies on large-scale labeled data, whose collection can be\nexpensive and time-consuming. Self-supervised pre-training is an effective and\ndesirable way to alleviate this dependence on extensive annotated data. In this\nwork, we present BEV-MAE, an efficient masked autoencoder pre-training\nframework for LiDAR-based 3D object detection in autonomous driving.\nSpecifically, we propose a bird's eye view (BEV) guided masking strategy to\nguide the 3D encoder learning feature representation in a BEV perspective and\navoid complex decoder design during pre-training. Furthermore, we introduce a\nlearnable point token to maintain a consistent receptive field size of the 3D\nencoder with fine-tuning for masked point cloud inputs. Based on the property\nof outdoor point clouds in autonomous driving scenarios, i.e., the point clouds\nof distant objects are more sparse, we propose point density prediction to\nenable the 3D encoder to learn location information, which is essential for\nobject detection. Experimental results show that BEV-MAE surpasses prior\nstate-of-the-art self-supervised methods and achieves a favorably pre-training\nefficiency. Furthermore, based on TransFusion-L, BEV-MAE achieves new\nstate-of-the-art LiDAR-based 3D object detection results, with 73.6 NDS and\n69.6 mAP on the nuScenes benchmark. The source code will be released at\nhttps://github.com/VDIGPKU/BEV-MAE\n", "date": "2022-12-12", "categories": []} {"id": "2303.07064", "title": "A Generalized Multi-Modal Fusion Detection Framework", "abstract": " LiDAR point clouds have become the most common data source in autonomous\ndriving. However, due to the sparsity of point clouds, accurate and reliable\ndetection cannot be achieved in specific scenarios. Because of their\ncomplementarity with point clouds, images are getting increasing attention.\nAlthough with some success, existing fusion methods either perform hard fusion\nor do not fuse in a direct manner. In this paper, we propose a generic 3D\ndetection framework called MMFusion, using multi-modal features. The framework\naims to achieve accurate fusion between LiDAR and images to improve 3D\ndetection in complex scenes. Our framework consists of two separate streams:\nthe LiDAR stream and the camera stream, which can be compatible with any\nsingle-modal feature extraction network. The Voxel Local Perception Module in\nthe LiDAR stream enhances local feature representation, and then the\nMulti-modal Feature Fusion Module selectively combines feature output from\ndifferent streams to achieve better fusion. Extensive experiments have shown\nthat our framework not only outperforms existing benchmarks but also improves\ntheir detection, especially for detecting cyclists and pedestrians on KITTI\nbenchmarks, with strong robustness and generalization capabilities. Hopefully,\nour work will stimulate more research into multi-modal fusion for autonomous\ndriving tasks.\n", "date": "2023-03-13", "categories": []} {"id": "2401.11718", "title": "MsSVT++: Mixed-scale Sparse Voxel Transformer with Center Voting for 3D\n Object Detection", "abstract": " Accurate 3D object detection in large-scale outdoor scenes, characterized by\nconsiderable variations in object scales, necessitates features rich in both\nlong-range and fine-grained information. While recent detectors have utilized\nwindow-based transformers to model long-range dependencies, they tend to\noverlook fine-grained details. To bridge this gap, we propose MsSVT++, an\ninnovative Mixed-scale Sparse Voxel Transformer that simultaneously captures\nboth types of information through a divide-and-conquer approach. This approach\ninvolves explicitly dividing attention heads into multiple groups, each\nresponsible for attending to information within a specific range. The outputs\nof these groups are subsequently merged to obtain final mixed-scale features.\nTo mitigate the computational complexity associated with applying a\nwindow-based transformer in 3D voxel space, we introduce a novel Chessboard\nSampling strategy and implement voxel sampling and gathering operations\nsparsely using a hash map. Moreover, an important challenge stems from the\nobservation that non-empty voxels are primarily located on the surface of\nobjects, which impedes the accurate estimation of bounding boxes. To overcome\nthis challenge, we introduce a Center Voting module that integrates newly voted\nvoxels enriched with mixed-scale contextual information towards the centers of\nthe objects, thereby improving precise object localization. Extensive\nexperiments demonstrate that our single-stage detector, built upon the\nfoundation of MsSVT++, consistently delivers exceptional performance across\ndiverse datasets.\n", "date": "2024-01-22", "categories": []} {"id": "2309.04105", "title": "Weakly Supervised Point Clouds Transformer for 3D Object Detection", "abstract": " The annotation of 3D datasets is required for semantic-segmentation and\nobject detection in scene understanding. In this paper we present a framework\nfor the weakly supervision of a point clouds transformer that is used for 3D\nobject detection. The aim is to decrease the required amount of supervision\nneeded for training, as a result of the high cost of annotating a 3D datasets.\nWe propose an Unsupervised Voting Proposal Module, which learns randomly preset\nanchor points and uses voting network to select prepared anchor points of high\nquality. Then it distills information into student and teacher network. In\nterms of student network, we apply ResNet network to efficiently extract local\ncharacteristics. However, it also can lose much global information. To provide\nthe input which incorporates the global and local information as the input of\nstudent networks, we adopt the self-attention mechanism of transformer to\nextract global features, and the ResNet layers to extract region proposals. The\nteacher network supervises the classification and regression of the student\nnetwork using the pre-trained model on ImageNet. On the challenging KITTI\ndatasets, the experimental results have achieved the highest level of average\nprecision compared with the most recent weakly supervised 3D object detectors.\n", "date": "2023-09-08", "categories": []} {"id": "2401.09049", "title": "Enhancing Lidar-based Object Detection in Adverse Weather using Offset\n Sequences in Time", "abstract": " Automated vehicles require an accurate perception of their surroundings for\nsafe and efficient driving. Lidar-based object detection is a widely used\nmethod for environment perception, but its performance is significantly\naffected by adverse weather conditions such as rain and fog. In this work, we\ninvestigate various strategies for enhancing the robustness of lidar-based\nobject detection by processing sequential data samples generated by lidar\nsensors. Our approaches leverage temporal information to improve a lidar object\ndetection model, without the need for additional filtering or pre-processing\nsteps. We compare $10$ different neural network architectures that process\npoint cloud sequences including a novel augmentation strategy introducing a\ntemporal offset between frames of a sequence during training and evaluate the\neffectiveness of all strategies on lidar point clouds under adverse weather\nconditions through experiments. Our research provides a comprehensive study of\neffective methods for mitigating the effects of adverse weather on the\nreliability of lidar-based object detection using sequential data that are\nevaluated using public datasets such as nuScenes, Dense, and the Canadian\nAdverse Driving Conditions Dataset. Our findings demonstrate that our novel\nmethod, involving temporal offset augmentation through randomized frame\nskipping in sequences, enhances object detection accuracy compared to both the\nbaseline model (Pillar-based Object Detection) and no augmentation.\n", "date": "2024-01-17", "categories": []} {"id": "2212.06244", "title": "PathFusion: Path-consistent Lidar-Camera Deep Feature Fusion", "abstract": " Fusing 3D LiDAR features with 2D camera features is a promising technique for\nenhancing the accuracy of 3D detection, thanks to their complementary physical\nproperties. While most of the existing methods focus on directly fusing camera\nfeatures with raw LiDAR point clouds or shallow-level 3D features, it is\nobserved that directly combining 2D and 3D features in deeper layers actually\nleads to a decrease in accuracy due to feature misalignment. The misalignment,\nwhich stems from the aggregation of features learned from large receptive\nfields, becomes increasingly more severe as we delve into deeper layers. In\nthis paper, we propose PathFusion as a solution to enable the alignment of\nsemantically coherent LiDAR-camera deep feature fusion. PathFusion introduces a\npath consistency loss at multiple stages within the network, encouraging the 2D\nbackbone and its fusion path to transform 2D features in a way that aligns\nsemantically with the transformation of the 3D backbone. This ensures semantic\nconsistency between 2D and 3D features, even in deeper layers, and amplifies\nthe usage of the network's learning capacity. We apply PathFusion to improve a\nprior-art fusion baseline, Focals Conv, and observe an improvement of over 1.6%\nin mAP on the nuScenes test split consistently with and without testing-time\ndata augmentations, and moreover, PathFusion also improves KITTI\n$\\text{AP}_{\\text{3D}}$ (R11) by about 0.6% on the moderate level.\n", "date": "2022-12-12", "categories": []} {"id": "2401.06994", "title": "UniVision: A Unified Framework for Vision-Centric 3D Perception", "abstract": " The past few years have witnessed the rapid development of vision-centric 3D\nperception in autonomous driving. Although the 3D perception models share many\nstructural and conceptual similarities, there still exist gaps in their feature\nrepresentations, data formats, and objectives, posing challenges for unified\nand efficient 3D perception framework design. In this paper, we present\nUniVision, a simple and efficient framework that unifies two major tasks in\nvision-centric 3D perception, \\ie, occupancy prediction and object detection.\nSpecifically, we propose an explicit-implicit view transform module for\ncomplementary 2D-3D feature transformation. We propose a local-global feature\nextraction and fusion module for efficient and adaptive voxel and BEV feature\nextraction, enhancement, and interaction. Further, we propose a joint\noccupancy-detection data augmentation strategy and a progressive loss weight\nadjustment strategy which enables the efficiency and stability of the\nmulti-task framework training. We conduct extensive experiments for different\nperception tasks on four public benchmarks, including nuScenes LiDAR\nsegmentation, nuScenes detection, OpenOccupancy, and Occ3D. UniVision achieves\nstate-of-the-art results with +1.5 mIoU, +1.8 NDS, +1.5 mIoU, and +1.8 mIoU\ngains on each benchmark, respectively. We believe that the UniVision framework\ncan serve as a high-performance baseline for the unified vision-centric 3D\nperception task. The code will be available at\n\\url{https://github.com/Cc-Hy/UniVision}.\n", "date": "2024-01-13", "categories": []} {"id": "2401.07477", "title": "CascadeV-Det: Cascade Point Voting for 3D Object Detection", "abstract": " Anchor-free object detectors are highly efficient in performing point-based\nprediction without the need for extra post-processing of anchors. However,\ndifferent from the 2D grids, the 3D points used in these detectors are often\nfar from the ground truth center, making it challenging to accurately regress\nthe bounding boxes. To address this issue, we propose a Cascade Voting\n(CascadeV) strategy that provides high-quality 3D object detection with\npoint-based prediction. Specifically, CascadeV performs cascade detection using\na novel Cascade Voting decoder that combines two new components: Instance Aware\nVoting (IA-Voting) and a Cascade Point Assignment (CPA) module. The IA-Voting\nmodule updates the object features of updated proposal points within the\nbounding box using conditional inverse distance weighting. This approach\nprevents features from being aggregated outside the instance and helps improve\nthe accuracy of object detection. Additionally, since model training can suffer\nfrom a lack of proposal points with high centerness, we have developed the CPA\nmodule to narrow down the positive assignment threshold with cascade stages.\nThis approach relaxes the dependence on proposal centerness in the early stages\nwhile ensuring an ample quantity of positives with high centerness in the later\nstages. Experiments show that FCAF3D with our CascadeV achieves\nstate-of-the-art 3D object detection results with 70.4\\% mAP@0.25 and 51.6\\%\nmAP@0.5 on SUN RGB-D and competitive results on ScanNet. Code will be released\nat https://github.com/Sharpiless/CascadeV-Det\n", "date": "2024-01-15", "categories": []} {"id": "2305.16829", "title": "BEV-IO: Enhancing Bird's-Eye-View 3D Detection with Instance Occupancy", "abstract": " A popular approach for constructing bird's-eye-view (BEV) representation in\n3D detection is to lift 2D image features onto the viewing frustum space based\non explicitly predicted depth distribution. However, depth distribution can\nonly characterize the 3D geometry of visible object surfaces but fails to\ncapture their internal space and overall geometric structure, leading to sparse\nand unsatisfactory 3D representations. To mitigate this issue, we present\nBEV-IO, a new 3D detection paradigm to enhance BEV representation with instance\noccupancy information. At the core of our method is the newly-designed instance\noccupancy prediction (IOP) module, which aims to infer point-level occupancy\nstatus for each instance in the frustum space. To ensure training efficiency\nwhile maintaining representational flexibility, it is trained using the\ncombination of both explicit and implicit supervision. With the predicted\noccupancy, we further design a geometry-aware feature propagation mechanism\n(GFP), which performs self-attention based on occupancy distribution along each\nray in frustum and is able to enforce instance-level feature consistency. By\nintegrating the IOP module with GFP mechanism, our BEV-IO detector is able to\nrender highly informative 3D scene structures with more comprehensive BEV\nrepresentations. Experimental results demonstrate that BEV-IO can outperform\nstate-of-the-art methods while only adding a negligible increase in parameters\n(0.2%) and computational overhead (0.24%in GFLOPs).\n", "date": "2023-05-26", "categories": []} {"id": "2310.08261", "title": "GraphAlign: Enhancing Accurate Feature Alignment by Graph matching for\n Multi-Modal 3D Object Detection", "abstract": " LiDAR and cameras are complementary sensors for 3D object detection in\nautonomous driving. However, it is challenging to explore the unnatural\ninteraction between point clouds and images, and the critical factor is how to\nconduct feature alignment of heterogeneous modalities. Currently, many methods\nachieve feature alignment by projection calibration only, without considering\nthe problem of coordinate conversion accuracy errors between sensors, leading\nto sub-optimal performance. In this paper, we present GraphAlign, a more\naccurate feature alignment strategy for 3D object detection by graph matching.\nSpecifically, we fuse image features from a semantic segmentation encoder in\nthe image branch and point cloud features from a 3D Sparse CNN in the LiDAR\nbranch. To save computation, we construct the nearest neighbor relationship by\ncalculating Euclidean distance within the subspaces that are divided into the\npoint cloud features. Through the projection calibration between the image and\npoint cloud, we project the nearest neighbors of point cloud features onto the\nimage features. Then by matching the nearest neighbors with a single point\ncloud to multiple images, we search for a more appropriate feature alignment.\nIn addition, we provide a self-attention module to enhance the weights of\nsignificant relations to fine-tune the feature alignment between heterogeneous\nmodalities. Extensive experiments on nuScenes benchmark demonstrate the\neffectiveness and efficiency of our GraphAlign.\n", "date": "2023-10-12", "categories": []} {"id": "2401.05011", "title": "Dual-Perspective Knowledge Enrichment for Semi-Supervised 3D Object\n Detection", "abstract": " Semi-supervised 3D object detection is a promising yet under-explored\ndirection to reduce data annotation costs, especially for cluttered indoor\nscenes. A few prior works, such as SESS and 3DIoUMatch, attempt to solve this\ntask by utilizing a teacher model to generate pseudo-labels for unlabeled\nsamples. However, the availability of unlabeled samples in the 3D domain is\nrelatively limited compared to its 2D counterpart due to the greater effort\nrequired to collect 3D data. Moreover, the loose consistency regularization in\nSESS and restricted pseudo-label selection strategy in 3DIoUMatch lead to\neither low-quality supervision or a limited amount of pseudo labels. To address\nthese issues, we present a novel Dual-Perspective Knowledge Enrichment approach\nnamed DPKE for semi-supervised 3D object detection. Our DPKE enriches the\nknowledge of limited training data, particularly unlabeled data, from two\nperspectives: data-perspective and feature-perspective. Specifically, from the\ndata-perspective, we propose a class-probabilistic data augmentation method\nthat augments the input data with additional instances based on the varying\ndistribution of class probabilities. Our DPKE achieves feature-perspective\nknowledge enrichment by designing a geometry-aware feature matching method that\nregularizes feature-level similarity between object proposals from the student\nand teacher models. Extensive experiments on the two benchmark datasets\ndemonstrate that our DPKE achieves superior performance over existing\nstate-of-the-art approaches under various label ratio conditions. The source\ncode will be made available to the public.\n", "date": "2024-01-10", "categories": []} {"id": "2312.01468", "title": "Exploring Adversarial Robustness of LiDAR-Camera Fusion Model in\n Autonomous Driving", "abstract": " Our study assesses the adversarial robustness of LiDAR-camera fusion models\nin 3D object detection. We introduce an attack technique that, by simply adding\na limited number of physically constrained adversarial points above a car, can\nmake the car undetectable by the fusion model. Experimental results reveal that\neven without changes to the image data channel, the fusion model can be\ndeceived solely by manipulating the LiDAR data channel. This finding raises\nsafety concerns in the field of autonomous driving. Further, we explore how the\nquantity of adversarial points, the distance between the front-near car and the\nLiDAR-equipped car, and various angular factors affect the attack success rate.\nWe believe our research can contribute to the understanding of multi-sensor\nrobustness, offering insights and guidance to enhance the safety of autonomous\ndriving.\n", "date": "2023-12-03", "categories": []} {"id": "2401.04230", "title": "SOAP: Cross-sensor Domain Adaptation for 3D Object Detection Using\n Stationary Object Aggregation Pseudo-labelling", "abstract": " We consider the problem of cross-sensor domain adaptation in the context of\nLiDAR-based 3D object detection and propose Stationary Object Aggregation\nPseudo-labelling (SOAP) to generate high quality pseudo-labels for stationary\nobjects. In contrast to the current state-of-the-art in-domain practice of\naggregating just a few input scans, SOAP aggregates entire sequences of point\nclouds at the input level to reduce the sensor domain gap. Then, by means of\nwhat we call quasi-stationary training and spatial consistency post-processing,\nthe SOAP model generates accurate pseudo-labels for stationary objects, closing\na minimum of 30.3% domain gap compared to few-frame detectors. Our results also\nshow that state-of-the-art domain adaptation approaches can achieve even\ngreater performance in combination with SOAP, in both the unsupervised and\nsemi-supervised settings.\n", "date": "2024-01-08", "categories": []} {"id": "2303.16818", "title": "SimDistill: Simulated Multi-modal Distillation for BEV 3D Object\n Detection", "abstract": " Multi-view camera-based 3D object detection has become popular due to its low\ncost, but accurately inferring 3D geometry solely from camera data remains\nchallenging and may lead to inferior performance. Although distilling precise\n3D geometry knowledge from LiDAR data could help tackle this challenge, the\nbenefits of LiDAR information could be greatly hindered by the significant\nmodality gap between different sensory modalities. To address this issue, we\npropose a Simulated multi-modal Distillation (SimDistill) method by carefully\ncrafting the model architecture and distillation strategy. Specifically, we\ndevise multi-modal architectures for both teacher and student models, including\na LiDAR-camera fusion-based teacher and a simulated fusion-based student. Owing\nto the ``identical'' architecture design, the student can mimic the teacher to\ngenerate multi-modal features with merely multi-view images as input, where a\ngeometry compensation module is introduced to bridge the modality gap.\nFurthermore, we propose a comprehensive multi-modal distillation scheme that\nsupports intra-modal, cross-modal, and multi-modal fusion distillation\nsimultaneously in the Bird's-eye-view space. Incorporating them together, our\nSimDistill can learn better feature representations for 3D object detection\nwhile maintaining a cost-effective camera-only deployment. Extensive\nexperiments validate the effectiveness and superiority of SimDistill over\nstate-of-the-art methods, achieving an improvement of 4.8\\% mAP and 4.1\\% NDS\nover the baseline detector. The source code will be released at\nhttps://github.com/ViTAE-Transformer/SimDistill.\n", "date": "2023-03-29", "categories": []} {"id": "2308.13794", "title": "SOGDet: Semantic-Occupancy Guided Multi-view 3D Object Detection", "abstract": " In the field of autonomous driving, accurate and comprehensive perception of\nthe 3D environment is crucial. Bird's Eye View (BEV) based methods have emerged\nas a promising solution for 3D object detection using multi-view images as\ninput. However, existing 3D object detection methods often ignore the physical\ncontext in the environment, such as sidewalk and vegetation, resulting in\nsub-optimal performance. In this paper, we propose a novel approach called\nSOGDet (Semantic-Occupancy Guided Multi-view 3D Object Detection), that\nleverages a 3D semantic-occupancy branch to improve the accuracy of 3D object\ndetection. In particular, the physical context modeled by semantic occupancy\nhelps the detector to perceive the scenes in a more holistic view. Our SOGDet\nis flexible to use and can be seamlessly integrated with most existing\nBEV-based methods. To evaluate its effectiveness, we apply this approach to\nseveral state-of-the-art baselines and conduct extensive experiments on the\nexclusive nuScenes dataset. Our results show that SOGDet consistently enhance\nthe performance of three baseline methods in terms of nuScenes Detection Score\n(NDS) and mean Average Precision (mAP). This indicates that the combination of\n3D object detection and 3D semantic occupancy leads to a more comprehensive\nperception of the 3D environment, thereby aiding build more robust autonomous\ndriving systems. The codes are available at: https://github.com/zhouqiu/SOGDet.\n", "date": "2023-08-26", "categories": []} {"id": "2401.01918", "title": "Distilling Temporal Knowledge with Masked Feature Reconstruction for 3D\n Object Detection", "abstract": " Striking a balance between precision and efficiency presents a prominent\nchallenge in the bird's-eye-view (BEV) 3D object detection. Although previous\ncamera-based BEV methods achieved remarkable performance by incorporating\nlong-term temporal information, most of them still face the problem of low\nefficiency. One potential solution is knowledge distillation. Existing\ndistillation methods only focus on reconstructing spatial features, while\noverlooking temporal knowledge. To this end, we propose TempDistiller, a\nTemporal knowledge Distiller, to acquire long-term memory from a teacher\ndetector when provided with a limited number of frames. Specifically, a\nreconstruction target is formulated by integrating long-term temporal knowledge\nthrough self-attention operation applied to feature teachers. Subsequently,\nnovel features are generated for masked student features via a generator.\nUltimately, we utilize this reconstruction target to reconstruct the student\nfeatures. In addition, we also explore temporal relational knowledge when\ninputting full frames for the student model. We verify the effectiveness of the\nproposed method on the nuScenes benchmark. The experimental results show our\nmethod obtain an enhancement of +1.6 mAP and +1.1 NDS compared to the baseline,\na speed improvement of approximately 6 FPS after compressing temporal\nknowledge, and the most accurate velocity estimation.\n", "date": "2024-01-03", "categories": []} {"id": "2304.14484", "title": "OriCon3D: Effective 3D Object Detection using Orientation and Confidence", "abstract": " In this paper, we propose an advanced methodology for the detection of 3D\nobjects and precise estimation of their spatial positions from a single image.\nUnlike conventional frameworks that rely solely on center-point and dimension\npredictions, our research leverages a deep convolutional neural network-based\n3D object weighted orientation regression paradigm. These estimates are then\nseamlessly integrated with geometric constraints obtained from a 2D bounding\nbox, resulting in derivation of a comprehensive 3D bounding box. Our novel\nnetwork design encompasses two key outputs. The first output involves the\nestimation of 3D object orientation through the utilization of a\ndiscrete-continuous loss function. Simultaneously, the second output predicts\nobjectivity-based confidence scores with minimal variance. Additionally, we\nalso introduce enhancements to our methodology through the incorporation of\nlightweight residual feature extractors. By combining the derived estimates\nwith the geometric constraints inherent in the 2D bounding box, our approach\nsignificantly improves the accuracy of 3D object pose determination, surpassing\nbaseline methodologies. Our method is rigorously evaluated on the KITTI 3D\nobject detection benchmark, demonstrating superior performance.\n", "date": "2023-04-27", "categories": []} {"id": "2401.01075", "title": "Depth-discriminative Metric Learning for Monocular 3D Object Detection", "abstract": " Monocular 3D object detection poses a significant challenge due to the lack\nof depth information in RGB images. Many existing methods strive to enhance the\nobject depth estimation performance by allocating additional parameters for\nobject depth estimation, utilizing extra modules or data. In contrast, we\nintroduce a novel metric learning scheme that encourages the model to extract\ndepth-discriminative features regardless of the visual attributes without\nincreasing inference time and model size. Our method employs the\ndistance-preserving function to organize the feature space manifold in relation\nto ground-truth object depth. The proposed (K, B, eps)-quasi-isometric loss\nleverages predetermined pairwise distance restriction as guidance for adjusting\nthe distance among object descriptors without disrupting the non-linearity of\nthe natural feature manifold. Moreover, we introduce an auxiliary head for\nobject-wise depth estimation, which enhances depth quality while maintaining\nthe inference time. The broad applicability of our method is demonstrated\nthrough experiments that show improvements in overall performance when\nintegrated into various baselines. The results show that our method\nconsistently improves the performance of various baselines by 23.51% and 5.78%\non average across KITTI and Waymo, respectively.\n", "date": "2024-01-02", "categories": []} {"id": "2312.17260", "title": "TimePillars: Temporally-Recurrent 3D LiDAR Object Detection", "abstract": " Object detection applied to LiDAR point clouds is a relevant task in\nrobotics, and particularly in autonomous driving. Single frame methods,\npredominant in the field, exploit information from individual sensor scans.\nRecent approaches achieve good performance, at relatively low inference time.\nNevertheless, given the inherent high sparsity of LiDAR data, these methods\nstruggle in long-range detection (e.g. 200m) which we deem to be critical in\nachieving safe automation. Aggregating multiple scans not only leads to a\ndenser point cloud representation, but it also brings time-awareness to the\nsystem, and provides information about how the environment is changing.\nSolutions of this kind, however, are often highly problem-specific, demand\ncareful data processing, and tend not to fulfil runtime requirements. In this\ncontext we propose TimePillars, a temporally-recurrent object detection\npipeline which leverages the pillar representation of LiDAR data across time,\nrespecting hardware integration efficiency constraints, and exploiting the\ndiversity and long-range information of the novel Zenseact Open Dataset (ZOD).\nThrough experimentation, we prove the benefits of having recurrency, and show\nhow basic building blocks are enough to achieve robust and efficient results.\n", "date": "2023-12-22", "categories": []} {"id": "2304.00670", "title": "CRN: Camera Radar Net for Accurate, Robust, Efficient 3D Perception", "abstract": " Autonomous driving requires an accurate and fast 3D perception system that\nincludes 3D object detection, tracking, and segmentation. Although recent\nlow-cost camera-based approaches have shown promising results, they are\nsusceptible to poor illumination or bad weather conditions and have a large\nlocalization error. Hence, fusing camera with low-cost radar, which provides\nprecise long-range measurement and operates reliably in all environments, is\npromising but has not yet been thoroughly investigated. In this paper, we\npropose Camera Radar Net (CRN), a novel camera-radar fusion framework that\ngenerates a semantically rich and spatially accurate bird's-eye-view (BEV)\nfeature map for various tasks. To overcome the lack of spatial information in\nan image, we transform perspective view image features to BEV with the help of\nsparse but accurate radar points. We further aggregate image and radar feature\nmaps in BEV using multi-modal deformable attention designed to tackle the\nspatial misalignment between inputs. CRN with real-time setting operates at 20\nFPS while achieving comparable performance to LiDAR detectors on nuScenes, and\neven outperforms at a far distance on 100m setting. Moreover, CRN with offline\nsetting yields 62.4% NDS, 57.5% mAP on nuScenes test set and ranks first among\nall camera and camera-radar 3D object detectors.\n", "date": "2023-04-03", "categories": []} {"id": "2310.11346", "title": "Towards Generalizable Multi-Camera 3D Object Detection via Perspective\n Debiasing", "abstract": " Detecting objects in 3D space using multiple cameras, known as Multi-Camera\n3D Object Detection (MC3D-Det), has gained prominence with the advent of\nbird's-eye view (BEV) approaches. However, these methods often struggle when\nfaced with unfamiliar testing environments due to the lack of diverse training\ndata encompassing various viewpoints and environments. To address this, we\npropose a novel method that aligns 3D detection with 2D camera plane results,\nensuring consistent and accurate detections. Our framework, anchored in\nperspective debiasing, helps the learning of features resilient to domain\nshifts. In our approach, we render diverse view maps from BEV features and\nrectify the perspective bias of these maps, leveraging implicit foreground\nvolumes to bridge the camera and BEV planes. This two-step process promotes the\nlearning of perspective- and context-independent features, crucial for accurate\nobject detection across varying viewpoints, camera parameters, and\nenvironmental conditions. Notably, our model-agnostic approach preserves the\noriginal network structure without incurring additional inference costs,\nfacilitating seamless integration across various models and simplifying\ndeployment. Furthermore, we also show our approach achieves satisfactory\nresults in real data when trained only with virtual datasets, eliminating the\nneed for real scene annotations. Experimental results on both Domain\nGeneralization (DG) and Unsupervised Domain Adaptation (UDA) clearly\ndemonstrate its effectiveness. The codes are available at\nhttps://github.com/EnVision-Research/Generalizable-BEV.\n", "date": "2023-10-17", "categories": []} {"id": "2312.15377", "title": "End-to-End 3D Object Detection using LiDAR Point Cloud", "abstract": " There has been significant progress made in the field of autonomous vehicles.\nObject detection and tracking are the primary tasks for any autonomous vehicle.\nThe task of object detection in autonomous vehicles relies on a variety of\nsensors like cameras, and LiDAR. Although image features are typically\npreferred, numerous approaches take spatial data as input. Exploiting this\ninformation we present an approach wherein, using a novel encoding of the LiDAR\npoint cloud we infer the location of different classes near the autonomous\nvehicles. This approach does not implement a bird's eye view approach, which is\ngenerally applied for this application and thus saves the extensive\npre-processing required. After studying the numerous networks and approaches\nused to solve this approach, we have implemented a novel model with the\nintention to inculcate their advantages and avoid their shortcomings. The\noutput is predictions about the location and orientation of objects in the\nscene in form of 3D bounding boxes and labels of scene objects.\n", "date": "2023-12-24", "categories": []} {"id": "2312.15449", "title": "iDet3D: Towards Efficient Interactive Object Detection for LiDAR Point\n Clouds", "abstract": " Accurately annotating multiple 3D objects in LiDAR scenes is laborious and\nchallenging. While a few previous studies have attempted to leverage\nsemi-automatic methods for cost-effective bounding box annotation, such methods\nhave limitations in efficiently handling numerous multi-class objects. To\neffectively accelerate 3D annotation pipelines, we propose iDet3D, an efficient\ninteractive 3D object detector. Supporting a user-friendly 2D interface, which\ncan ease the cognitive burden of exploring 3D space to provide click\ninteractions, iDet3D enables users to annotate the entire objects in each scene\nwith minimal interactions. Taking the sparse nature of 3D point clouds into\naccount, we design a negative click simulation (NCS) to improve accuracy by\nreducing false-positive predictions. In addition, iDet3D incorporates two click\npropagation techniques to take full advantage of user interactions: (1) dense\nclick guidance (DCG) for keeping user-provided information throughout the\nnetwork and (2) spatial click propagation (SCP) for detecting other instances\nof the same class based on the user-specified objects. Through our extensive\nexperiments, we present that our method can construct precise annotations in a\nfew clicks, which shows the practicality as an efficient annotation tool for 3D\nobject detection.\n", "date": "2023-12-24", "categories": []} {"id": "2312.16141", "title": "VirtualPainting: Addressing Sparsity with Virtual Points and\n Distance-Aware Data Augmentation for 3D Object Detection", "abstract": " In recent times, there has been a notable surge in multimodal approaches that\ndecorates raw LiDAR point clouds with camera-derived features to improve object\ndetection performance. However, we found that these methods still grapple with\nthe inherent sparsity of LiDAR point cloud data, primarily because fewer points\nare enriched with camera-derived features for sparsely distributed objects. We\npresent an innovative approach that involves the generation of virtual LiDAR\npoints using camera images and enhancing these virtual points with semantic\nlabels obtained from image-based segmentation networks to tackle this issue and\nfacilitate the detection of sparsely distributed objects, particularly those\nthat are occluded or distant. Furthermore, we integrate a distance aware data\naugmentation (DADA) technique to enhance the models capability to recognize\nthese sparsely distributed objects by generating specialized training samples.\nOur approach offers a versatile solution that can be seamlessly integrated into\nvarious 3D frameworks and 2D semantic segmentation methods, resulting in\nsignificantly improved overall detection accuracy. Evaluation on the KITTI and\nnuScenes datasets demonstrates substantial enhancements in both 3D and birds\neye view (BEV) detection benchmarks\n", "date": "2023-12-26", "categories": []} {"id": "2312.14465", "title": "FM-OV3D: Foundation Model-based Cross-modal Knowledge Blending for\n Open-Vocabulary 3D Detection", "abstract": " The superior performances of pre-trained foundation models in various visual\ntasks underscore their potential to enhance the 2D models' open-vocabulary\nability. Existing methods explore analogous applications in the 3D space.\nHowever, most of them only center around knowledge extraction from singular\nfoundation models, which limits the open-vocabulary ability of 3D models. We\nhypothesize that leveraging complementary pre-trained knowledge from various\nfoundation models can improve knowledge transfer from 2D pre-trained visual\nlanguage models to the 3D space. In this work, we propose FM-OV3D, a method of\nFoundation Model-based Cross-modal Knowledge Blending for Open-Vocabulary 3D\nDetection, which improves the open-vocabulary localization and recognition\nabilities of 3D model by blending knowledge from multiple pre-trained\nfoundation models, achieving true open-vocabulary without facing constraints\nfrom original 3D datasets. Specifically, to learn the open-vocabulary 3D\nlocalization ability, we adopt the open-vocabulary localization knowledge of\nthe Grounded-Segment-Anything model. For open-vocabulary 3D recognition\nability, We leverage the knowledge of generative foundation models, including\nGPT-3 and Stable Diffusion models, and cross-modal discriminative models like\nCLIP. The experimental results on two popular benchmarks for open-vocabulary 3D\nobject detection show that our model efficiently learns knowledge from multiple\nfoundation models to enhance the open-vocabulary ability of the 3D model and\nsuccessfully achieves state-of-the-art performance in open-vocabulary 3D object\ndetection tasks. Code is released at\nhttps://github.com/dmzhang0425/FM-OV3D.git.\n", "date": "2023-12-22", "categories": []} {"id": "2312.13641", "title": "SPGroup3D: Superpoint Grouping Network for Indoor 3D Object Detection", "abstract": " Current 3D object detection methods for indoor scenes mainly follow the\nvoting-and-grouping strategy to generate proposals. However, most methods\nutilize instance-agnostic groupings, such as ball query, leading to\ninconsistent semantic information and inaccurate regression of the proposals.\nTo this end, we propose a novel superpoint grouping network for indoor\nanchor-free one-stage 3D object detection. Specifically, we first adopt an\nunsupervised manner to partition raw point clouds into superpoints, areas with\nsemantic consistency and spatial similarity. Then, we design a geometry-aware\nvoting module that adapts to the centerness in anchor-free detection by\nconstraining the spatial relationship between superpoints and object centers.\nNext, we present a superpoint-based grouping module to explore the consistent\nrepresentation within proposals. This module includes a superpoint attention\nlayer to learn feature interaction between neighboring superpoints, and a\nsuperpoint-voxel fusion layer to propagate the superpoint-level information to\nthe voxel level. Finally, we employ effective multiple matching to capitalize\non the dynamic receptive fields of proposals based on superpoints during the\ntraining. Experimental results demonstrate our method achieves state-of-the-art\nperformance on ScanNet V2, SUN RGB-D, and S3DIS datasets in the indoor\none-stage 3D object detection. Source code is available at\nhttps://github.com/zyrant/SPGroup3D.\n", "date": "2023-12-21", "categories": []} {"id": "2312.11578", "title": "Diffusion-Based Particle-DETR for BEV Perception", "abstract": " The Bird-Eye-View (BEV) is one of the most widely-used scene representations\nfor visual perception in Autonomous Vehicles (AVs) due to its well suited\ncompatibility to downstream tasks. For the enhanced safety of AVs, modeling\nperception uncertainty in BEV is crucial. Recent diffusion-based methods offer\na promising approach to uncertainty modeling for visual perception but fail to\neffectively detect small objects in the large coverage of the BEV. Such\ndegradation of performance can be attributed primarily to the specific network\narchitectures and the matching strategy used when training. Here, we address\nthis problem by combining the diffusion paradigm with current state-of-the-art\n3D object detectors in BEV. We analyze the unique challenges of this approach,\nwhich do not exist with deterministic detectors, and present a simple technique\nbased on object query interpolation that allows the model to learn positional\ndependencies even in the presence of the diffusion noise. Based on this, we\npresent a diffusion-based DETR model for object detection that bears\nsimilarities to particle methods. Abundant experimentation on the NuScenes\ndataset shows equal or better performance for our generative approach, compared\nto deterministic state-of-the-art methods. Our source code will be made\npublicly available.\n", "date": "2023-12-18", "categories": []} {"id": "2312.12144", "title": "M-BEV: Masked BEV Perception for Robust Autonomous Driving", "abstract": " 3D perception is a critical problem in autonomous driving. Recently, the\nBird-Eye-View (BEV) approach has attracted extensive attention, due to low-cost\ndeployment and desirable vision detection capacity. However, the existing\nmodels ignore a realistic scenario during the driving procedure, i.e., one or\nmore view cameras may be failed, which largely deteriorates the performance. To\ntackle this problem, we propose a generic Masked BEV (M-BEV) perception\nframework, which can effectively improve robustness to this challenging\nscenario, by random masking and reconstructing camera views in the end-to-end\ntraining. More specifically, we develop a novel Masked View Reconstruction\n(MVR) module for M-BEV. It mimics various missing cases by randomly masking\nfeatures of different camera views, then leverages the original features of\nthese views as self-supervision, and reconstructs the masked ones with the\ndistinct spatio-temporal context across views. Via such a plug-and-play MVR,\nour M-BEV is capable of learning the missing views from the resting ones, and\nthus well generalized for robust view recovery and accurate perception in the\ntesting. We perform extensive experiments on the popular NuScenes benchmark,\nwhere our framework can significantly boost 3D perception performance of the\nstate-of-the-art models on various missing view cases, e.g., for the absence of\nback view, our M-BEV promotes the PETRv2 model with 10.3% mAP gain.\n", "date": "2023-12-19", "categories": []} {"id": "2308.09616", "title": "Far3D: Expanding the Horizon for Surround-view 3D Object Detection", "abstract": " Recently 3D object detection from surround-view images has made notable\nadvancements with its low deployment cost. However, most works have primarily\nfocused on close perception range while leaving long-range detection less\nexplored. Expanding existing methods directly to cover long distances poses\nchallenges such as heavy computation costs and unstable convergence. To address\nthese limitations, this paper proposes a novel sparse query-based framework,\ndubbed Far3D. By utilizing high-quality 2D object priors, we generate 3D\nadaptive queries that complement the 3D global queries. To efficiently capture\ndiscriminative features across different views and scales for long-range\nobjects, we introduce a perspective-aware aggregation module. Additionally, we\npropose a range-modulated 3D denoising approach to address query error\npropagation and mitigate convergence issues in long-range tasks. Significantly,\nFar3D demonstrates SoTA performance on the challenging Argoverse 2 dataset,\ncovering a wide range of 150 meters, surpassing several LiDAR-based approaches.\nMeanwhile, Far3D exhibits superior performance compared to previous methods on\nthe nuScenes dataset. The code is available at\nhttps://github.com/megvii-research/Far3D.\n", "date": "2023-08-18", "categories": []} {"id": "2312.10390", "title": "Not Every Side Is Equal: Localization Uncertainty Estimation for\n Semi-Supervised 3D Object Detection", "abstract": " Semi-supervised 3D object detection from point cloud aims to train a detector\nwith a small number of labeled data and a large number of unlabeled data. The\ncore of existing methods lies in how to select high-quality pseudo-labels using\nthe designed quality evaluation criterion. However, these methods treat each\npseudo bounding box as a whole and assign equal importance to each side during\ntraining, which is detrimental to model performance due to many sides having\npoor localization quality. Besides, existing methods filter out a large number\nof low-quality pseudo-labels, which also contain some correct regression values\nthat can help with model training. To address the above issues, we propose a\nside-aware framework for semi-supervised 3D object detection consisting of\nthree key designs: a 3D bounding box parameterization method, an uncertainty\nestimation module, and a pseudo-label selection strategy. These modules work\ntogether to explicitly estimate the localization quality of each side and\nassign different levels of importance during the training phase. Extensive\nexperiment results demonstrate that the proposed method can consistently\noutperform baseline models under different scenes and evaluation criteria.\nMoreover, our method achieves state-of-the-art performance on three datasets\nwith different labeled ratios.\n", "date": "2023-12-16", "categories": []} {"id": "2312.08876", "title": "OpenSight: A Simple Open-Vocabulary Framework for LiDAR-Based Object\n Detection", "abstract": " Traditional LiDAR-based object detection research primarily focuses on\nclosed-set scenarios, which falls short in complex real-world applications.\nDirectly transferring existing 2D open-vocabulary models with some known LiDAR\nclasses for open-vocabulary ability, however, tends to suffer from over-fitting\nproblems: The obtained model will detect the known objects, even presented with\na novel category. In this paper, we propose OpenSight, a more advanced 2D-3D\nmodeling framework for LiDAR-based open-vocabulary detection. OpenSight\nutilizes 2D-3D geometric priors for the initial discernment and localization of\ngeneric objects, followed by a more specific semantic interpretation of the\ndetected objects. The process begins by generating 2D boxes for generic objects\nfrom the accompanying camera images of LiDAR. These 2D boxes, together with\nLiDAR points, are then lifted back into the LiDAR space to estimate\ncorresponding 3D boxes. For better generic object perception, our framework\nintegrates both temporal and spatial-aware constraints. Temporal awareness\ncorrelates the predicted 3D boxes across consecutive timestamps, recalibrating\nthe missed or inaccurate boxes. The spatial awareness randomly places some\n``precisely'' estimated 3D boxes at varying distances, increasing the\nvisibility of generic objects. To interpret the specific semantics of detected\nobjects, we develop a cross-modal alignment and fusion module to first align 3D\nfeatures with 2D image embeddings and then fuse the aligned 3D-2D features for\nsemantic decoding. Our experiments indicate that our method establishes\nstate-of-the-art open-vocabulary performance on widely used 3D detection\nbenchmarks and effectively identifies objects for new categories of interest.\n", "date": "2023-12-12", "categories": []} {"id": "2312.09082", "title": "Learned Fusion: 3D Object Detection using Calibration-Free Transformer\n Feature Fusion", "abstract": " The state of the art in 3D object detection using sensor fusion heavily\nrelies on calibration quality, which is difficult to maintain in large scale\ndeployment outside a lab environment. We present the first calibration-free\napproach for 3D object detection. Thus, eliminating the need for complex and\ncostly calibration procedures. Our approach uses transformers to map the\nfeatures between multiple views of different sensors at multiple abstraction\nlevels. In an extensive evaluation for object detection, we not only show that\nour approach outperforms single modal setups by 14.1% in BEV mAP, but also that\nthe transformer indeed learns mapping. By showing calibration is not necessary\nfor sensor fusion, we hope to motivate other researchers following the\ndirection of calibration-free fusion. Additionally, resulting approaches have a\nsubstantial resilience against rotation and translation changes.\n", "date": "2023-12-14", "categories": []} {"id": "2302.02367", "title": "FastPillars: A Deployment-friendly Pillar-based 3D Detector", "abstract": " The deployment of 3D detectors strikes one of the major challenges in\nreal-world self-driving scenarios. Existing BEV-based (i.e., Bird Eye View)\ndetectors favor sparse convolutions (known as SPConv) to speed up training and\ninference, which puts a hard barrier for deployment, especially for on-device\napplications. In this paper, to tackle the challenge of efficient 3D object\ndetection from an industry perspective, we devise a deployment-friendly\npillar-based 3D detector, termed FastPillars. First, we introduce a novel\nlightweight Max-and-Attention Pillar Encoding (MAPE) module specially for\nenhancing small 3D objects. Second, we propose a simple yet effective principle\nfor designing a backbone in pillar-based 3D detection. We construct FastPillars\nbased on these designs, achieving high performance and low latency without\nSPConv. Extensive experiments on two large-scale datasets demonstrate the\neffectiveness and efficiency of FastPillars for on-device 3D detection\nregarding both performance and speed. Specifically, FastPillars delivers\nstate-of-the-art accuracy on Waymo Open Dataset with 1.8X speed up and 3.8\nmAPH/L2 improvement over CenterPoint (SPConv-based). Our code is publicly\navailable at: https://github.com/StiphyJay/FastPillars.\n", "date": "2023-02-05", "categories": []} {"id": "2312.08004", "title": "Instance-aware Multi-Camera 3D Object Detection with Structural Priors\n Mining and Self-Boosting Learning", "abstract": " Camera-based bird-eye-view (BEV) perception paradigm has made significant\nprogress in the autonomous driving field. Under such a paradigm, accurate BEV\nrepresentation construction relies on reliable depth estimation for\nmulti-camera images. However, existing approaches exhaustively predict depths\nfor every pixel without prioritizing objects, which are precisely the entities\nrequiring detection in the 3D space. To this end, we propose IA-BEV, which\nintegrates image-plane instance awareness into the depth estimation process\nwithin a BEV-based detector. First, a category-specific structural priors\nmining approach is proposed for enhancing the efficacy of monocular depth\ngeneration. Besides, a self-boosting learning strategy is further proposed to\nencourage the model to place more emphasis on challenging objects in\ncomputation-expensive temporal stereo matching. Together they provide advanced\ndepth estimation results for high-quality BEV features construction, benefiting\nthe ultimate 3D detection. The proposed method achieves state-of-the-art\nperformances on the challenging nuScenes benchmark, and extensive experimental\nresults demonstrate the effectiveness of our designs.\n", "date": "2023-12-13", "categories": []} {"id": "2312.05277", "title": "3D Copy-Paste: Physically Plausible Object Insertion for Monocular 3D\n Detection", "abstract": " A major challenge in monocular 3D object detection is the limited diversity\nand quantity of objects in real datasets. While augmenting real scenes with\nvirtual objects holds promise to improve both the diversity and quantity of the\nobjects, it remains elusive due to the lack of an effective 3D object insertion\nmethod in complex real captured scenes. In this work, we study augmenting\ncomplex real indoor scenes with virtual objects for monocular 3D object\ndetection. The main challenge is to automatically identify plausible physical\nproperties for virtual assets (e.g., locations, appearances, sizes, etc.) in\ncluttered real scenes. To address this challenge, we propose a physically\nplausible indoor 3D object insertion approach to automatically copy virtual\nobjects and paste them into real scenes. The resulting objects in scenes have\n3D bounding boxes with plausible physical locations and appearances. In\nparticular, our method first identifies physically feasible locations and poses\nfor the inserted objects to prevent collisions with the existing room layout.\nSubsequently, it estimates spatially-varying illumination for the insertion\nlocation, enabling the immersive blending of the virtual objects into the\noriginal scene with plausible appearances and cast shadows. We show that our\naugmentation method significantly improves existing monocular 3D object models\nand achieves state-of-the-art performance. For the first time, we demonstrate\nthat a physically plausible 3D object insertion, serving as a generative data\naugmentation technique, can lead to significant improvements for discriminative\ndownstream tasks such as monocular 3D object detection. Project website:\nhttps://gyhandy.github.io/3D-Copy-Paste/\n", "date": "2023-12-08", "categories": []} {"id": "2312.06113", "title": "SimMining-3D: Altitude-Aware 3D Object Detection in Complex Mining\n Environments: A Novel Dataset and ROS-Based Automatic Annotation Pipeline", "abstract": " Accurate and efficient object detection is crucial for safe and efficient\noperation of earth-moving equipment in mining. Traditional 2D image-based\nmethods face limitations in dynamic and complex mine environments. To overcome\nthese challenges, 3D object detection using point cloud data has emerged as a\ncomprehensive approach. However, training models for mining scenarios is\nchallenging due to sensor height variations, viewpoint changes, and the need\nfor diverse annotated datasets. This paper presents novel contributions to\naddress these challenges. We introduce a synthetic dataset SimMining 3D [1]\nspecifically designed for 3D object detection in mining environments. The\ndataset captures objects and sensors positioned at various heights within mine\nbenches, accurately reflecting authentic mining scenarios. An automatic\nannotation pipeline through ROS interface reduces manual labor and accelerates\ndataset creation. We propose evaluation metrics accounting for sensor-to-object\nheight variations and point cloud density, enabling accurate model assessment\nin mining scenarios. Real data tests validate our models effectiveness in\nobject prediction. Our ablation study emphasizes the importance of altitude and\nheight variation augmentations in improving accuracy and reliability. The\npublicly accessible synthetic dataset [1] serves as a benchmark for supervised\nlearning and advances object detection techniques in mining with complimentary\npointwise annotations for each scene. In conclusion, our work bridges the gap\nbetween synthetic and real data, addressing the domain shift challenge in 3D\nobject detection for mining. We envision robust object detection systems\nenhancing safety and efficiency in mining and related domains.\n", "date": "2023-12-11", "categories": []} {"id": "2312.03808", "title": "SurfaceAug: Closing the Gap in Multimodal Ground Truth Sampling", "abstract": " Despite recent advances in both model architectures and data augmentation,\nmultimodal object detectors still barely outperform their LiDAR-only\ncounterparts. This shortcoming has been attributed to a lack of sufficiently\npowerful multimodal data augmentation. To address this, we present SurfaceAug,\na novel ground truth sampling algorithm. SurfaceAug pastes objects by\nresampling both images and point clouds, enabling object-level transformations\nin both modalities. We evaluate our algorithm by training a multimodal detector\non KITTI and compare its performance to previous works. We show experimentally\nthat SurfaceAug outperforms existing methods on car detection tasks and\nestablishes a new state of the art for multimodal ground truth sampling.\n", "date": "2023-12-06", "categories": []} {"id": "2302.02858", "title": "TR3D: Towards Real-Time Indoor 3D Object Detection", "abstract": " Recently, sparse 3D convolutions have changed 3D object detection. Performing\non par with the voting-based approaches, 3D CNNs are memory-efficient and scale\nto large scenes better. However, there is still room for improvement. With a\nconscious, practice-oriented approach to problem-solving, we analyze the\nperformance of such methods and localize the weaknesses. Applying modifications\nthat resolve the found issues one by one, we end up with TR3D: a fast\nfully-convolutional 3D object detection model trained end-to-end, that achieves\nstate-of-the-art results on the standard benchmarks, ScanNet v2, SUN RGB-D, and\nS3DIS. Moreover, to take advantage of both point cloud and RGB inputs, we\nintroduce an early fusion of 2D and 3D features. We employ our fusion module to\nmake conventional 3D object detection methods multimodal and demonstrate an\nimpressive boost in performance. Our model with early feature fusion, which we\nrefer to as TR3D+FF, outperforms existing 3D object detection approaches on the\nSUN RGB-D dataset. Overall, besides being accurate, both TR3D and TR3D+FF\nmodels are lightweight, memory-efficient, and fast, thereby marking another\nmilestone on the way toward real-time 3D object detection. Code is available at\nhttps://github.com/SamsungLabs/tr3d .\n", "date": "2023-02-06", "categories": []} {"id": "2312.02966", "title": "Diffusion-SS3D: Diffusion Model for Semi-supervised 3D Object Detection", "abstract": " Semi-supervised object detection is crucial for 3D scene understanding,\nefficiently addressing the limitation of acquiring large-scale 3D bounding box\nannotations. Existing methods typically employ a teacher-student framework with\npseudo-labeling to leverage unlabeled point clouds. However, producing reliable\npseudo-labels in a diverse 3D space still remains challenging. In this work, we\npropose Diffusion-SS3D, a new perspective of enhancing the quality of\npseudo-labels via the diffusion model for semi-supervised 3D object detection.\nSpecifically, we include noises to produce corrupted 3D object size and class\nlabel distributions, and then utilize the diffusion model as a denoising\nprocess to obtain bounding box outputs. Moreover, we integrate the diffusion\nmodel into the teacher-student framework, so that the denoised bounding boxes\ncan be used to improve pseudo-label generation, as well as the entire\nsemi-supervised learning process. We conduct experiments on the ScanNet and SUN\nRGB-D benchmark datasets to demonstrate that our approach achieves\nstate-of-the-art performance against existing methods. We also present\nextensive analysis to understand how our diffusion model design affects\nperformance in semi-supervised learning.\n", "date": "2023-12-05", "categories": []} {"id": "2308.03982", "title": "PARTNER: Level up the Polar Representation for LiDAR 3D Object Detection", "abstract": " Recently, polar-based representation has shown promising properties in\nperceptual tasks. In addition to Cartesian-based approaches, which separate\npoint clouds unevenly, representing point clouds as polar grids has been\nrecognized as an alternative due to (1) its advantage in robust performance\nunder different resolutions and (2) its superiority in streaming-based\napproaches. However, state-of-the-art polar-based detection methods inevitably\nsuffer from the feature distortion problem because of the non-uniform division\nof polar representation, resulting in a non-negligible performance gap compared\nto Cartesian-based approaches. To tackle this issue, we present PARTNER, a\nnovel 3D object detector in the polar coordinate. PARTNER alleviates the\ndilemma of feature distortion with global representation re-alignment and\nfacilitates the regression by introducing instance-level geometric information\ninto the detection head. Extensive experiments show overwhelming advantages in\nstreaming-based detection and different resolutions. Furthermore, our method\noutperforms the previous polar-based works with remarkable margins of 3.68% and\n9.15% on Waymo and ONCE validation set, thus achieving competitive results over\nthe state-of-the-art methods.\n", "date": "2023-08-08", "categories": []} {"id": "2312.01148", "title": "Has Anything Changed? 3D Change Detection by 2D Segmentation Masks", "abstract": " As capturing devices become common, 3D scans of interior spaces are acquired\non a daily basis. Through scene comparison over time, information about objects\nin the scene and their changes is inferred. This information is important for\nrobots and AR and VR devices, in order to operate in an immersive virtual\nexperience. We thus propose an unsupervised object discovery method that\nidentifies added, moved, or removed objects without any prior knowledge of what\nobjects exist in the scene. We model this problem as a combination of a 3D\nchange detection and a 2D segmentation task. Our algorithm leverages generic 2D\nsegmentation masks to refine an initial but incomplete set of 3D change\ndetections. The initial changes, acquired through render-and-compare likely\ncorrespond to movable objects. The incomplete detections are refined through\ngraph optimization, distilling the information of the 2D segmentation masks in\nthe 3D space. Experiments on the 3Rscan dataset prove that our method\noutperforms competitive baselines, with SoTA results.\n", "date": "2023-12-02", "categories": []} {"id": "2310.17462", "title": "Towards Learning Monocular 3D Object Localization From 2D Labels using\n the Physical Laws of Motion", "abstract": " We present a novel method for precise 3D object localization in single images\nfrom a single calibrated camera using only 2D labels. No expensive 3D labels\nare needed. Thus, instead of using 3D labels, our model is trained with\neasy-to-annotate 2D labels along with the physical knowledge of the object's\nmotion. Given this information, the model can infer the latent third dimension,\neven though it has never seen this information during training. Our method is\nevaluated on both synthetic and real-world datasets, and we are able to achieve\na mean distance error of just 6 cm in our experiments on real data. The results\nindicate the method's potential as a step towards learning 3D object location\nestimation, where collecting 3D data for training is not feasible.\n", "date": "2023-10-26", "categories": []} {"id": "2311.17770", "title": "PillarNeSt: Embracing Backbone Scaling and Pretraining for Pillar-based\n 3D Object Detection", "abstract": " This paper shows the effectiveness of 2D backbone scaling and pretraining for\npillar-based 3D object detectors. Pillar-based methods mainly employ randomly\ninitialized 2D convolution neural network (ConvNet) for feature extraction and\nfail to enjoy the benefits from the backbone scaling and pretraining in the\nimage domain. To show the scaling-up capacity in point clouds, we introduce the\ndense ConvNet pretrained on large-scale image datasets (e.g., ImageNet) as the\n2D backbone of pillar-based detectors. The ConvNets are adaptively designed\nbased on the model size according to the specific features of point clouds,\nsuch as sparsity and irregularity. Equipped with the pretrained ConvNets, our\nproposed pillar-based detector, termed PillarNeSt, outperforms the existing 3D\nobject detectors by a large margin on the nuScenes and Argoversev2 datasets.\nOur code shall be released upon acceptance.\n", "date": "2023-11-29", "categories": []} {"id": "2303.12976", "title": "NVAutoNet: Fast and Accurate 360$^{\\circ}$ 3D Visual Perception For Self\n Driving", "abstract": " Achieving robust and real-time 3D perception is fundamental for autonomous\nvehicles. While most existing 3D perception methods prioritize detection\naccuracy, they often overlook critical aspects such as computational\nefficiency, onboard chip deployment friendliness, resilience to sensor mounting\ndeviations, and adaptability to various vehicle types. To address these\nchallenges, we present NVAutoNet: a specialized Bird's-Eye-View (BEV)\nperception network tailored explicitly for automated vehicles. NVAutoNet takes\nsynchronized camera images as input and predicts 3D signals like obstacles,\nfreespaces, and parking spaces. The core of NVAutoNet's architecture (image and\nBEV backbones) relies on efficient convolutional networks, optimized for high\nperformance using TensorRT. More importantly, our image-to-BEV transformation\nemploys simple linear layers and BEV look-up tables, ensuring rapid inference\nspeed. Trained on an extensive proprietary dataset, NVAutoNet consistently\nachieves elevated perception accuracy, operating remarkably at 53 frames per\nsecond on the NVIDIA DRIVE Orin SoC. Notably, NVAutoNet demonstrates resilience\nto sensor mounting deviations arising from diverse car models. Moreover,\nNVAutoNet excels in adapting to varied vehicle types, facilitated by\ninexpensive model fine-tuning procedures that expedite compatibility\nadjustments.\n", "date": "2023-03-23", "categories": []} {"id": "2311.15615", "title": "Technical Report for Argoverse Challenges on Unified Sensor-based\n Detection, Tracking, and Forecasting", "abstract": " This report presents our Le3DE2E solution for unified sensor-based detection,\ntracking, and forecasting in Argoverse Challenges at CVPR 2023 Workshop on\nAutonomous Driving (WAD). We propose a unified network that incorporates three\ntasks, including detection, tracking, and forecasting. This solution adopts a\nstrong Bird's Eye View (BEV) encoder with spatial and temporal fusion and\ngenerates unified representations for multi-tasks. The solution was tested in\nthe Argoverse 2 sensor dataset to evaluate the detection, tracking, and\nforecasting of 26 object categories. We achieved 1st place in Detection,\nTracking, and Forecasting on the E2E Forecasting track in Argoverse Challenges\nat CVPR 2023 WAD.\n", "date": "2023-11-27", "categories": []} {"id": "2311.06054", "title": "Refining the ONCE Benchmark with Hyperparameter Tuning", "abstract": " In response to the growing demand for 3D object detection in applications\nsuch as autonomous driving, robotics, and augmented reality, this work focuses\non the evaluation of semi-supervised learning approaches for point cloud data.\nThe point cloud representation provides reliable and consistent observations\nregardless of lighting conditions, thanks to advances in LiDAR sensors. Data\nannotation is of paramount importance in the context of LiDAR applications, and\nautomating 3D data annotation with semi-supervised methods is a pivotal\nchallenge that promises to reduce the associated workload and facilitate the\nemergence of cost-effective LiDAR solutions. Nevertheless, the task of\nsemi-supervised learning in the context of unordered point cloud data remains\nformidable due to the inherent sparsity and incomplete shapes that hinder the\ngeneration of accurate pseudo-labels. In this study, we consider these\nchallenges by posing the question: \"To what extent does unlabelled data\ncontribute to the enhancement of model performance?\" We show that improvements\nfrom previous semi-supervised methods may not be as profound as previously\nthought. Our results suggest that simple grid search hyperparameter tuning\napplied to a supervised model can lead to state-of-the-art performance on the\nONCE dataset, while the contribution of unlabelled data appears to be\ncomparatively less exceptional.\n", "date": "2023-11-10", "categories": []} {"id": "2302.06494", "title": "Explicit3D: Graph Network with Spatial Inference for Single Image 3D\n Object Detection", "abstract": " Indoor 3D object detection is an essential task in single image scene\nunderstanding, impacting spatial cognition fundamentally in visual reasoning.\nExisting works on 3D object detection from a single image either pursue this\ngoal through independent predictions of each object or implicitly reason over\nall possible objects, failing to harness relational geometric information\nbetween objects. To address this problem, we propose a dynamic sparse graph\npipeline named Explicit3D based on object geometry and semantics features.\nTaking the efficiency into consideration, we further define a relatedness score\nand design a novel dynamic pruning algorithm followed by a cluster sampling\nmethod for sparse scene graph generation and updating. Furthermore, our\nExplicit3D introduces homogeneous matrices and defines new relative loss and\ncorner loss to model the spatial difference between target pairs explicitly.\nInstead of using ground-truth labels as direct supervision, our relative and\ncorner loss are derived from the homogeneous transformation, which renders the\nmodel to learn the geometric consistency between objects. The experimental\nresults on the SUN RGB-D dataset demonstrate that our Explicit3D achieves\nbetter performance balance than the-state-of-the-art.\n", "date": "2023-02-13", "categories": []} {"id": "2311.10261", "title": "Vision meets mmWave Radar: 3D Object Perception Benchmark for Autonomous\n Driving", "abstract": " Sensor fusion is crucial for an accurate and robust perception system on\nautonomous vehicles. Most existing datasets and perception solutions focus on\nfusing cameras and LiDAR. However, the collaboration between camera and radar\nis significantly under-exploited. The incorporation of rich semantic\ninformation from the camera, and reliable 3D information from the radar can\npotentially achieve an efficient, cheap, and portable solution for 3D object\nperception tasks. It can also be robust to different lighting or all-weather\ndriving scenarios due to the capability of mmWave radars. In this paper, we\nintroduce the CRUW3D dataset, including 66K synchronized and well-calibrated\ncamera, radar, and LiDAR frames in various driving scenarios. Unlike other\nlarge-scale autonomous driving datasets, our radar data is in the format of\nradio frequency (RF) tensors that contain not only 3D location information but\nalso spatio-temporal semantic information. This kind of radar format can enable\nmachine learning models to generate more reliable object perception results\nafter interacting and fusing the information or features between the camera and\nradar.\n", "date": "2023-11-17", "categories": []} {"id": "2310.17842", "title": "What You See Is What You Detect: Towards better Object Densification in\n 3D detection", "abstract": " Recent works have demonstrated the importance of object completion in 3D\nPerception from Lidar signal. Several methods have been proposed in which\nmodules were used to densify the point clouds produced by laser scanners,\nleading to better recall and more accurate results. Pursuing in that direction,\nwe present, in this work, a counter-intuitive perspective: the widely-used\nfull-shape completion approach actually leads to a higher error-upper bound\nespecially for far away objects and small objects like pedestrians. Based on\nthis observation, we introduce a visible part completion method that requires\nonly 11.3\\% of the prediction points that previous methods generate. To recover\nthe dense representation, we propose a mesh-deformation-based method to augment\nthe point set associated with visible foreground objects. Considering that our\napproach focuses only on the visible part of the foreground objects to achieve\naccurate 3D detection, we named our method What You See Is What You Detect\n(WYSIWYD). Our proposed method is thus a detector-independent model that\nconsists of 2 parts: an Intra-Frustum Segmentation Transformer (IFST) and a\nMesh Depth Completion Network(MDCNet) that predicts the foreground depth from\nmesh deformation. This way, our model does not require the time-consuming\nfull-depth completion task used by most pseudo-lidar-based methods. Our\nexperimental evaluation shows that our approach can provide up to 12.2\\%\nperformance improvements over most of the public baseline models on the KITTI\nand NuScenes dataset bringing the state-of-the-art to a new level. The codes\nwill be available at\n\\textcolor[RGB]{0,0,255}{\\url{{https://github.com/Orbis36/WYSIWYD}}\n", "date": "2023-10-27", "categories": []} {"id": "2306.17551", "title": "Comparative study of subset selection methods for rapid prototyping of\n 3D object detection algorithms", "abstract": " Object detection in 3D is a crucial aspect in the context of autonomous\nvehicles and drones. However, prototyping detection algorithms is\ntime-consuming and costly in terms of energy and environmental impact. To\naddress these challenges, one can check the effectiveness of different models\nby training on a subset of the original training set. In this paper, we present\na comparison of three algorithms for selecting such a subset - random sampling,\nrandom per class sampling, and our proposed MONSPeC (Maximum Object Number\nSampling per Class). We provide empirical evidence for the superior\neffectiveness of random per class sampling and MONSPeC over basic random\nsampling. By replacing random sampling with one of the more efficient\nalgorithms, the results obtained on the subset are more likely to transfer to\nthe results on the entire dataset. The code is available at:\nhttps://github.com/vision-agh/monspec.\n", "date": "2023-06-30", "categories": []} {"id": "2311.07152", "title": "Detecting As Labeling: Rethinking LiDAR-camera Fusion in 3D Object\n Detection", "abstract": " 3D object Detection with LiDAR-camera encounters overfitting in algorithm\ndevelopment which is derived from the violation of some fundamental rules. We\nrefer to the data annotation in dataset construction for theory complementing\nand argue that the regression task prediction should not involve the feature\nfrom the camera branch. By following the cutting-edge perspective of 'Detecting\nAs Labeling', we propose a novel paradigm dubbed DAL. With the most classical\nelementary algorithms, a simple predicting pipeline is constructed by imitating\nthe data annotation process. Then we train it in the simplest way to minimize\nits dependency and strengthen its portability. Though simple in construction\nand training, the proposed DAL paradigm not only substantially pushes the\nperformance boundary but also provides a superior trade-off between speed and\naccuracy among all existing methods. With comprehensive superiority, DAL is an\nideal baseline for both future work development and practical deployment. The\ncode has been released to facilitate future work on\nhttps://github.com/HuangJunJie2017/BEVDet.\n", "date": "2023-11-13", "categories": []} {"id": "2311.06043", "title": "Deep learning for 3D Object Detection and Tracking in Autonomous\n Driving: A Brief Survey", "abstract": " Object detection and tracking are vital and fundamental tasks for autonomous\ndriving, aiming at identifying and locating objects from those predefined\ncategories in a scene. 3D point cloud learning has been attracting more and\nmore attention among all other forms of self-driving data. Currently, there are\nmany deep learning methods for 3D object detection. However, the tasks of\nobject detection and tracking for point clouds still need intensive study due\nto the unique characteristics of point cloud data. To help get a good grasp of\nthe present situation of this research, this paper shows recent advances in\ndeep learning methods for 3D object detection and tracking.\n", "date": "2023-11-10", "categories": []} {"id": "2311.04391", "title": "3DiffTection: 3D Object Detection with Geometry-Aware Diffusion Features", "abstract": " We present 3DiffTection, a state-of-the-art method for 3D object detection\nfrom single images, leveraging features from a 3D-aware diffusion model.\nAnnotating large-scale image data for 3D detection is resource-intensive and\ntime-consuming. Recently, pretrained large image diffusion models have become\nprominent as effective feature extractors for 2D perception tasks. However,\nthese features are initially trained on paired text and image data, which are\nnot optimized for 3D tasks, and often exhibit a domain gap when applied to the\ntarget data. Our approach bridges these gaps through two specialized tuning\nstrategies: geometric and semantic. For geometric tuning, we fine-tune a\ndiffusion model to perform novel view synthesis conditioned on a single image,\nby introducing a novel epipolar warp operator. This task meets two essential\ncriteria: the necessity for 3D awareness and reliance solely on posed image\ndata, which are readily available (e.g., from videos) and does not require\nmanual annotation. For semantic refinement, we further train the model on\ntarget data with detection supervision. Both tuning phases employ ControlNet to\npreserve the integrity of the original feature capabilities. In the final step,\nwe harness these enhanced capabilities to conduct a test-time prediction\nensemble across multiple virtual viewpoints. Through our methodology, we obtain\n3D-aware features that are tailored for 3D detection and excel in identifying\ncross-view point correspondences. Consequently, our model emerges as a powerful\n3D detector, substantially surpassing previous benchmarks, e.g., Cube-RCNN, a\nprecedent in single-view 3D detection by 9.43\\% in AP3D on the\nOmni3D-ARkitscene dataset. Furthermore, 3DiffTection showcases robust data\nefficiency and generalization to cross-domain data.\n", "date": "2023-11-07", "categories": []} {"id": "2206.08171", "title": "K-Radar: 4D Radar Object Detection for Autonomous Driving in Various\n Weather Conditions", "abstract": " Unlike RGB cameras that use visible light bands (384$\\sim$769 THz) and Lidars\nthat use infrared bands (361$\\sim$331 THz), Radars use relatively longer\nwavelength radio bands (77$\\sim$81 GHz), resulting in robust measurements in\nadverse weathers. Unfortunately, existing Radar datasets only contain a\nrelatively small number of samples compared to the existing camera and Lidar\ndatasets. This may hinder the development of sophisticated data-driven deep\nlearning techniques for Radar-based perception. Moreover, most of the existing\nRadar datasets only provide 3D Radar tensor (3DRT) data that contain power\nmeasurements along the Doppler, range, and azimuth dimensions. As there is no\nelevation information, it is challenging to estimate the 3D bounding box of an\nobject from 3DRT. In this work, we introduce KAIST-Radar (K-Radar), a novel\nlarge-scale object detection dataset and benchmark that contains 35K frames of\n4D Radar tensor (4DRT) data with power measurements along the Doppler, range,\nazimuth, and elevation dimensions, together with carefully annotated 3D\nbounding box labels of objects on the roads. K-Radar includes challenging\ndriving conditions such as adverse weathers (fog, rain, and snow) on various\nroad structures (urban, suburban roads, alleyways, and highways). In addition\nto the 4DRT, we provide auxiliary measurements from carefully calibrated\nhigh-resolution Lidars, surround stereo cameras, and RTK-GPS. We also provide\n4DRT-based object detection baseline neural networks (baseline NNs) and show\nthat the height information is crucial for 3D object detection. And by\ncomparing the baseline NN with a similarly-structured Lidar-based neural\nnetwork, we demonstrate that 4D Radar is a more robust sensor for adverse\nweather conditions. All codes are available at\nhttps://github.com/kaist-avelab/k-radar.\n", "date": "2022-06-16", "categories": []} {"id": "2310.18620", "title": "ODM3D: Alleviating Foreground Sparsity for Semi-Supervised Monocular 3D\n Object Detection", "abstract": " Monocular 3D object detection (M3OD) is a significant yet inherently\nchallenging task in autonomous driving due to absence of explicit depth cues in\na single RGB image. In this paper, we strive to boost currently underperforming\nmonocular 3D object detectors by leveraging an abundance of unlabelled data via\nsemi-supervised learning. Our proposed ODM3D framework entails cross-modal\nknowledge distillation at various levels to inject LiDAR-domain knowledge into\na monocular detector during training. By identifying foreground sparsity as the\nmain culprit behind existing methods' suboptimal training, we exploit the\nprecise localisation information embedded in LiDAR points to enable more\nforeground-attentive and efficient distillation via the proposed BEV occupancy\nguidance mask, leading to notably improved knowledge transfer and M3OD\nperformance. Besides, motivated by insights into why existing cross-modal\nGT-sampling techniques fail on our task at hand, we further design a novel\ncross-modal object-wise data augmentation strategy for effective RGB-LiDAR\njoint learning. Our method ranks 1st in both KITTI validation and test\nbenchmarks, significantly surpassing all existing monocular methods, supervised\nor semi-supervised, on both BEV and 3D detection metrics.\n", "date": "2023-10-28", "categories": []} {"id": "2311.03620", "title": "FusionViT: Hierarchical 3D Object Detection via LiDAR-Camera Vision\n Transformer Fusion", "abstract": " For 3D object detection, both camera and lidar have been demonstrated to be\nuseful sensory devices for providing complementary information about the same\nscenery with data representations in different modalities, e.g., 2D RGB image\nvs 3D point cloud. An effective representation learning and fusion of such\nmulti-modal sensor data is necessary and critical for better 3D object\ndetection performance. To solve the problem, in this paper, we will introduce a\nnovel vision transformer-based 3D object detection model, namely FusionViT.\nDifferent from the existing 3D object detection approaches, FusionViT is a\npure-ViT based framework, which adopts a hierarchical architecture by extending\nthe transformer model to embed both images and point clouds for effective\nrepresentation learning. Such multi-modal data embedding representations will\nbe further fused together via a fusion vision transformer model prior to\nfeeding the learned features to the object detection head for both detection\nand localization of the 3D objects in the input scenery. To demonstrate the\neffectiveness of FusionViT, extensive experiments have been done on real-world\ntraffic object detection benchmark datasets KITTI and Waymo Open. Notably, our\nFusionViT model can achieve state-of-the-art performance and outperforms not\nonly the existing baseline methods that merely rely on camera images or lidar\npoint clouds, but also the latest multi-modal image-point cloud deep fusion\napproaches.\n", "date": "2023-11-07", "categories": []} {"id": "2311.03742", "title": "3DifFusionDet: Diffusion Model for 3D Object Detection with Robust\n LiDAR-Camera Fusion", "abstract": " Good 3D object detection performance from LiDAR-Camera sensors demands\nseamless feature alignment and fusion strategies. We propose the 3DifFusionDet\nframework in this paper, which structures 3D object detection as a denoising\ndiffusion process from noisy 3D boxes to target boxes. In this framework,\nground truth boxes diffuse in a random distribution for training, and the model\nlearns to reverse the noising process. During inference, the model gradually\nrefines a set of boxes that were generated at random to the outcomes. Under the\nfeature align strategy, the progressive refinement method could make a\nsignificant contribution to robust LiDAR-Camera fusion. The iterative\nrefinement process could also demonstrate great adaptability by applying the\nframework to various detecting circumstances where varying levels of accuracy\nand speed are required. Extensive experiments on KITTI, a benchmark for\nreal-world traffic object identification, revealed that 3DifFusionDet is able\nto perform favorably in comparison to earlier, well-respected detectors.\n", "date": "2023-11-07", "categories": []} {"id": "2311.04058", "title": "mmFUSION: Multimodal Fusion for 3D Objects Detection", "abstract": " Multi-sensor fusion is essential for accurate 3D object detection in\nself-driving systems. Camera and LiDAR are the most commonly used sensors, and\nusually, their fusion happens at the early or late stages of 3D detectors with\nthe help of regions of interest (RoIs). On the other hand, fusion at the\nintermediate level is more adaptive because it does not need RoIs from\nmodalities but is complex as the features of both modalities are presented from\ndifferent points of view. In this paper, we propose a new intermediate-level\nmulti-modal fusion (mmFUSION) approach to overcome these challenges. First, the\nmmFUSION uses separate encoders for each modality to compute features at a\ndesired lower space volume. Second, these features are fused through\ncross-modality and multi-modality attention mechanisms proposed in mmFUSION.\nThe mmFUSION framework preserves multi-modal information and learns to\ncomplement modalities' deficiencies through attention weights. The strong\nmulti-modal features from the mmFUSION framework are fed to a simple 3D\ndetection head for 3D predictions. We evaluate mmFUSION on the KITTI and\nNuScenes dataset where it performs better than available early, intermediate,\nlate, and even two-stage based fusion schemes. The code with the mmdetection3D\nproject plugin will be publicly available soon.\n", "date": "2023-11-07", "categories": []} {"id": "2301.02364", "title": "Object as Query: Lifting any 2D Object Detector to 3D Detection", "abstract": " 3D object detection from multi-view images has drawn much attention over the\npast few years. Existing methods mainly establish 3D representations from\nmulti-view images and adopt a dense detection head for object detection, or\nemploy object queries distributed in 3D space to localize objects. In this\npaper, we design Multi-View 2D Objects guided 3D Object Detector (MV2D), which\ncan lift any 2D object detector to multi-view 3D object detection. Since 2D\ndetections can provide valuable priors for object existence, MV2D exploits 2D\ndetectors to generate object queries conditioned on the rich image semantics.\nThese dynamically generated queries help MV2D to recall objects in the field of\nview and show a strong capability of localizing 3D objects. For the generated\nqueries, we design a sparse cross attention module to force them to focus on\nthe features of specific objects, which suppresses interference from noises.\nThe evaluation results on the nuScenes dataset demonstrate the dynamic object\nqueries and sparse feature aggregation can promote 3D detection capability.\nMV2D also exhibits a state-of-the-art performance among existing methods. We\nhope MV2D can serve as a new baseline for future research. Code is available at\n\\url{https://github.com/tusen-ai/MV2D}.\n", "date": "2023-01-06", "categories": []} {"id": "2309.03734", "title": "ClusterFusion: Leveraging Radar Spatial Features for Radar-Camera 3D\n Object Detection in Autonomous Vehicles", "abstract": " Thanks to the complementary nature of millimeter wave radar and camera, deep\nlearning-based radar-camera 3D object detection methods may reliably produce\naccurate detections even in low-visibility conditions. This makes them\npreferable to use in autonomous vehicles' perception systems, especially as the\ncombined cost of both sensors is cheaper than the cost of a lidar. Recent\nradar-camera methods commonly perform feature-level fusion which often involves\nprojecting the radar points onto the same plane as the image features and\nfusing the extracted features from both modalities. While performing fusion on\nthe image plane is generally simpler and faster, projecting radar points onto\nthe image plane flattens the depth dimension of the point cloud which might\nlead to information loss and makes extracting the spatial features of the point\ncloud harder. We proposed ClusterFusion, an architecture that leverages the\nlocal spatial features of the radar point cloud by clustering the point cloud\nand performing feature extraction directly on the point cloud clusters before\nprojecting the features onto the image plane. ClusterFusion achieved the\nstate-of-the-art performance among all radar-monocular camera methods on the\ntest slice of the nuScenes dataset with 48.7% nuScenes detection score (NDS).\nWe also investigated the performance of different radar feature extraction\nstrategies on point cloud clusters: a handcrafted strategy, a learning-based\nstrategy, and a combination of both, and found that the handcrafted strategy\nyielded the best performance. The main goal of this work is to explore the use\nof radar's local spatial and point-wise features by extracting them directly\nfrom radar point cloud clusters for a radar-monocular camera 3D object\ndetection method that performs cross-modal feature fusion on the image plane.\n", "date": "2023-09-07", "categories": []} {"id": "2311.01423", "title": "CenterRadarNet: Joint 3D Object Detection and Tracking Framework using\n 4D FMCW Radar", "abstract": " Robust perception is a vital component for ensuring safe autonomous and\nassisted driving. Automotive radar (77 to 81 GHz), which offers\nweather-resilient sensing, provides a complementary capability to the vision-\nor LiDAR-based autonomous driving systems. Raw radio-frequency (RF) radar\ntensors contain rich spatiotemporal semantics besides 3D location information.\nThe majority of previous methods take in 3D (Doppler-range-azimuth) RF radar\ntensors, allowing prediction of an object's location, heading angle, and size\nin bird's-eye-view (BEV). However, they lack the ability to at the same time\ninfer objects' size, orientation, and identity in the 3D space. To overcome\nthis limitation, we propose an efficient joint architecture called\nCenterRadarNet, designed to facilitate high-resolution representation learning\nfrom 4D (Doppler-range-azimuth-elevation) radar data for 3D object detection\nand re-identification (re-ID) tasks. As a single-stage 3D object detector,\nCenterRadarNet directly infers the BEV object distribution confidence maps,\ncorresponding 3D bounding box attributes, and appearance embedding for each\npixel. Moreover, we build an online tracker utilizing the learned appearance\nembedding for re-ID. CenterRadarNet achieves the state-of-the-art result on the\nK-Radar 3D object detection benchmark. In addition, we present the first 3D\nobject-tracking result using radar on the K-Radar dataset V2. In diverse\ndriving scenarios, CenterRadarNet shows consistent, robust performance,\nemphasizing its wide applicability.\n", "date": "2023-11-02", "categories": []} {"id": "2311.02007", "title": "Towards Unsupervised Object Detection From LiDAR Point Clouds", "abstract": " In this paper, we study the problem of unsupervised object detection from 3D\npoint clouds in self-driving scenes. We present a simple yet effective method\nthat exploits (i) point clustering in near-range areas where the point clouds\nare dense, (ii) temporal consistency to filter out noisy unsupervised\ndetections, (iii) translation equivariance of CNNs to extend the auto-labels to\nlong range, and (iv) self-supervision for improving on its own. Our approach,\nOYSTER (Object Discovery via Spatio-Temporal Refinement), does not impose\nconstraints on data collection (such as repeated traversals of the same\nlocation), is able to detect objects in a zero-shot manner without supervised\nfinetuning (even in sparse, distant regions), and continues to self-improve\ngiven more rounds of iterative self-training. To better measure model\nperformance in self-driving scenarios, we propose a new planning-centric\nperception metric based on distance-to-collision. We demonstrate that our\nunsupervised object detector significantly outperforms unsupervised baselines\non PandaSet and Argoverse 2 Sensor dataset, showing promise that\nself-supervision combined with object priors can enable object discovery in the\nwild. For more information, visit the project website:\nhttps://waabi.ai/research/oyster\n", "date": "2023-11-03", "categories": []} {"id": "2311.00986", "title": "M&M3D: Multi-Dataset Training and Efficient Network for Multi-view 3D\n Object Detection", "abstract": " In this research, I proposed a network structure for multi-view 3D object\ndetection using camera-only data and a Bird's-Eye-View map. My work is based on\na current key challenge domain adaptation and visual data transfer. Although\nmany excellent camera-only 3D object detection has been continuously proposed,\nmany research work risk dramatic performance drop when the networks are trained\non the source domain but tested on a different target domain. Then I found it\nis very surprising that predictions on bounding boxes and classes are still\nreplied to on 2D networks. Based on the domain gap assumption on various 3D\ndatasets, I found they still shared a similar data extraction on the same BEV\nmap size and camera data transfer. Therefore, to analyze the domain gap\ninfluence on the current method and to make good use of 3D space information\namong the dataset and the real world, I proposed a transfer learning method and\nTransformer construction to study the 3D object detection on NuScenes-mini and\nLyft. Through multi-dataset training and a detection head from the Transformer,\nthe network demonstrated good data migration performance and efficient\ndetection performance by using 3D anchor query and 3D positional information.\nRelying on only a small amount of source data and the existing large model\npre-training weights, the efficient network manages to achieve competitive\nresults on the new target domain. Moreover, my study utilizes 3D information as\navailable semantic information and 2D multi-view image features blending into\nthe visual-language transfer design. In the final 3D anchor box prediction and\nobject classification, my network achieved good results on standard metrics of\n3D object detection, which differs from dataset-specific models on each\ntraining domain without any fine-tuning.\n", "date": "2023-11-02", "categories": []} {"id": "2309.07084", "title": "SupFusion: Supervised LiDAR-Camera Fusion for 3D Object Detection", "abstract": " In this paper, we propose a novel training strategy called SupFusion, which\nprovides an auxiliary feature level supervision for effective LiDAR-Camera\nfusion and significantly boosts detection performance. Our strategy involves a\ndata enhancement method named Polar Sampling, which densifies sparse objects\nand trains an assistant model to generate high-quality features as the\nsupervision. These features are then used to train the LiDAR-Camera fusion\nmodel, where the fusion feature is optimized to simulate the generated\nhigh-quality features. Furthermore, we propose a simple yet effective deep\nfusion module, which contiguously gains superior performance compared with\nprevious fusion methods with SupFusion strategy. In such a manner, our proposal\nshares the following advantages. Firstly, SupFusion introduces auxiliary\nfeature-level supervision which could boost LiDAR-Camera detection performance\nwithout introducing extra inference costs. Secondly, the proposed deep fusion\ncould continuously improve the detector's abilities. Our proposed SupFusion and\ndeep fusion module is plug-and-play, we make extensive experiments to\ndemonstrate its effectiveness. Specifically, we gain around 2% 3D mAP\nimprovements on KITTI benchmark based on multiple LiDAR-Camera 3D detectors.\n", "date": "2023-09-13", "categories": []} {"id": "2310.20234", "title": "HEDNet: A Hierarchical Encoder-Decoder Network for 3D Object Detection\n in Point Clouds", "abstract": " 3D object detection in point clouds is important for autonomous driving\nsystems. A primary challenge in 3D object detection stems from the sparse\ndistribution of points within the 3D scene. Existing high-performance methods\ntypically employ 3D sparse convolutional neural networks with small kernels to\nextract features. To reduce computational costs, these methods resort to\nsubmanifold sparse convolutions, which prevent the information exchange among\nspatially disconnected features. Some recent approaches have attempted to\naddress this problem by introducing large-kernel convolutions or self-attention\nmechanisms, but they either achieve limited accuracy improvements or incur\nexcessive computational costs. We propose HEDNet, a hierarchical\nencoder-decoder network for 3D object detection, which leverages\nencoder-decoder blocks to capture long-range dependencies among features in the\nspatial space, particularly for large and distant objects. We conducted\nextensive experiments on the Waymo Open and nuScenes datasets. HEDNet achieved\nsuperior detection accuracy on both datasets than previous state-of-the-art\nmethods with competitive efficiency. The code is available at\nhttps://github.com/zhanggang001/HEDNet.\n", "date": "2023-10-31", "categories": []} {"id": "2310.20319", "title": "GACE: Geometry Aware Confidence Enhancement for Black-Box 3D Object\n Detectors on LiDAR-Data", "abstract": " Widely-used LiDAR-based 3D object detectors often neglect fundamental\ngeometric information readily available from the object proposals in their\nconfidence estimation. This is mostly due to architectural design choices,\nwhich were often adopted from the 2D image domain, where geometric context is\nrarely available. In 3D, however, considering the object properties and its\nsurroundings in a holistic way is important to distinguish between true and\nfalse positive detections, e.g. occluded pedestrians in a group. To address\nthis, we present GACE, an intuitive and highly efficient method to improve the\nconfidence estimation of a given black-box 3D object detector. We aggregate\ngeometric cues of detections and their spatial relationships, which enables us\nto properly assess their plausibility and consequently, improve the confidence\nestimation. This leads to consistent performance gains over a variety of\nstate-of-the-art detectors. Across all evaluated detectors, GACE proves to be\nespecially beneficial for the vulnerable road user classes, i.e. pedestrians\nand cyclists.\n", "date": "2023-10-31", "categories": []} {"id": "2310.19405", "title": "Radar-Lidar Fusion for Object Detection by Designing Effective\n Convolution Networks", "abstract": " Object detection is a core component of perception systems, providing the ego\nvehicle with information about its surroundings to ensure safe route planning.\nWhile cameras and Lidar have significantly advanced perception systems, their\nperformance can be limited in adverse weather conditions. In contrast,\nmillimeter-wave technology enables radars to function effectively in such\nconditions. However, relying solely on radar for building a perception system\ndoesn't fully capture the environment due to the data's sparse nature. To\naddress this, sensor fusion strategies have been introduced. We propose a\ndual-branch framework to integrate radar and Lidar data for enhanced object\ndetection. The primary branch focuses on extracting radar features, while the\nauxiliary branch extracts Lidar features. These are then combined using\nadditive attention. Subsequently, the integrated features are processed through\na novel Parallel Forked Structure (PFS) to manage scale variations. A region\nproposal head is then utilized for object detection. We evaluated the\neffectiveness of our proposed method on the Radiate dataset using COCO metrics.\nThe results show that it surpasses state-of-the-art methods by $1.89\\%$ and\n$2.61\\%$ in favorable and adverse weather conditions, respectively. This\nunderscores the value of radar-Lidar fusion in achieving precise object\ndetection and localization, especially in challenging weather conditions.\n", "date": "2023-10-30", "categories": []} {"id": "2310.17659", "title": "RTNH+: Enhanced 4D Radar Object Detection Network using Combined\n CFAR-based Two-level Preprocessing and Vertical Encoding", "abstract": " Four-dimensional (4D) Radar is a useful sensor for 3D object detection and\nthe relative radial speed estimation of surrounding objects under various\nweather conditions. However, since Radar measurements are corrupted with\ninvalid components such as noise, interference, and clutter, it is necessary to\nemploy a preprocessing algorithm before the 3D object detection with neural\nnetworks. In this paper, we propose RTNH+ that is an enhanced version of RTNH,\na 4D Radar object detection network, by two novel algorithms. The first\nalgorithm is the combined constant false alarm rate (CFAR)-based two-level\npreprocessing (CCTP) algorithm that generates two filtered measurements of\ndifferent characteristics using the same 4D Radar measurements, which can\nenrich the information of the input to the 4D Radar object detection network.\nThe second is the vertical encoding (VE) algorithm that effectively encodes\nvertical features of the road objects from the CCTP outputs. We provide details\nof the RTNH+, and demonstrate that RTNH+ achieves significant performance\nimprovement of 10.14\\% in ${{AP}_{3D}^{IoU=0.3}}$ and 16.12\\% in\n${{AP}_{3D}^{IoU=0.5}}$ over RTNH.\n", "date": "2023-10-19", "categories": []} {"id": "2310.16349", "title": "DiffRef3D: A Diffusion-based Proposal Refinement Framework for 3D Object\n Detection", "abstract": " Denoising diffusion models show remarkable performances in generative tasks,\nand their potential applications in perception tasks are gaining interest. In\nthis paper, we introduce a novel framework named DiffRef3D which adopts the\ndiffusion process on 3D object detection with point clouds for the first time.\nSpecifically, we formulate the proposal refinement stage of two-stage 3D object\ndetectors as a conditional diffusion process. During training, DiffRef3D\ngradually adds noise to the residuals between proposals and target objects,\nthen applies the noisy residuals to proposals to generate hypotheses. The\nrefinement module utilizes these hypotheses to denoise the noisy residuals and\ngenerate accurate box predictions. In the inference phase, DiffRef3D generates\ninitial hypotheses by sampling noise from a Gaussian distribution as residuals\nand refines the hypotheses through iterative steps. DiffRef3D is a versatile\nproposal refinement framework that consistently improves the performance of\nexisting 3D object detection models. We demonstrate the significance of\nDiffRef3D through extensive experiments on the KITTI benchmark. Code will be\navailable.\n", "date": "2023-10-25", "categories": []} {"id": "2310.16389", "title": "MVFAN: Multi-View Feature Assisted Network for 4D Radar Object Detection", "abstract": " 4D radar is recognized for its resilience and cost-effectiveness under\nadverse weather conditions, thus playing a pivotal role in autonomous driving.\nWhile cameras and LiDAR are typically the primary sensors used in perception\nmodules for autonomous vehicles, radar serves as a valuable supplementary\nsensor. Unlike LiDAR and cameras, radar remains unimpaired by harsh weather\nconditions, thereby offering a dependable alternative in challenging\nenvironments. Developing radar-based 3D object detection not only augments the\ncompetency of autonomous vehicles but also provides economic benefits. In\nresponse, we propose the Multi-View Feature Assisted Network (\\textit{MVFAN}),\nan end-to-end, anchor-free, and single-stage framework for 4D-radar-based 3D\nobject detection for autonomous vehicles. We tackle the issue of insufficient\nfeature utilization by introducing a novel Position Map Generation module to\nenhance feature learning by reweighing foreground and background points, and\ntheir features, considering the irregular distribution of radar point clouds.\nAdditionally, we propose a pioneering backbone, the Radar Feature Assisted\nbackbone, explicitly crafted to fully exploit the valuable Doppler velocity and\nreflectivity data provided by the 4D radar sensor. Comprehensive experiments\nand ablation studies carried out on Astyx and VoD datasets attest to the\nefficacy of our framework. The incorporation of Doppler velocity and RCS\nreflectivity dramatically improves the detection performance for small moving\nobjects such as pedestrians and cyclists. Consequently, our approach culminates\nin a highly optimized 4D-radar-based 3D object detection capability for\nautonomous driving systems, setting a new standard in the field.\n", "date": "2023-10-25", "categories": []} {"id": "2310.15670", "title": "Leveraging Vision-Centric Multi-Modal Expertise for 3D Object Detection", "abstract": " Current research is primarily dedicated to advancing the accuracy of\ncamera-only 3D object detectors (apprentice) through the knowledge transferred\nfrom LiDAR- or multi-modal-based counterparts (expert). However, the presence\nof the domain gap between LiDAR and camera features, coupled with the inherent\nincompatibility in temporal fusion, significantly hinders the effectiveness of\ndistillation-based enhancements for apprentices. Motivated by the success of\nuni-modal distillation, an apprentice-friendly expert model would predominantly\nrely on camera features, while still achieving comparable performance to\nmulti-modal models. To this end, we introduce VCD, a framework to improve the\ncamera-only apprentice model, including an apprentice-friendly multi-modal\nexpert and temporal-fusion-friendly distillation supervision. The multi-modal\nexpert VCD-E adopts an identical structure as that of the camera-only\napprentice in order to alleviate the feature disparity, and leverages LiDAR\ninput as a depth prior to reconstruct the 3D scene, achieving the performance\non par with other heterogeneous multi-modal experts. Additionally, a\nfine-grained trajectory-based distillation module is introduced with the\npurpose of individually rectifying the motion misalignment for each object in\nthe scene. With those improvements, our camera-only apprentice VCD-A sets new\nstate-of-the-art on nuScenes with a score of 63.1% NDS.\n", "date": "2023-10-24", "categories": []} {"id": "2310.13951", "title": "Fuzzy-NMS: Improving 3D Object Detection with Fuzzy Classification in\n NMS", "abstract": " Non-maximum suppression (NMS) is an essential post-processing module used in\nmany 3D object detection frameworks to remove overlapping candidate bounding\nboxes. However, an overreliance on classification scores and difficulties in\ndetermining appropriate thresholds can affect the resulting accuracy directly.\nTo address these issues, we introduce fuzzy learning into NMS and propose a\nnovel generalized Fuzzy-NMS module to achieve finer candidate bounding box\nfiltering. The proposed Fuzzy-NMS module combines the volume and clustering\ndensity of candidate bounding boxes, refining them with a fuzzy classification\nmethod and optimizing the appropriate suppression thresholds to reduce\nuncertainty in the NMS process. Adequate validation experiments are conducted\nusing the mainstream KITTI and large-scale Waymo 3D object detection\nbenchmarks. The results of these tests demonstrate the proposed Fuzzy-NMS\nmodule can improve the accuracy of numerous recently NMS-based detectors\nsignificantly, including PointPillars, PV-RCNN, and IA-SSD, etc. This effect is\nparticularly evident for small objects such as pedestrians and bicycles. As a\nplug-and-play module, Fuzzy-NMS does not need to be retrained and produces no\nobvious increases in inference time.\n", "date": "2023-10-21", "categories": []} {"id": "2301.09077", "title": "Unleash the Potential of Image Branch for Cross-modal 3D Object\n Detection", "abstract": " To achieve reliable and precise scene understanding, autonomous vehicles\ntypically incorporate multiple sensing modalities to capitalize on their\ncomplementary attributes. However, existing cross-modal 3D detectors do not\nfully utilize the image domain information to address the bottleneck issues of\nthe LiDAR-based detectors. This paper presents a new cross-modal 3D object\ndetector, namely UPIDet, which aims to unleash the potential of the image\nbranch from two aspects. First, UPIDet introduces a new 2D auxiliary task\ncalled normalized local coordinate map estimation. This approach enables the\nlearning of local spatial-aware features from the image modality to supplement\nsparse point clouds. Second, we discover that the representational capability\nof the point cloud backbone can be enhanced through the gradients\nbackpropagated from the training objectives of the image branch, utilizing a\nsuccinct and effective point-to-pixel module. Extensive experiments and\nablation studies validate the effectiveness of our method. Notably, we achieved\nthe top rank in the highly competitive cyclist class of the KITTI benchmark at\nthe time of submission. The source code is available at\nhttps://github.com/Eaphan/UPIDet.\n", "date": "2023-01-22", "categories": []} {"id": "2308.15479", "title": "3D Adversarial Augmentations for Robust Out-of-Domain Predictions", "abstract": " Since real-world training datasets cannot properly sample the long tail of\nthe underlying data distribution, corner cases and rare out-of-domain samples\ncan severely hinder the performance of state-of-the-art models. This problem\nbecomes even more severe for dense tasks, such as 3D semantic segmentation,\nwhere points of non-standard objects can be confidently associated to the wrong\nclass. In this work, we focus on improving the generalization to out-of-domain\ndata. We achieve this by augmenting the training set with adversarial examples.\nFirst, we learn a set of vectors that deform the objects in an adversarial\nfashion. To prevent the adversarial examples from being too far from the\nexisting data distribution, we preserve their plausibility through a series of\nconstraints, ensuring sensor-awareness and shapes smoothness. Then, we perform\nadversarial augmentation by applying the learned sample-independent vectors to\nthe available objects when training a model. We conduct extensive experiments\nacross a variety of scenarios on data from KITTI, Waymo, and CrashD for 3D\nobject detection, and on data from SemanticKITTI, Waymo, and nuScenes for 3D\nsemantic segmentation. Despite training on a standard single dataset, our\napproach substantially improves the robustness and generalization of both 3D\nobject detection and 3D semantic segmentation methods to out-of-domain data.\n", "date": "2023-08-29", "categories": []} {"id": "2212.10230", "title": "A Comprehensive Study of the Robustness for LiDAR-based 3D Object\n Detectors against Adversarial Attacks", "abstract": " Recent years have witnessed significant advancements in deep learning-based\n3D object detection, leading to its widespread adoption in numerous\napplications. As 3D object detectors become increasingly crucial for\nsecurity-critical tasks, it is imperative to understand their robustness\nagainst adversarial attacks. This paper presents the first comprehensive\nevaluation and analysis of the robustness of LiDAR-based 3D detectors under\nadversarial attacks. Specifically, we extend three distinct adversarial attacks\nto the 3D object detection task, benchmarking the robustness of\nstate-of-the-art LiDAR-based 3D object detectors against attacks on the KITTI\nand Waymo datasets. We further analyze the relationship between robustness and\ndetector properties. Additionally, we explore the transferability of\ncross-model, cross-task, and cross-data attacks. Thorough experiments on\ndefensive strategies for 3D detectors are conducted, demonstrating that simple\ntransformations like flipping provide little help in improving robustness when\nthe applied transformation strategy is exposed to attackers. \\revise{Finally,\nwe propose balanced adversarial focal training, based on conventional\nadversarial training, to strike a balance between accuracy and robustness.} Our\nfindings will facilitate investigations into understanding and defending\nagainst adversarial attacks on LiDAR-based 3D object detectors, thus advancing\nthe field. The source code is publicly available at\n\\url{https://github.com/Eaphan/Robust3DOD}.\n", "date": "2022-12-20", "categories": []} {"id": "2310.11316", "title": "MonoSKD: General Distillation Framework for Monocular 3D Object\n Detection via Spearman Correlation Coefficient", "abstract": " Monocular 3D object detection is an inherently ill-posed problem, as it is\nchallenging to predict accurate 3D localization from a single image. Existing\nmonocular 3D detection knowledge distillation methods usually project the LiDAR\nonto the image plane and train the teacher network accordingly. Transferring\nLiDAR-based model knowledge to RGB-based models is more complex, so a general\ndistillation strategy is needed. To alleviate cross-modal prob-lem, we propose\nMonoSKD, a novel Knowledge Distillation framework for Monocular 3D detection\nbased on Spearman correlation coefficient, to learn the relative correlation\nbetween cross-modal features. Considering the large gap between these features,\nstrict alignment of features may mislead the training, so we propose a looser\nSpearman loss. Furthermore, by selecting appropriate distillation locations and\nremoving redundant modules, our scheme saves more GPU resources and trains\nfaster than existing methods. Extensive experiments are performed to verify the\neffectiveness of our framework on the challenging KITTI 3D object detection\nbenchmark. Our method achieves state-of-the-art performance until submission\nwith no additional inference computational cost. Our codes are available at\nhttps://github.com/Senwang98/MonoSKD\n", "date": "2023-10-17", "categories": []} {"id": "2303.05078", "title": "Efficient Transformer-based 3D Object Detection with Dynamic Token\n Halting", "abstract": " Balancing efficiency and accuracy is a long-standing problem for deploying\ndeep learning models. The trade-off is even more important for real-time\nsafety-critical systems like autonomous vehicles. In this paper, we propose an\neffective approach for accelerating transformer-based 3D object detectors by\ndynamically halting tokens at different layers depending on their contribution\nto the detection task. Although halting a token is a non-differentiable\noperation, our method allows for differentiable end-to-end learning by\nleveraging an equivalent differentiable forward-pass. Furthermore, our\nframework allows halted tokens to be reused to inform the model's predictions\nthrough a straightforward token recycling mechanism. Our method significantly\nimproves the Pareto frontier of efficiency versus accuracy when compared with\nthe existing approaches. By halting tokens and increasing model capacity, we\nare able to improve the baseline model's performance without increasing the\nmodel's latency on the Waymo Open Dataset.\n", "date": "2023-03-09", "categories": []} {"id": "2303.09800", "title": "GOOD: General Optimization-based Fusion for 3D Object Detection via\n LiDAR-Camera Object Candidates", "abstract": " 3D object detection serves as the core basis of the perception tasks in\nautonomous driving. Recent years have seen the rapid progress of multi-modal\nfusion strategies for more robust and accurate 3D object detection. However,\ncurrent researches for robust fusion are all learning-based frameworks, which\ndemand a large amount of training data and are inconvenient to implement in new\nscenes. In this paper, we propose GOOD, a general optimization-based fusion\nframework that can achieve satisfying detection without training additional\nmodels and is available for any combinations of 2D and 3D detectors to improve\nthe accuracy and robustness of 3D detection. First we apply the mutual-sided\nnearest-neighbor probability model to achieve the 3D-2D data association. Then\nwe design an optimization pipeline that can optimize different kinds of\ninstances separately based on the matching result. Apart from this, the 3D MOT\nmethod is also introduced to enhance the performance aided by previous frames.\nTo the best of our knowledge, this is the first optimization-based late fusion\nframework for multi-modal 3D object detection which can be served as a baseline\nfor subsequent research. Experiments on both nuScenes and KITTI datasets are\ncarried out and the results show that GOOD outperforms by 9.1\\% on mAP score\ncompared with PointPillars and achieves competitive results with the\nlearning-based late fusion CLOCs.\n", "date": "2023-03-17", "categories": []} {"id": "2310.05447", "title": "Towards Fair and Comprehensive Comparisons for Image-Based 3D Object\n Detection", "abstract": " In this work, we build a modular-designed codebase, formulate strong training\nrecipes, design an error diagnosis toolbox, and discuss current methods for\nimage-based 3D object detection. In particular, different from other highly\nmature tasks, e.g., 2D object detection, the community of image-based 3D object\ndetection is still evolving, where methods often adopt different training\nrecipes and tricks resulting in unfair evaluations and comparisons. What is\nworse, these tricks may overwhelm their proposed designs in performance, even\nleading to wrong conclusions. To address this issue, we build a module-designed\ncodebase and formulate unified training standards for the community.\nFurthermore, we also design an error diagnosis toolbox to measure the detailed\ncharacterization of detection models. Using these tools, we analyze current\nmethods in-depth under varying settings and provide discussions for some open\nquestions, e.g., discrepancies in conclusions on KITTI-3D and nuScenes\ndatasets, which have led to different dominant methods for these datasets. We\nhope that this work will facilitate future research in image-based 3D object\ndetection. Our codes will be released at\n\\url{https://github.com/OpenGVLab/3dodi}\n", "date": "2023-10-09", "categories": []} {"id": "2309.05257", "title": "FusionFormer: A Multi-sensory Fusion in Bird's-Eye-View and Temporal\n Consistent Transformer for 3D Object Detection", "abstract": " Multi-sensor modal fusion has demonstrated strong advantages in 3D object\ndetection tasks. However, existing methods that fuse multi-modal features\nrequire transforming features into the bird's eye view space and may lose\ncertain information on Z-axis, thus leading to inferior performance. To this\nend, we propose a novel end-to-end multi-modal fusion transformer-based\nframework, dubbed FusionFormer, that incorporates deformable attention and\nresidual structures within the fusion encoding module. Specifically, by\ndeveloping a uniform sampling strategy, our method can easily sample from 2D\nimage and 3D voxel features spontaneously, thus exploiting flexible\nadaptability and avoiding explicit transformation to the bird's eye view space\nduring the feature concatenation process. We further implement a residual\nstructure in our feature encoder to ensure the model's robustness in case of\nmissing an input modality. Through extensive experiments on a popular\nautonomous driving benchmark dataset, nuScenes, our method achieves\nstate-of-the-art single model performance of 72.6% mAP and 75.1% NDS in the 3D\nobject detection task without test time augmentation.\n", "date": "2023-09-11", "categories": []} {"id": "2310.05366", "title": "Rotation Matters: Generalized Monocular 3D Object Detection for Various\n Camera Systems", "abstract": " Research on monocular 3D object detection is being actively studied, and as a\nresult, performance has been steadily improving. However, 3D object detection\nperformance is significantly reduced when applied to a camera system different\nfrom the system used to capture the training datasets. For example, a 3D\ndetector trained on datasets from a passenger car mostly fails to regress\naccurate 3D bounding boxes for a camera mounted on a bus. In this paper, we\nconduct extensive experiments to analyze the factors that cause performance\ndegradation. We find that changing the camera pose, especially camera\norientation, relative to the road plane caused performance degradation. In\naddition, we propose a generalized 3D object detection method that can be\nuniversally applied to various camera systems. We newly design a compensation\nmodule that corrects the estimated 3D bounding box location and heading\ndirection. The proposed module can be applied to most of the recent 3D object\ndetection networks. It increases AP3D score (KITTI moderate, IoU $> 70\\%$)\nabout 6-to-10-times above the baselines without additional training. Both\nquantitative and qualitative results show the effectiveness of the proposed\nmethod.\n", "date": "2023-10-09", "categories": []} {"id": "2310.05600", "title": "Care3D: An Active 3D Object Detection Dataset of Real Robotic-Care\n Environments", "abstract": " As labor shortage increases in the health sector, the demand for assistive\nrobotics grows. However, the needed test data to develop those robots is\nscarce, especially for the application of active 3D object detection, where no\nreal data exists at all. This short paper counters this by introducing such an\nannotated dataset of real environments. The captured environments represent\nareas which are already in use in the field of robotic health care research. We\nfurther provide ground truth data within one room, for assessing SLAM\nalgorithms running directly on a health care robot.\n", "date": "2023-10-09", "categories": []} {"id": "2310.05699", "title": "Uni3DETR: Unified 3D Detection Transformer", "abstract": " Existing point cloud based 3D detectors are designed for the particular\nscene, either indoor or outdoor ones. Because of the substantial differences in\nobject distribution and point density within point clouds collected from\nvarious environments, coupled with the intricate nature of 3D metrics, there is\nstill a lack of a unified network architecture that can accommodate diverse\nscenes. In this paper, we propose Uni3DETR, a unified 3D detector that\naddresses indoor and outdoor 3D detection within the same framework.\nSpecifically, we employ the detection transformer with point-voxel interaction\nfor object prediction, which leverages voxel features and points for\ncross-attention and behaves resistant to the discrepancies from data. We then\npropose the mixture of query points, which sufficiently exploits global\ninformation for dense small-range indoor scenes and local information for\nlarge-range sparse outdoor ones. Furthermore, our proposed decoupled IoU\nprovides an easy-to-optimize training target for localization by disentangling\nthe xy and z space. Extensive experiments validate that Uni3DETR exhibits\nexcellent performance consistently on both indoor and outdoor 3D detection. In\ncontrast to previous specialized detectors, which may perform well on some\nparticular datasets but suffer a substantial degradation on different scenes,\nUni3DETR demonstrates the strong generalization ability under heterogeneous\nconditions (Fig. 1).\n Codes are available at\n\\href{https://github.com/zhenyuw16/Uni3DETR}{https://github.com/zhenyuw16/Uni3DETR}.\n", "date": "2023-10-09", "categories": []} {"id": "2310.05785", "title": "Joint object detection and re-identification for 3D obstacle\n multi-camera systems", "abstract": " In recent years, the field of autonomous driving has witnessed remarkable\nadvancements, driven by the integration of a multitude of sensors, including\ncameras and LiDAR systems, in different prototypes. However, with the\nproliferation of sensor data comes the pressing need for more sophisticated\ninformation processing techniques. This research paper introduces a novel\nmodification to an object detection network that uses camera and lidar\ninformation, incorporating an additional branch designed for the task of\nre-identifying objects across adjacent cameras within the same vehicle while\nelevating the quality of the baseline 3D object detection outcomes. The\nproposed methodology employs a two-step detection pipeline: initially, an\nobject detection network is employed, followed by a 3D box estimator that\noperates on the filtered point cloud generated from the network's detections.\nExtensive experimental evaluations encompassing both 2D and 3D domains validate\nthe effectiveness of the proposed approach and the results underscore the\nsuperiority of this method over traditional Non-Maximum Suppression (NMS)\ntechniques, with an improvement of more than 5\\% in the car category in the\noverlapping areas.\n", "date": "2023-10-09", "categories": []} {"id": "2307.10784", "title": "SMURF: Spatial Multi-Representation Fusion for 3D Object Detection with\n 4D Imaging Radar", "abstract": " The 4D Millimeter wave (mmWave) radar is a promising technology for vehicle\nsensing due to its cost-effectiveness and operability in adverse weather\nconditions. However, the adoption of this technology has been hindered by\nsparsity and noise issues in radar point cloud data. This paper introduces\nspatial multi-representation fusion (SMURF), a novel approach to 3D object\ndetection using a single 4D imaging radar. SMURF leverages multiple\nrepresentations of radar detection points, including pillarization and density\nfeatures of a multi-dimensional Gaussian mixture distribution through kernel\ndensity estimation (KDE). KDE effectively mitigates measurement inaccuracy\ncaused by limited angular resolution and multi-path propagation of radar\nsignals. Additionally, KDE helps alleviate point cloud sparsity by capturing\ndensity features. Experimental evaluations on View-of-Delft (VoD) and\nTJ4DRadSet datasets demonstrate the effectiveness and generalization ability of\nSMURF, outperforming recently proposed 4D imaging radar-based\nsingle-representation models. Moreover, while using 4D imaging radar only,\nSMURF still achieves comparable performance to the state-of-the-art 4D imaging\nradar and camera fusion-based method, with an increase of 1.22% in the mean\naverage precision on bird's-eye view of TJ4DRadSet dataset and 1.32% in the 3D\nmean average precision on the entire annotated area of VoD dataset. Our\nproposed method demonstrates impressive inference time and addresses the\nchallenges of real-time detection, with the inference time no more than 0.05\nseconds for most scans on both datasets. This research highlights the benefits\nof 4D mmWave radar and is a strong benchmark for subsequent works regarding 3D\nobject detection with 4D imaging radar.\n", "date": "2023-07-20", "categories": []} {"id": "2310.00944", "title": "Towards Robust 3D Object Detection In Rainy Conditions", "abstract": " LiDAR sensors are used in autonomous driving applications to accurately\nperceive the environment. However, they are affected by adverse weather\nconditions such as snow, fog, and rain. These everyday phenomena introduce\nunwanted noise into the measurements, severely degrading the performance of\nLiDAR-based perception systems. In this work, we propose a framework for\nimproving the robustness of LiDAR-based 3D object detectors against road spray.\nOur approach uses a state-of-the-art adverse weather detection network to\nfilter out spray from the LiDAR point cloud, which is then used as input for\nthe object detector. In this way, the detected objects are less affected by the\nadverse weather in the scene, resulting in a more accurate perception of the\nenvironment. In addition to adverse weather filtering, we explore the use of\nradar targets to further filter false positive detections. Tests on real-world\ndata show that our approach improves the robustness to road spray of several\npopular 3D object detectors.\n", "date": "2023-10-02", "categories": []} {"id": "2310.02532", "title": "ShaSTA-Fuse: Camera-LiDAR Sensor Fusion to Model Shape and\n Spatio-Temporal Affinities for 3D Multi-Object Tracking", "abstract": " 3D multi-object tracking (MOT) is essential for an autonomous mobile agent to\nsafely navigate a scene. In order to maximize the perception capabilities of\nthe autonomous agent, we aim to develop a 3D MOT framework that fuses camera\nand LiDAR sensor information. Building on our prior LiDAR-only work, ShaSTA,\nwhich models shape and spatio-temporal affinities for 3D MOT, we propose a\nnovel camera-LiDAR fusion approach for learning affinities. At its core, this\nwork proposes a fusion technique that generates a rich sensory signal\nincorporating information about depth and distant objects to enhance affinity\nestimation for improved data association, track lifecycle management,\nfalse-positive elimination, false-negative propagation, and track confidence\nscore refinement. Our main contributions include a novel fusion approach for\ncombining camera and LiDAR sensory signals to learn affinities, and a\nfirst-of-its-kind multimodal sequential track confidence refinement technique\nthat fuses 2D and 3D detections. Additionally, we perform an ablative analysis\non each fusion step to demonstrate the added benefits of incorporating the\ncamera sensor, particular for small, distant objects that tend to suffer from\nthe depth-sensing limits and sparsity of LiDAR sensors. In sum, our technique\nachieves state-of-the-art performance on the nuScenes benchmark amongst\nmultimodal 3D MOT algorithms using CenterPoint detections.\n", "date": "2023-10-04", "categories": []} {"id": "2310.02960", "title": "CoDA: Collaborative Novel Box Discovery and Cross-modal Alignment for\n Open-vocabulary 3D Object Detection", "abstract": " Open-vocabulary 3D Object Detection (OV-3DDet) aims to detect objects from an\narbitrary list of categories within a 3D scene, which remains seldom explored\nin the literature. There are primarily two fundamental problems in OV-3DDet,\ni.e., localizing and classifying novel objects. This paper aims at addressing\nthe two problems simultaneously via a unified framework, under the condition of\nlimited base categories. To localize novel 3D objects, we propose an effective\n3D Novel Object Discovery strategy, which utilizes both the 3D box geometry\npriors and 2D semantic open-vocabulary priors to generate pseudo box labels of\nthe novel objects. To classify novel object boxes, we further develop a\ncross-modal alignment module based on discovered novel boxes, to align feature\nspaces between 3D point cloud and image/text modalities. Specifically, the\nalignment process contains a class-agnostic and a class-discriminative\nalignment, incorporating not only the base objects with annotations but also\nthe increasingly discovered novel objects, resulting in an iteratively enhanced\nalignment. The novel box discovery and crossmodal alignment are jointly learned\nto collaboratively benefit each other. The novel object discovery can directly\nimpact the cross-modal alignment, while a better feature alignment can, in\nturn, boost the localization capability, leading to a unified OV-3DDet\nframework, named CoDA, for simultaneous novel object localization and\nclassification. Extensive experiments on two challenging datasets (i.e.,\nSUN-RGBD and ScanNet) demonstrate the effectiveness of our method and also show\na significant mAP improvement upon the best-performing alternative method by\n80%. Codes and pre-trained models are released on the project page.\n", "date": "2023-10-04", "categories": []} {"id": "2307.00724", "title": "LXL: LiDAR Excluded Lean 3D Object Detection with 4D Imaging Radar and\n Camera Fusion", "abstract": " As an emerging technology and a relatively affordable device, the 4D imaging\nradar has already been confirmed effective in performing 3D object detection in\nautonomous driving. Nevertheless, the sparsity and noisiness of 4D radar point\nclouds hinder further performance improvement, and in-depth studies about its\nfusion with other modalities are lacking. On the other hand, as a new image\nview transformation strategy, \"sampling\" has been applied in a few image-based\ndetectors and shown to outperform the widely applied \"depth-based splatting\"\nproposed in Lift-Splat-Shoot (LSS), even without image depth prediction.\nHowever, the potential of \"sampling\" is not fully unleashed. This paper\ninvestigates the \"sampling\" view transformation strategy on the camera and 4D\nimaging radar fusion-based 3D object detection. LiDAR Excluded Lean (LXL)\nmodel, predicted image depth distribution maps and radar 3D occupancy grids are\ngenerated from image perspective view (PV) features and radar bird's eye view\n(BEV) features, respectively. They are sent to the core of LXL, called \"radar\noccupancy-assisted depth-based sampling\", to aid image view transformation. We\ndemonstrated that more accurate view transformation can be performed by\nintroducing image depths and radar information to enhance the \"sampling\"\nstrategy. Experiments on VoD and TJ4DRadSet datasets show that the proposed\nmethod outperforms the state-of-the-art 3D object detection methods by a\nsignificant margin without bells and whistles. Ablation studies demonstrate\nthat our method performs the best among different enhancement settings.\n", "date": "2023-07-03", "categories": []} {"id": "2309.13549", "title": "Towards Robust Robot 3D Perception in Urban Environments: The UT Campus\n Object Dataset", "abstract": " We introduce the UT Campus Object Dataset (CODa), a mobile robot egocentric\nperception dataset collected on the University of Texas Austin Campus. Our\ndataset contains 8.5 hours of multimodal sensor data: synchronized 3D point\nclouds and stereo RGB video from a 128-channel 3D LiDAR and two 1.25MP RGB\ncameras at 10 fps; RGB-D videos from an additional 0.5MP sensor at 7 fps, and a\n9-DOF IMU sensor at 40 Hz. We provide 58 minutes of ground-truth annotations\ncontaining 1.3 million 3D bounding boxes with instance IDs for 53 semantic\nclasses, 5000 frames of 3D semantic annotations for urban terrain, and\npseudo-ground truth localization. We repeatedly traverse identical geographic\nlocations for a wide range of indoor and outdoor areas, weather conditions, and\ntimes of the day. Using CODa, we empirically demonstrate that: 1) 3D object\ndetection performance in urban settings is significantly higher when trained\nusing CODa compared to existing datasets even when employing state-of-the-art\ndomain adaptation approaches, 2) sensor-specific fine-tuning improves 3D object\ndetection accuracy and 3) pretraining on CODa improves cross-dataset 3D object\ndetection performance in urban settings compared to pretraining on AV datasets.\nUsing our dataset and annotations, we release benchmarks for 3D object\ndetection and 3D semantic segmentation using established metrics. In the\nfuture, the CODa benchmark will include additional tasks like unsupervised\nobject discovery and re-identification. We publicly release CODa on the Texas\nData Repository, pre-trained models, dataset development package, and\ninteractive dataset viewer on our website at https://amrl.cs.utexas.edu/coda.\nWe expect CODa to be a valuable dataset for research in egocentric 3D\nperception and planning for autonomous navigation in urban environments.\n", "date": "2023-09-24", "categories": []} {"id": "2310.00400", "title": "MonoGAE: Roadside Monocular 3D Object Detection with Ground-Aware\n Embeddings", "abstract": " Although the majority of recent autonomous driving systems concentrate on\ndeveloping perception methods based on ego-vehicle sensors, there is an\noverlooked alternative approach that involves leveraging intelligent roadside\ncameras to help extend the ego-vehicle perception ability beyond the visual\nrange. We discover that most existing monocular 3D object detectors rely on the\nego-vehicle prior assumption that the optical axis of the camera is parallel to\nthe ground. However, the roadside camera is installed on a pole with a pitched\nangle, which makes the existing methods not optimal for roadside scenes. In\nthis paper, we introduce a novel framework for Roadside Monocular 3D object\ndetection with ground-aware embeddings, named MonoGAE. Specifically, the ground\nplane is a stable and strong prior knowledge due to the fixed installation of\ncameras in roadside scenarios. In order to reduce the domain gap between the\nground geometry information and high-dimensional image features, we employ a\nsupervised training paradigm with a ground plane to predict high-dimensional\nground-aware embeddings. These embeddings are subsequently integrated with\nimage features through cross-attention mechanisms. Furthermore, to improve the\ndetector's robustness to the divergences in cameras' installation poses, we\nreplace the ground plane depth map with a novel pixel-level refined ground\nplane equation map. Our approach demonstrates a substantial performance\nadvantage over all previous monocular 3D object detectors on widely recognized\n3D detection benchmarks for roadside cameras. The code and pre-trained models\nwill be released soon.\n", "date": "2023-09-30", "categories": []} {"id": "2310.01401", "title": "Pixel-Aligned Recurrent Queries for Multi-View 3D Object Detection", "abstract": " We present PARQ - a multi-view 3D object detector with transformer and\npixel-aligned recurrent queries. Unlike previous works that use learnable\nfeatures or only encode 3D point positions as queries in the decoder, PARQ\nleverages appearance-enhanced queries initialized from reference points in 3D\nspace and updates their 3D location with recurrent cross-attention operations.\nIncorporating pixel-aligned features and cross attention enables the model to\nencode the necessary 3D-to-2D correspondences and capture global contextual\ninformation of the input images. PARQ outperforms prior best methods on the\nScanNet and ARKitScenes datasets, learns and detects faster, is more robust to\ndistribution shifts in reference points, can leverage additional input views\nwithout retraining, and can adapt inference compute by changing the number of\nrecurrent iterations.\n", "date": "2023-10-02", "categories": []} {"id": "2309.16870", "title": "LEF: Late-to-Early Temporal Fusion for LiDAR 3D Object Detection", "abstract": " We propose a late-to-early recurrent feature fusion scheme for 3D object\ndetection using temporal LiDAR point clouds. Our main motivation is fusing\nobject-aware latent embeddings into the early stages of a 3D object detector.\nThis feature fusion strategy enables the model to better capture the shapes and\nposes for challenging objects, compared with learning from raw points directly.\nOur method conducts late-to-early feature fusion in a recurrent manner. This is\nachieved by enforcing window-based attention blocks upon temporally calibrated\nand aligned sparse pillar tokens. Leveraging bird's eye view foreground pillar\nsegmentation, we reduce the number of sparse history features that our model\nneeds to fuse into its current frame by 10$\\times$. We also propose a\nstochastic-length FrameDrop training technique, which generalizes the model to\nvariable frame lengths at inference for improved performance without\nretraining. We evaluate our method on the widely adopted Waymo Open Dataset and\ndemonstrate improvement on 3D object detection against the baseline model,\nespecially for the challenging category of large objects.\n", "date": "2023-09-28", "categories": []} {"id": "2309.16179", "title": "BEVHeight++: Toward Robust Visual Centric 3D Object Detection", "abstract": " While most recent autonomous driving system focuses on developing perception\nmethods on ego-vehicle sensors, people tend to overlook an alternative approach\nto leverage intelligent roadside cameras to extend the perception ability\nbeyond the visual range. We discover that the state-of-the-art vision-centric\nbird's eye view detection methods have inferior performances on roadside\ncameras. This is because these methods mainly focus on recovering the depth\nregarding the camera center, where the depth difference between the car and the\nground quickly shrinks while the distance increases. In this paper, we propose\na simple yet effective approach, dubbed BEVHeight++, to address this issue. In\nessence, we regress the height to the ground to achieve a distance-agnostic\nformulation to ease the optimization process of camera-only perception methods.\nBy incorporating both height and depth encoding techniques, we achieve a more\naccurate and robust projection from 2D to BEV spaces. On popular 3D detection\nbenchmarks of roadside cameras, our method surpasses all previous\nvision-centric methods by a significant margin. In terms of the ego-vehicle\nscenario, our BEVHeight++ possesses superior over depth-only methods.\nSpecifically, it yields a notable improvement of +1.9% NDS and +1.1% mAP over\nBEVDepth when evaluated on the nuScenes validation set. Moreover, on the\nnuScenes test set, our method achieves substantial advancements, with an\nincrease of +2.8% NDS and +1.7% mAP, respectively.\n", "date": "2023-09-28", "categories": []} {"id": "2309.15465", "title": "Cross-Dataset Experimental Study of Radar-Camera Fusion in Bird's-Eye\n View", "abstract": " By exploiting complementary sensor information, radar and camera fusion\nsystems have the potential to provide a highly robust and reliable perception\nsystem for advanced driver assistance systems and automated driving functions.\nRecent advances in camera-based object detection offer new radar-camera fusion\npossibilities with bird's eye view feature maps. In this work, we propose a\nnovel and flexible fusion network and evaluate its performance on two datasets:\nnuScenes and View-of-Delft. Our experiments reveal that while the camera branch\nneeds large and diverse training data, the radar branch benefits more from a\nhigh-performance radar. Using transfer learning, we improve the camera's\nperformance on the smaller dataset. Our results further demonstrate that the\nradar-camera fusion approach significantly outperforms the camera-only and\nradar-only baselines.\n", "date": "2023-09-27", "categories": []} {"id": "2309.15572", "title": "HPL-ViT: A Unified Perception Framework for Heterogeneous Parallel\n LiDARs in V2V", "abstract": " To develop the next generation of intelligent LiDARs, we propose a novel\nframework of parallel LiDARs and construct a hardware prototype in our\nexperimental platform, DAWN (Digital Artificial World for Natural). It\nemphasizes the tight integration of physical and digital space in LiDAR\nsystems, with networking being one of its supported core features. In the\ncontext of autonomous driving, V2V (Vehicle-to-Vehicle) technology enables\nefficient information sharing between different agents which significantly\npromotes the development of LiDAR networks. However, current research operates\nunder an ideal situation where all vehicles are equipped with identical LiDAR,\nignoring the diversity of LiDAR categories and operating frequencies. In this\npaper, we first utilize OpenCDA and RLS (Realistic LiDAR Simulation) to\nconstruct a novel heterogeneous LiDAR dataset named OPV2V-HPL. Additionally, we\npresent HPL-ViT, a pioneering architecture designed for robust feature fusion\nin heterogeneous and dynamic scenarios. It uses a graph-attention Transformer\nto extract domain-specific features for each agent, coupled with a\ncross-attention mechanism for the final fusion. Extensive experiments on\nOPV2V-HPL demonstrate that HPL-ViT achieves SOTA (state-of-the-art) performance\nin all settings and exhibits outstanding generalization capabilities.\n", "date": "2023-09-27", "categories": []} {"id": "2308.09421", "title": "MonoNeRD: NeRF-like Representations for Monocular 3D Object Detection", "abstract": " In the field of monocular 3D detection, it is common practice to utilize\nscene geometric clues to enhance the detector's performance. However, many\nexisting works adopt these clues explicitly such as estimating a depth map and\nback-projecting it into 3D space. This explicit methodology induces sparsity in\n3D representations due to the increased dimensionality from 2D to 3D, and leads\nto substantial information loss, especially for distant and occluded objects.\nTo alleviate this issue, we propose MonoNeRD, a novel detection framework that\ncan infer dense 3D geometry and occupancy. Specifically, we model scenes with\nSigned Distance Functions (SDF), facilitating the production of dense 3D\nrepresentations. We treat these representations as Neural Radiance Fields\n(NeRF) and then employ volume rendering to recover RGB images and depth maps.\nTo the best of our knowledge, this work is the first to introduce volume\nrendering for M3D, and demonstrates the potential of implicit reconstruction\nfor image-based 3D perception. Extensive experiments conducted on the KITTI-3D\nbenchmark and Waymo Open Dataset demonstrate the effectiveness of MonoNeRD.\nCodes are available at https://github.com/cskkxjk/MonoNeRD.\n", "date": "2023-08-18", "categories": []} {"id": "2309.15109", "title": "DistillBEV: Boosting Multi-Camera 3D Object Detection with Cross-Modal\n Knowledge Distillation", "abstract": " 3D perception based on the representations learned from multi-camera\nbird's-eye-view (BEV) is trending as cameras are cost-effective for mass\nproduction in autonomous driving industry. However, there exists a distinct\nperformance gap between multi-camera BEV and LiDAR based 3D object detection.\nOne key reason is that LiDAR captures accurate depth and other geometry\nmeasurements, while it is notoriously challenging to infer such 3D information\nfrom merely image input. In this work, we propose to boost the representation\nlearning of a multi-camera BEV based student detector by training it to imitate\nthe features of a well-trained LiDAR based teacher detector. We propose\neffective balancing strategy to enforce the student to focus on learning the\ncrucial features from the teacher, and generalize knowledge transfer to\nmulti-scale layers with temporal fusion. We conduct extensive evaluations on\nmultiple representative models of multi-camera BEV. Experiments reveal that our\napproach renders significant improvement over the student models, leading to\nthe state-of-the-art performance on the popular benchmark nuScenes.\n", "date": "2023-09-26", "categories": []} {"id": "2309.11804", "title": "FGFusion: Fine-Grained Lidar-Camera Fusion for 3D Object Detection", "abstract": " Lidars and cameras are critical sensors that provide complementary\ninformation for 3D detection in autonomous driving. While most prevalent\nmethods progressively downscale the 3D point clouds and camera images and then\nfuse the high-level features, the downscaled features inevitably lose low-level\ndetailed information. In this paper, we propose Fine-Grained Lidar-Camera\nFusion (FGFusion) that make full use of multi-scale features of image and point\ncloud and fuse them in a fine-grained way. First, we design a dual pathway\nhierarchy structure to extract both high-level semantic and low-level detailed\nfeatures of the image. Second, an auxiliary network is introduced to guide\npoint cloud features to better learn the fine-grained spatial information.\nFinally, we propose multi-scale fusion (MSF) to fuse the last N feature maps of\nimage and point cloud. Extensive experiments on two popular autonomous driving\nbenchmarks, i.e. KITTI and Waymo, demonstrate the effectiveness of our method.\n", "date": "2023-09-21", "categories": []} {"id": "2309.12140", "title": "Unsupervised Domain Adaptation for Self-Driving from Past Traversal\n Features", "abstract": " The rapid development of 3D object detection systems for self-driving cars\nhas significantly improved accuracy. However, these systems struggle to\ngeneralize across diverse driving environments, which can lead to\nsafety-critical failures in detecting traffic participants. To address this, we\npropose a method that utilizes unlabeled repeated traversals of multiple\nlocations to adapt object detectors to new driving environments. By\nincorporating statistics computed from repeated LiDAR scans, we guide the\nadaptation process effectively. Our approach enhances LiDAR-based detection\nmodels using spatial quantized historical features and introduces a lightweight\nregression head to leverage the statistics for feature regularization.\nAdditionally, we leverage the statistics for a novel self-training process to\nstabilize the training. The framework is detector model-agnostic and\nexperiments on real-world datasets demonstrate significant improvements,\nachieving up to a 20-point performance gain, especially in detecting\npedestrians and distant objects. Code is available at\nhttps://github.com/zhangtravis/Hist-DA.\n", "date": "2023-09-21", "categories": []} {"id": "2301.01283", "title": "Cross Modal Transformer: Towards Fast and Robust 3D Object Detection", "abstract": " In this paper, we propose a robust 3D detector, named Cross Modal Transformer\n(CMT), for end-to-end 3D multi-modal detection. Without explicit view\ntransformation, CMT takes the image and point clouds tokens as inputs and\ndirectly outputs accurate 3D bounding boxes. The spatial alignment of\nmulti-modal tokens is performed by encoding the 3D points into multi-modal\nfeatures. The core design of CMT is quite simple while its performance is\nimpressive. It achieves 74.1\\% NDS (state-of-the-art with single model) on\nnuScenes test set while maintaining fast inference speed. Moreover, CMT has a\nstrong robustness even if the LiDAR is missing. Code is released at\nhttps://github.com/junjie18/CMT.\n", "date": "2023-01-03", "categories": []} {"id": "2303.17297", "title": "Understanding the Robustness of 3D Object Detection with Bird's-Eye-View\n Representations in Autonomous Driving", "abstract": " 3D object detection is an essential perception task in autonomous driving to\nunderstand the environments. The Bird's-Eye-View (BEV) representations have\nsignificantly improved the performance of 3D detectors with camera inputs on\npopular benchmarks. However, there still lacks a systematic understanding of\nthe robustness of these vision-dependent BEV models, which is closely related\nto the safety of autonomous driving systems. In this paper, we evaluate the\nnatural and adversarial robustness of various representative models under\nextensive settings, to fully understand their behaviors influenced by explicit\nBEV features compared with those without BEV. In addition to the classic\nsettings, we propose a 3D consistent patch attack by applying adversarial\npatches in the 3D space to guarantee the spatiotemporal consistency, which is\nmore realistic for the scenario of autonomous driving. With substantial\nexperiments, we draw several findings: 1) BEV models tend to be more stable\nthan previous methods under different natural conditions and common corruptions\ndue to the expressive spatial representations; 2) BEV models are more\nvulnerable to adversarial noises, mainly caused by the redundant BEV features;\n3) Camera-LiDAR fusion models have superior performance under different\nsettings with multi-modal inputs, but BEV fusion model is still vulnerable to\nadversarial noises of both point cloud and image. These findings alert the\nsafety issue in the applications of BEV detectors and could facilitate the\ndevelopment of more robust models.\n", "date": "2023-03-30", "categories": []} {"id": "2304.08304", "title": "SDVRF: Sparse-to-Dense Voxel Region Fusion for Multi-modal 3D Object\n Detection", "abstract": " In the perception task of autonomous driving, multi-modal methods have become\na trend due to the complementary characteristics of LiDAR point clouds and\nimage data. However, the performance of multi-modal methods is usually limited\nby the sparsity of the point cloud or the noise problem caused by the\nmisalignment between LiDAR and the camera. To solve these two problems, we\npresent a new concept, Voxel Region (VR), which is obtained by projecting the\nsparse local point clouds in each voxel dynamically. And we propose a novel\nfusion method named Sparse-to-Dense Voxel Region Fusion (SDVRF). Specifically,\nmore pixels of the image feature map inside the VR are gathered to supplement\nthe voxel feature extracted from sparse points and achieve denser fusion.\nMeanwhile, different from prior methods, which project the size-fixed grids,\nour strategy of generating dynamic regions achieves better alignment and avoids\nintroducing too much background noise. Furthermore, we propose a multi-scale\nfusion framework to extract more contextual information and capture the\nfeatures of objects of different sizes. Experiments on the KITTI dataset show\nthat our method improves the performance of different baselines, especially on\nclasses of small size, including Pedestrian and Cyclist.\n", "date": "2023-04-17", "categories": []} {"id": "2309.08932", "title": "Semantics-aware LiDAR-Only Pseudo Point Cloud Generation for 3D Object\n Detection", "abstract": " Although LiDAR sensors are crucial for autonomous systems due to providing\nprecise depth information, they struggle with capturing fine object details,\nespecially at a distance, due to sparse and non-uniform data. Recent advances\nintroduced pseudo-LiDAR, i.e., synthetic dense point clouds, using additional\nmodalities such as cameras to enhance 3D object detection. We present a novel\nLiDAR-only framework that augments raw scans with denser pseudo point clouds by\nsolely relying on LiDAR sensors and scene semantics, omitting the need for\ncameras. Our framework first utilizes a segmentation model to extract scene\nsemantics from raw point clouds, and then employs a multi-modal domain\ntranslator to generate synthetic image segments and depth cues without real\ncameras. This yields a dense pseudo point cloud enriched with semantic\ninformation. We also introduce a new semantically guided projection method,\nwhich enhances detection performance by retaining only relevant pseudo points.\nWe applied our framework to different advanced 3D object detection methods and\nreported up to 2.9% performance upgrade. We also obtained comparable results on\nthe KITTI 3D object detection dataset, in contrast to other state-of-the-art\nLiDAR-only detectors.\n", "date": "2023-09-16", "categories": []} {"id": "2309.09456", "title": "Object2Scene: Putting Objects in Context for Open-Vocabulary 3D\n Detection", "abstract": " Point cloud-based open-vocabulary 3D object detection aims to detect 3D\ncategories that do not have ground-truth annotations in the training set. It is\nextremely challenging because of the limited data and annotations (bounding\nboxes with class labels or text descriptions) of 3D scenes. Previous approaches\nleverage large-scale richly-annotated image datasets as a bridge between 3D and\ncategory semantics but require an extra alignment process between 2D images and\n3D points, limiting the open-vocabulary ability of 3D detectors. Instead of\nleveraging 2D images, we propose Object2Scene, the first approach that\nleverages large-scale large-vocabulary 3D object datasets to augment existing\n3D scene datasets for open-vocabulary 3D object detection. Object2Scene inserts\nobjects from different sources into 3D scenes to enrich the vocabulary of 3D\nscene datasets and generates text descriptions for the newly inserted objects.\nWe further introduce a framework that unifies 3D detection and visual\ngrounding, named L3Det, and propose a cross-domain category-level contrastive\nlearning approach to mitigate the domain gap between 3D objects from different\ndatasets. Extensive experiments on existing open-vocabulary 3D object detection\nbenchmarks show that Object2Scene obtains superior performance over existing\nmethods. We further verify the effectiveness of Object2Scene on a new benchmark\nOV-ScanNet-200, by holding out all rare categories as novel categories not seen\nduring training.\n", "date": "2023-09-18", "categories": []} {"id": "2309.09593", "title": "Mutual Information-calibrated Conformal Feature Fusion for\n Uncertainty-Aware Multimodal 3D Object Detection at the Edge", "abstract": " In the expanding landscape of AI-enabled robotics, robust quantification of\npredictive uncertainties is of great importance. Three-dimensional (3D) object\ndetection, a critical robotics operation, has seen significant advancements;\nhowever, the majority of current works focus only on accuracy and ignore\nuncertainty quantification. Addressing this gap, our novel study integrates the\nprinciples of conformal inference (CI) with information theoretic measures to\nperform lightweight, Monte Carlo-free uncertainty estimation within a\nmultimodal framework. Through a multivariate Gaussian product of the latent\nvariables in a Variational Autoencoder (VAE), features from RGB camera and\nLiDAR sensor data are fused to improve the prediction accuracy. Normalized\nmutual information (NMI) is leveraged as a modulator for calibrating\nuncertainty bounds derived from CI based on a weighted loss function. Our\nsimulation results show an inverse correlation between inherent predictive\nuncertainty and NMI throughout the model's training. The framework demonstrates\ncomparable or better performance in KITTI 3D object detection benchmarks to\nsimilar methods that are not uncertainty-aware, making it suitable for\nreal-time edge robotics.\n", "date": "2023-09-18", "categories": []} {"id": "2309.08369", "title": "An Efficient Wide-Range Pseudo-3D Vehicle Detection Using A Single\n Camera", "abstract": " Wide-range and fine-grained vehicle detection plays a critical role in\nenabling active safety features in intelligent driving systems. However,\nexisting vehicle detection methods based on rectangular bounding boxes (BBox)\noften struggle with perceiving wide-range objects, especially small objects at\nlong distances. And BBox expression cannot provide detailed geometric shape and\npose information of vehicles. This paper proposes a novel wide-range Pseudo-3D\nVehicle Detection method based on images from a single camera and incorporates\nefficient learning methods. This model takes a spliced image as input, which is\nobtained by combining two sub-window images from a high-resolution image. This\nimage format maximizes the utilization of limited image resolution to retain\nessential information about wide-range vehicle objects. To detect pseudo-3D\nobjects, our model adopts specifically designed detection heads. These heads\nsimultaneously output extended BBox and Side Projection Line (SPL)\nrepresentations, which capture vehicle shapes and poses, enabling\nhigh-precision detection. To further enhance the performance of detection, a\njoint constraint loss combining both the object box and SPL is designed during\nmodel training, improving the efficiency, stability, and prediction accuracy of\nthe model. Experimental results on our self-built dataset demonstrate that our\nmodel achieves favorable performance in wide-range pseudo-3D vehicle detection\nacross multiple evaluation metrics. Our demo video has been placed at\nhttps://www.youtube.com/watch?v=1gk1PmsQ5Q8.\n", "date": "2023-09-15", "categories": []} {"id": "2211.07108", "title": "Recursive Cross-View: Use Only 2D Detectors to Achieve 3D Object\n Detection without 3D Annotations", "abstract": " Heavily relying on 3D annotations limits the real-world application of 3D\nobject detection. In this paper, we propose a method that does not demand any\n3D annotation, while being able to predict fully oriented 3D bounding boxes.\nOur method, called Recursive Cross-View (RCV), utilizes the three-view\nprinciple to convert 3D detection into multiple 2D detection tasks, requiring\nonly a subset of 2D labels. We propose a recursive paradigm, in which instance\nsegmentation and 3D bounding box generation by Cross-View are implemented\nrecursively until convergence. Specifically, our proposed method involves the\nuse of a frustum for each 2D bounding box, which is then followed by the\nrecursive paradigm that ultimately generates a fully oriented 3D box, along\nwith its corresponding class and score. Note that, class and score are given by\nthe 2D detector. Estimated on the SUN RGB-D and KITTI datasets, our method\noutperforms existing image-based approaches. To justify that our method can be\nquickly used to new tasks, we implement it on two real-world scenarios, namely\n3D human detection and 3D hand detection. As a result, two new 3D annotated\ndatasets are obtained, which means that RCV can be viewed as a (semi-)\nautomatic 3D annotator. Furthermore, we deploy RCV on a depth sensor, which\nachieves detection at 7 fps on a live RGB-D stream. RCV is the first 3D\ndetection method that yields fully oriented 3D boxes without consuming 3D\nlabels.\n", "date": "2022-11-14", "categories": []} {"id": "2211.09302", "title": "You Only Label Once: 3D Box Adaptation from Point Cloud to Image via\n Semi-Supervised Learning", "abstract": " The image-based 3D object detection task expects that the predicted 3D\nbounding box has a ``tightness'' projection (also referred to as cuboid), which\nfits the object contour well on the image while still keeping the geometric\nattribute on the 3D space, e.g., physical dimension, pairwise orthogonal, etc.\nThese requirements bring significant challenges to the annotation. Simply\nprojecting the Lidar-labeled 3D boxes to the image leads to non-trivial\nmisalignment, while directly drawing a cuboid on the image cannot access the\noriginal 3D information. In this work, we propose a learning-based 3D box\nadaptation approach that automatically adjusts minimum parameters of the\n360$^{\\circ}$ Lidar 3D bounding box to perfectly fit the image appearance of\npanoramic cameras. With only a few 2D boxes annotation as guidance during the\ntraining phase, our network can produce accurate image-level cuboid annotations\nwith 3D properties from Lidar boxes. We call our method ``you only label\nonce'', which means labeling on the point cloud once and automatically adapting\nto all surrounding cameras. As far as we know, we are the first to focus on\nimage-level cuboid refinement, which balances the accuracy and efficiency well\nand dramatically reduces the labeling effort for accurate cuboid annotation.\nExtensive experiments on the public Waymo and NuScenes datasets show that our\nmethod can produce human-level cuboid annotation on the image without needing\nmanual adjustment.\n", "date": "2022-11-17", "categories": []} {"id": "2308.10161", "title": "ThermRad: A Multi-modal Dataset for Robust 3D Object Detection under\n Challenging Conditions", "abstract": " Robust 3D object detection in extreme weather and illumination conditions is\na challenging task. While radars and thermal cameras are known for their\nresilience to these conditions, few studies have been conducted on\nradar-thermal fusion due to the lack of corresponding datasets. To address this\ngap, we first present a new multi-modal dataset called ThermRad, which includes\na 3D LiDAR, a 4D radar, an RGB camera and a thermal camera. This dataset is\nunique because it includes data from all four sensors in extreme weather\nconditions, providing a valuable resource for future research in this area. To\nvalidate the robustness of 4D radars and thermal cameras for 3D object\ndetection in challenging weather conditions, we propose a new multi-modal\nfusion method called RTDF-RCNN, which leverages the complementary strengths of\n4D radars and thermal cameras to boost object detection performance. To further\nprove the effectiveness of our proposed framework, we re-implement\nstate-of-the-art (SOTA) 3D detectors on our dataset as benchmarks for\nevaluation. Our method achieves significant enhancements in detecting cars,\npedestrians, and cyclists, with improvements of over 7.98%, 24.27%, and 27.15%,\nrespectively, while achieving comparable results to LiDAR-based approaches. Our\ncontributions in both the ThermRad dataset and the new multi-modal fusion\nmethod provide a new approach to robust 3D object detection in adverse weather\nand illumination conditions. The ThermRad dataset will be released.\n", "date": "2023-08-20", "categories": []} {"id": "2309.05810", "title": "SHIFT3D: Synthesizing Hard Inputs For Tricking 3D Detectors", "abstract": " We present SHIFT3D, a differentiable pipeline for generating 3D shapes that\nare structurally plausible yet challenging to 3D object detectors. In\nsafety-critical applications like autonomous driving, discovering such novel\nchallenging objects can offer insight into unknown vulnerabilities of 3D\ndetectors. By representing objects with a signed distanced function (SDF), we\nshow that gradient error signals allow us to smoothly deform the shape or pose\nof a 3D object in order to confuse a downstream 3D detector. Importantly, the\nobjects generated by SHIFT3D physically differ from the baseline object yet\nretain a semantically recognizable shape. Our approach provides interpretable\nfailure modes for modern 3D object detectors, and can aid in preemptive\ndiscovery of potential safety risks within 3D perception systems before these\nrisks become critical failures.\n", "date": "2023-09-11", "categories": []} {"id": "2309.06199", "title": "SCP: Scene Completion Pre-training for 3D Object Detection", "abstract": " 3D object detection using LiDAR point clouds is a fundamental task in the\nfields of computer vision, robotics, and autonomous driving. However, existing\n3D detectors heavily rely on annotated datasets, which are both time-consuming\nand prone to errors during the process of labeling 3D bounding boxes. In this\npaper, we propose a Scene Completion Pre-training (SCP) method to enhance the\nperformance of 3D object detectors with less labeled data. SCP offers three key\nadvantages: (1) Improved initialization of the point cloud model. By completing\nthe scene point clouds, SCP effectively captures the spatial and semantic\nrelationships among objects within urban environments. (2) Elimination of the\nneed for additional datasets. SCP serves as a valuable auxiliary network that\ndoes not impose any additional efforts or data requirements on the 3D\ndetectors. (3) Reduction of the amount of labeled data for detection. With the\nhelp of SCP, the existing state-of-the-art 3D detectors can achieve comparable\nperformance while only relying on 20% labeled data.\n", "date": "2023-09-12", "categories": []} {"id": "2302.09221", "title": "Moby: Empowering 2D Models for Efficient Point Cloud Analytics on the\n Edge", "abstract": " 3D object detection plays a pivotal role in many applications, most notably\nautonomous driving and robotics. These applications are commonly deployed on\nedge devices to promptly interact with the environment, and often require near\nreal-time response. With limited computation power, it is challenging to\nexecute 3D detection on the edge using highly complex neural networks. Common\napproaches such as offloading to the cloud induce significant latency overheads\ndue to the large amount of point cloud data during transmission. To resolve the\ntension between wimpy edge devices and compute-intensive inference workloads,\nwe explore the possibility of empowering fast 2D detection to extrapolate 3D\nbounding boxes. To this end, we present Moby, a novel system that demonstrates\nthe feasibility and potential of our approach. We design a transformation\npipeline for Moby that generates 3D bounding boxes efficiently and accurately\nbased on 2D detection results without running 3D detectors. Further, we devise\na frame offloading scheduler that decides when to launch the 3D detector\njudiciously in the cloud to avoid the errors from accumulating. Extensive\nevaluations on NVIDIA Jetson TX2 with real-world autonomous driving datasets\ndemonstrate that Moby offers up to 91.9% latency improvement with modest\naccuracy loss over state of the art.\n", "date": "2023-02-18", "categories": []} {"id": "2303.17597", "title": "Robo3D: Towards Robust and Reliable 3D Perception against Corruptions", "abstract": " The robustness of 3D perception systems under natural corruptions from\nenvironments and sensors is pivotal for safety-critical applications. Existing\nlarge-scale 3D perception datasets often contain data that are meticulously\ncleaned. Such configurations, however, cannot reflect the reliability of\nperception models during the deployment stage. In this work, we present Robo3D,\nthe first comprehensive benchmark heading toward probing the robustness of 3D\ndetectors and segmentors under out-of-distribution scenarios against natural\ncorruptions that occur in real-world environments. Specifically, we consider\neight corruption types stemming from severe weather conditions, external\ndisturbances, and internal sensor failure. We uncover that, although promising\nresults have been progressively achieved on standard benchmarks,\nstate-of-the-art 3D perception models are at risk of being vulnerable to\ncorruptions. We draw key observations on the use of data representations,\naugmentation schemes, and training strategies, that could severely affect the\nmodel's performance. To pursue better robustness, we propose a\ndensity-insensitive training framework along with a simple flexible\nvoxelization strategy to enhance the model resiliency. We hope our benchmark\nand approach could inspire future research in designing more robust and\nreliable 3D perception models. Our robustness benchmark suite is publicly\navailable.\n", "date": "2023-03-30", "categories": []} {"id": "2305.15836", "title": "Improved Multi-Scale Grid Rendering of Point Clouds for Radar Object\n Detection Networks", "abstract": " Architectures that first convert point clouds to a grid representation and\nthen apply convolutional neural networks achieve good performance for\nradar-based object detection. However, the transfer from irregular point cloud\ndata to a dense grid structure is often associated with a loss of information,\ndue to the discretization and aggregation of points. In this paper, we propose\na novel architecture, multi-scale KPPillarsBEV, that aims to mitigate the\nnegative effects of grid rendering. Specifically, we propose a novel grid\nrendering method, KPBEV, which leverages the descriptive power of kernel point\nconvolutions to improve the encoding of local point cloud contexts during grid\nrendering. In addition, we propose a general multi-scale grid rendering\nformulation to incorporate multi-scale feature maps into convolutional\nbackbones of detection networks with arbitrary grid rendering methods. We\nperform extensive experiments on the nuScenes dataset and evaluate the methods\nin terms of detection performance and computational complexity. The proposed\nmulti-scale KPPillarsBEV architecture outperforms the baseline by 5.37% and the\nprevious state of the art by 2.88% in Car AP4.0 (average precision for a\nmatching threshold of 4 meters) on the nuScenes validation set. Moreover, the\nproposed single-scale KPBEV grid rendering improves the Car AP4.0 by 2.90% over\nthe baseline while maintaining the same inference speed.\n", "date": "2023-05-25", "categories": []} {"id": "2306.08528", "title": "Predict to Detect: Prediction-guided 3D Object Detection using\n Sequential Images", "abstract": " Recent camera-based 3D object detection methods have introduced sequential\nframes to improve the detection performance hoping that multiple frames would\nmitigate the large depth estimation error. Despite improved detection\nperformance, prior works rely on naive fusion methods (e.g., concatenation) or\nare limited to static scenes (e.g., temporal stereo), neglecting the importance\nof the motion cue of objects. These approaches do not fully exploit the\npotential of sequential images and show limited performance improvements. To\naddress this limitation, we propose a novel 3D object detection model, P2D\n(Predict to Detect), that integrates a prediction scheme into a detection\nframework to explicitly extract and leverage motion features. P2D predicts\nobject information in the current frame using solely past frames to learn\ntemporal motion features. We then introduce a novel temporal feature\naggregation method that attentively exploits Bird's-Eye-View (BEV) features\nbased on predicted object information, resulting in accurate 3D object\ndetection. Experimental results demonstrate that P2D improves mAP and NDS by\n3.0% and 3.7% compared to the sequential image-based baseline, illustrating\nthat incorporating a prediction scheme can significantly improve detection\naccuracy.\n", "date": "2023-06-14", "categories": []} {"id": "2308.09244", "title": "SparseBEV: High-Performance Sparse 3D Object Detection from Multi-Camera\n Videos", "abstract": " Camera-based 3D object detection in BEV (Bird's Eye View) space has drawn\ngreat attention over the past few years. Dense detectors typically follow a\ntwo-stage pipeline by first constructing a dense BEV feature and then\nperforming object detection in BEV space, which suffers from complex view\ntransformations and high computation cost. On the other side, sparse detectors\nfollow a query-based paradigm without explicit dense BEV feature construction,\nbut achieve worse performance than the dense counterparts. In this paper, we\nfind that the key to mitigate this performance gap is the adaptability of the\ndetector in both BEV and image space. To achieve this goal, we propose\nSparseBEV, a fully sparse 3D object detector that outperforms the dense\ncounterparts. SparseBEV contains three key designs, which are (1)\nscale-adaptive self attention to aggregate features with adaptive receptive\nfield in BEV space, (2) adaptive spatio-temporal sampling to generate sampling\nlocations under the guidance of queries, and (3) adaptive mixing to decode the\nsampled features with dynamic weights from the queries. On the test split of\nnuScenes, SparseBEV achieves the state-of-the-art performance of 67.5 NDS. On\nthe val split, SparseBEV achieves 55.8 NDS while maintaining a real-time\ninference speed of 23.5 FPS. Code is available at\nhttps://github.com/MCG-NJU/SparseBEV.\n", "date": "2023-08-18", "categories": []} {"id": "2309.01106", "title": "AdvMono3D: Advanced Monocular 3D Object Detection with Depth-Aware\n Robust Adversarial Training", "abstract": " Monocular 3D object detection plays a pivotal role in the field of autonomous\ndriving and numerous deep learning-based methods have made significant\nbreakthroughs in this area. Despite the advancements in detection accuracy and\nefficiency, these models tend to fail when faced with such attacks, rendering\nthem ineffective. Therefore, bolstering the adversarial robustness of 3D\ndetection models has become a crucial issue that demands immediate attention\nand innovative solutions. To mitigate this issue, we propose a depth-aware\nrobust adversarial training method for monocular 3D object detection, dubbed\nDART3D. Specifically, we first design an adversarial attack that iteratively\ndegrades the 2D and 3D perception capabilities of 3D object detection\nmodels(IDP), serves as the foundation for our subsequent defense mechanism. In\nresponse to this attack, we propose an uncertainty-based residual learning\nmethod for adversarial training. Our adversarial training approach capitalizes\non the inherent uncertainty, enabling the model to significantly improve its\nrobustness against adversarial attacks. We conducted extensive experiments on\nthe KITTI 3D datasets, demonstrating that DART3D surpasses direct adversarial\ntraining (the most popular approach) under attacks in 3D object detection\n$AP_{R40}$ of car category for the Easy, Moderate, and Hard settings, with\nimprovements of 4.415%, 4.112%, and 3.195%, respectively.\n", "date": "2023-09-03", "categories": []} {"id": "2309.02049", "title": "Diffusion-based 3D Object Detection with Random Boxes", "abstract": " 3D object detection is an essential task for achieving autonomous driving.\nExisting anchor-based detection methods rely on empirical heuristics setting of\nanchors, which makes the algorithms lack elegance. In recent years, we have\nwitnessed the rise of several generative models, among which diffusion models\nshow great potential for learning the transformation of two distributions. Our\nproposed Diff3Det migrates the diffusion model to proposal generation for 3D\nobject detection by considering the detection boxes as generative targets.\nDuring training, the object boxes diffuse from the ground truth boxes to the\nGaussian distribution, and the decoder learns to reverse this noise process. In\nthe inference stage, the model progressively refines a set of random boxes to\nthe prediction results. We provide detailed experiments on the KITTI benchmark\nand achieve promising performance compared to classical anchor-based 3D\ndetection methods.\n", "date": "2023-09-05", "categories": []} {"id": "2305.07270", "title": "SSD-MonoDETR: Supervised Scale-aware Deformable Transformer for\n Monocular 3D Object Detection", "abstract": " Transformer-based methods have demonstrated superior performance for\nmonocular 3D object detection recently, which aims at predicting 3D attributes\nfrom a single 2D image. Most existing transformer-based methods leverage both\nvisual and depth representations to explore valuable query points on objects,\nand the quality of the learned query points has a great impact on detection\naccuracy. Unfortunately, existing unsupervised attention mechanisms in\ntransformers are prone to generate low-quality query features due to inaccurate\nreceptive fields, especially on hard objects. To tackle this problem, this\npaper proposes a novel \"Supervised Scale-aware Deformable Attention\" (SSDA) for\nmonocular 3D object detection. Specifically, SSDA presets several masks with\ndifferent scales and utilizes depth and visual features to adaptively learn a\nscale-aware filter for object query augmentation. Imposing the scale awareness,\nSSDA could well predict the accurate receptive field of an object query to\nsupport robust query feature generation. Aside from this, SSDA is assigned with\na Weighted Scale Matching (WSM) loss to supervise scale prediction, which\npresents more confident results as compared to the unsupervised attention\nmechanisms. Extensive experiments on the KITTI and Waymo Open datasets\ndemonstrate that SSDA significantly improves the detection accuracy, especially\non moderate and hard objects, yielding state-of-the-art performance as compared\nto the existing approaches. Our code will be made publicly available at\nhttps://github.com/mikasa3lili/SSD-MonoDETR.\n", "date": "2023-05-12", "categories": []} {"id": "2303.12743", "title": "DR.CPO: Diversified and Realistic 3D Augmentation via Iterative\n Construction, Random Placement, and HPR Occlusion", "abstract": " In autonomous driving, data augmentation is commonly used for improving 3D\nobject detection. The most basic methods include insertion of copied objects\nand rotation and scaling of the entire training frame. Numerous variants have\nbeen developed as well. The existing methods, however, are considerably limited\nwhen compared to the variety of the real world possibilities. In this work, we\ndevelop a diversified and realistic augmentation method that can flexibly\nconstruct a whole-body object, freely locate and rotate the object, and apply\nself-occlusion and external-occlusion accordingly. To improve the diversity of\nthe whole-body object construction, we develop an iterative method that\nstochastically combines multiple objects observed from the real world into a\nsingle object. Unlike the existing augmentation methods, the constructed\nobjects can be randomly located and rotated in the training frame because\nproper occlusions can be reflected to the whole-body objects in the final step.\nFinally, proper self-occlusion at each local object level and\nexternal-occlusion at the global frame level are applied using the Hidden Point\nRemoval (HPR) algorithm that is computationally efficient. HPR is also used for\nadaptively controlling the point density of each object according to the\nobject's distance from the LiDAR. Experiment results show that the proposed\nDR.CPO algorithm is data-efficient and model-agnostic without incurring any\ncomputational overhead. Also, DR.CPO can improve mAP performance by 2.08% when\ncompared to the best 3D detection result known for KITTI dataset. The code is\navailable at https://github.com/SNU-DRL/DRCPO.git\n", "date": "2023-03-20", "categories": []} {"id": "2303.17895", "title": "EA-LSS: Edge-aware Lift-splat-shot Framework for 3D BEV Object Detection", "abstract": " In recent years, great progress has been made in the Lift-Splat-Shot-based\n(LSS-based) 3D object detection method. However, inaccurate depth estimation\nremains an important constraint to the accuracy of camera-only and multi-model\n3D object detection models, especially in regions where the depth changes\nsignificantly (i.e., the \"depth jump\" problem). In this paper, we proposed a\nnovel Edge-aware Lift-splat-shot (EA-LSS) framework. Specifically, edge-aware\ndepth fusion (EADF) module is proposed to alleviate the \"depth jump\" problem\nand fine-grained depth (FGD) module to further enforce refined supervision on\ndepth. Our EA-LSS framework is compatible for any LSS-based 3D object detection\nmodels, and effectively boosts their performances with negligible increment of\ninference time. Experiments on nuScenes benchmarks demonstrate that EA-LSS is\neffective in either camera-only or multi-model models. It is worth mentioning\nthat EA-LSS achieved the state-of-the-art performance on nuScenes test\nbenchmarks with mAP and NDS of 76.5% and 77.6%, respectively.\n", "date": "2023-03-31", "categories": []} {"id": "2308.15357", "title": "Ego-Motion Estimation and Dynamic Motion Separation from 3D Point Clouds\n for Accumulating Data and Improving 3D Object Detection", "abstract": " New 3+1D high-resolution radar sensors are gaining importance for 3D object\ndetection in the automotive domain due to their relative affordability and\nimproved detection compared to classic low-resolution radar sensors. One\nlimitation of high-resolution radar sensors, compared to lidar sensors, is the\nsparsity of the generated point cloud. This sparsity could be partially\novercome by accumulating radar point clouds of subsequent time steps. This\ncontribution analyzes limitations of accumulating radar point clouds on the\nView-of-Delft dataset. By employing different ego-motion estimation approaches,\nthe dataset's inherent constraints, and possible solutions are analyzed.\nAdditionally, a learning-based instance motion estimation approach is deployed\nto investigate the influence of dynamic motion on the accumulated point cloud\nfor object detection. Experiments document an improved object detection\nperformance by applying an ego-motion estimation and dynamic motion correction\napproach.\n", "date": "2023-08-29", "categories": []} {"id": "2303.11325", "title": "GeoMIM: Towards Better 3D Knowledge Transfer via Masked Image Modeling\n for Multi-view 3D Understanding", "abstract": " Multi-view camera-based 3D detection is a challenging problem in computer\nvision. Recent works leverage a pretrained LiDAR detection model to transfer\nknowledge to a camera-based student network. However, we argue that there is a\nmajor domain gap between the LiDAR BEV features and the camera-based BEV\nfeatures, as they have different characteristics and are derived from different\nsources. In this paper, we propose Geometry Enhanced Masked Image Modeling\n(GeoMIM) to transfer the knowledge of the LiDAR model in a pretrain-finetune\nparadigm for improving the multi-view camera-based 3D detection. GeoMIM is a\nmulti-camera vision transformer with Cross-View Attention (CVA) blocks that\nuses LiDAR BEV features encoded by the pretrained BEV model as learning\ntargets. During pretraining, GeoMIM's decoder has a semantic branch completing\ndense perspective-view features and the other geometry branch reconstructing\ndense perspective-view depth maps. The depth branch is designed to be\ncamera-aware by inputting the camera's parameters for better transfer\ncapability. Extensive results demonstrate that GeoMIM outperforms existing\nmethods on nuScenes benchmark, achieving state-of-the-art performance for\ncamera-based 3D object detection and 3D segmentation. Code and pretrained\nmodels are available at https://github.com/Sense-X/GeoMIM.\n", "date": "2023-03-20", "categories": []} {"id": "2308.14481", "title": "Group Regression for Query Based Object Detection and Tracking", "abstract": " Group regression is commonly used in 3D object detection to predict box\nparameters of similar classes in a joint head, aiming to benefit from\nsimilarities while separating highly dissimilar classes. For query-based\nperception methods, this has, so far, not been feasible. We close this gap and\npresent a method to incorporate multi-class group regression, especially\ndesigned for the 3D domain in the context of autonomous driving, into existing\nattention and query-based perception approaches. We enhance a transformer based\njoint object detection and tracking model with this approach, and thoroughly\nevaluate its behavior and performance. For group regression, the classes of the\nnuScenes dataset are divided into six groups of similar shape and prevalence,\neach being regressed by a dedicated head. We show that the proposed method is\napplicable to many existing transformer based perception approaches and can\nbring potential benefits. The behavior of query group regression is thoroughly\nanalyzed in comparison to a unified regression head, e.g. in terms of\nclass-switching behavior and distribution of the output parameters. The\nproposed method offers many possibilities for further research, such as in the\ndirection of deep multi-hypotheses tracking.\n", "date": "2023-08-28", "categories": []} {"id": "2307.11482", "title": "R2Det: Redemption from Range-view for Accurate 3D Object Detection", "abstract": " LiDAR-based 3D object detection is of paramount importance for autonomous\ndriving. Recent trends show a remarkable improvement for bird's-eye-view (BEV)\nbased and point-based methods as they demonstrate superior performance compared\nto range-view counterparts. This paper presents an insight that leverages\nrange-view representation to enhance 3D points for accurate 3D object\ndetection. Specifically, we introduce a Redemption from Range-view Module\n(R2M), a plug-and-play approach for 3D surface texture enhancement from the 2D\nrange view to the 3D point view. R2M comprises BasicBlock for 2D feature\nextraction, Hierarchical-dilated (HD) Meta Kernel for expanding the 3D\nreceptive field, and Feature Points Redemption (FPR) for recovering 3D surface\ntexture information. R2M can be seamlessly integrated into state-of-the-art\nLiDAR-based 3D object detectors as preprocessing and achieve appealing\nimprovement, e.g., 1.39%, 1.67%, and 1.97% mAP improvement on easy, moderate,\nand hard difficulty level of KITTI val set, respectively. Based on R2M, we\nfurther propose R2Detector (R2Det) with the Synchronous-Grid RoI Pooling for\naccurate box refinement. R2Det outperforms existing range-view-based methods by\na significant margin on both the KITTI benchmark and the Waymo Open Dataset.\nCodes will be made publicly available.\n", "date": "2023-07-21", "categories": []} {"id": "2308.12512", "title": "I3DOD: Towards Incremental 3D Object Detection via Prompting", "abstract": " 3D object detection has achieved significant performance in many fields,\ne.g., robotics system, autonomous driving, and augmented reality. However, most\nexisting methods could cause catastrophic forgetting of old classes when\nperforming on the class-incremental scenarios. Meanwhile, the current\nclass-incremental 3D object detection methods neglect the relationships between\nthe object localization information and category semantic information and\nassume all the knowledge of old model is reliable. To address the above\nchallenge, we present a novel Incremental 3D Object Detection framework with\nthe guidance of prompting, i.e., I3DOD. Specifically, we propose a task-shared\nprompts mechanism to learn the matching relationships between the object\nlocalization information and category semantic information. After training on\nthe current task, these prompts will be stored in our prompt pool, and perform\nthe relationship of old classes in the next task. Moreover, we design a\nreliable distillation strategy to transfer knowledge from two aspects: a\nreliable dynamic distillation is developed to filter out the negative knowledge\nand transfer the reliable 3D knowledge to new detection model; the relation\nfeature is proposed to capture the responses relation in feature space and\nprotect plasticity of the model when learning novel 3D classes. To the end, we\nconduct comprehensive experiments on two benchmark datasets and our method\noutperforms the state-of-the-art object detection methods by 0.6% - 2.7% in\nterms of mAP@0.25.\n", "date": "2023-08-24", "categories": []} {"id": "2308.12779", "title": "On Offline Evaluation of 3D Object Detection for Autonomous Driving", "abstract": " Prior work in 3D object detection evaluates models using offline metrics like\naverage precision since closed-loop online evaluation on the downstream driving\ntask is costly. However, it is unclear how indicative offline results are of\ndriving performance. In this work, we perform the first empirical evaluation\nmeasuring how predictive different detection metrics are of driving performance\nwhen detectors are integrated into a full self-driving stack. We conduct\nextensive experiments on urban driving in the CARLA simulator using 16 object\ndetection models. We find that the nuScenes Detection Score has a higher\ncorrelation to driving performance than the widely used average precision\nmetric. In addition, our results call for caution on the exclusive reliance on\nthe emerging class of `planner-centric' metrics.\n", "date": "2023-08-24", "categories": []} {"id": "2308.12938", "title": "Perspective-aware Convolution for Monocular 3D Object Detection", "abstract": " Monocular 3D object detection is a crucial and challenging task for\nautonomous driving vehicle, while it uses only a single camera image to infer\n3D objects in the scene. To address the difficulty of predicting depth using\nonly pictorial clue, we propose a novel perspective-aware convolutional layer\nthat captures long-range dependencies in images. By enforcing convolutional\nkernels to extract features along the depth axis of every image pixel, we\nincorporates perspective information into network architecture. We integrate\nour perspective-aware convolutional layer into a 3D object detector and\ndemonstrate improved performance on the KITTI3D dataset, achieving a 23.9\\%\naverage precision in the easy benchmark. These results underscore the\nimportance of modeling scene clues for accurate depth inference and highlight\nthe benefits of incorporating scene structure in network design. Our\nperspective-aware convolutional layer has the potential to enhance object\ndetection accuracy by providing more precise and context-aware feature\nextraction.\n", "date": "2023-08-24", "categories": []} {"id": "2308.11327", "title": "Object Detection Difficulty: Suppressing Over-aggregation for Faster and\n Better Video Object Detection", "abstract": " Current video object detection (VOD) models often encounter issues with\nover-aggregation due to redundant aggregation strategies, which perform feature\naggregation on every frame. This results in suboptimal performance and\nincreased computational complexity. In this work, we propose an image-level\nObject Detection Difficulty (ODD) metric to quantify the difficulty of\ndetecting objects in a given image. The derived ODD scores can be used in the\nVOD process to mitigate over-aggregation. Specifically, we train an ODD\npredictor as an auxiliary head of a still-image object detector to compute the\nODD score for each image based on the discrepancies between detection results\nand ground-truth bounding boxes. The ODD score enhances the VOD system in two\nways: 1) it enables the VOD system to select superior global reference frames,\nthereby improving overall accuracy; and 2) it serves as an indicator in the\nnewly designed ODD Scheduler to eliminate the aggregation of frames that are\neasy to detect, thus accelerating the VOD process. Comprehensive experiments\ndemonstrate that, when utilized for selecting global reference frames, ODD-VOD\nconsistently enhances the accuracy of Global-frame-based VOD models. When\nemployed for acceleration, ODD-VOD consistently improves the frames per second\n(FPS) by an average of 73.3% across 8 different VOD models without sacrificing\naccuracy. When combined, ODD-VOD attains state-of-the-art performance when\ncompeting with many VOD methods in both accuracy and speed. Our work represents\na significant advancement towards making VOD more practical for real-world\napplications.\n", "date": "2023-08-22", "categories": []} {"id": "2308.10308", "title": "Representation Disparity-aware Distillation for 3D Object Detection", "abstract": " In this paper, we focus on developing knowledge distillation (KD) for compact\n3D detectors. We observe that off-the-shelf KD methods manifest their efficacy\nonly when the teacher model and student counterpart share similar intermediate\nfeature representations. This might explain why they are less effective in\nbuilding extreme-compact 3D detectors where significant representation\ndisparity arises due primarily to the intrinsic sparsity and irregularity in 3D\npoint clouds. This paper presents a novel representation disparity-aware\ndistillation (RDD) method to address the representation disparity issue and\nreduce performance gap between compact students and over-parameterized\nteachers. This is accomplished by building our RDD from an innovative\nperspective of information bottleneck (IB), which can effectively minimize the\ndisparity of proposal region pairs from student and teacher in features and\nlogits. Extensive experiments are performed to demonstrate the superiority of\nour RDD over existing KD methods. For example, our RDD increases mAP of\nCP-Voxel-S to 57.1% on nuScenes dataset, which even surpasses teacher\nperformance while taking up only 42% FLOPs.\n", "date": "2023-08-20", "categories": []} {"id": "2308.10515", "title": "QD-BEV : Quantization-aware View-guided Distillation for Multi-view 3D\n Object Detection", "abstract": " Multi-view 3D detection based on BEV (bird-eye-view) has recently achieved\nsignificant improvements. However, the huge memory consumption of\nstate-of-the-art models makes it hard to deploy them on vehicles, and the\nnon-trivial latency will affect the real-time perception of streaming\napplications. Despite the wide application of quantization to lighten models,\nwe show in our paper that directly applying quantization in BEV tasks will 1)\nmake the training unstable, and 2) lead to intolerable performance degradation.\nTo solve these issues, our method QD-BEV enables a novel view-guided\ndistillation (VGD) objective, which can stabilize the quantization-aware\ntraining (QAT) while enhancing the model performance by leveraging both image\nfeatures and BEV features. Our experiments show that QD-BEV achieves similar or\neven better accuracy than previous methods with significant efficiency gains.\nOn the nuScenes datasets, the 4-bit weight and 6-bit activation quantized\nQD-BEV-Tiny model achieves 37.2% NDS with only 15.8 MB model size,\noutperforming BevFormer-Tiny by 1.8% with an 8x model compression. On the Small\nand Base variants, QD-BEV models also perform superbly and achieve 47.9% NDS\n(28.2 MB) and 50.9% NDS (32.9 MB), respectively.\n", "date": "2023-08-21", "categories": []} {"id": "2212.04675", "title": "SemanticBEVFusion: Rethink LiDAR-Camera Fusion in Unified Bird's-Eye\n View Representation for 3D Object Detection", "abstract": " LiDAR and camera are two essential sensors for 3D object detection in\nautonomous driving. LiDAR provides accurate and reliable 3D geometry\ninformation while the camera provides rich texture with color. Despite the\nincreasing popularity of fusing these two complementary sensors, the challenge\nremains in how to effectively fuse 3D LiDAR point cloud with 2D camera images.\nRecent methods focus on point-level fusion which paints the LiDAR point cloud\nwith camera features in the perspective view or bird's-eye view (BEV)-level\nfusion which unifies multi-modality features in the BEV representation. In this\npaper, we rethink these previous fusion strategies and analyze their\ninformation loss and influences on geometric and semantic features. We present\nSemanticBEVFusion to deeply fuse camera features with LiDAR features in a\nunified BEV representation while maintaining per-modality strengths for 3D\nobject detection. Our method achieves state-of-the-art performance on the\nlarge-scale nuScenes dataset, especially for challenging distant objects. The\ncode will be made publicly available.\n", "date": "2022-12-09", "categories": []} {"id": "2306.06023", "title": "DetZero: Rethinking Offboard 3D Object Detection with Long-term\n Sequential Point Clouds", "abstract": " Existing offboard 3D detectors always follow a modular pipeline design to\ntake advantage of unlimited sequential point clouds. We have found that the\nfull potential of offboard 3D detectors is not explored mainly due to two\nreasons: (1) the onboard multi-object tracker cannot generate sufficient\ncomplete object trajectories, and (2) the motion state of objects poses an\ninevitable challenge for the object-centric refining stage in leveraging the\nlong-term temporal context representation. To tackle these problems, we propose\na novel paradigm of offboard 3D object detection, named DetZero. Concretely, an\noffline tracker coupled with a multi-frame detector is proposed to focus on the\ncompleteness of generated object tracks. An attention-mechanism refining module\nis proposed to strengthen contextual information interaction across long-term\nsequential point clouds for object refining with decomposed regression methods.\nExtensive experiments on Waymo Open Dataset show our DetZero outperforms all\nstate-of-the-art onboard and offboard 3D detection methods. Notably, DetZero\nranks 1st place on Waymo 3D object detection leaderboard with 85.15 mAPH (L2)\ndetection performance. Further experiments validate the application of taking\nthe place of human labels with such high-quality results. Our empirical study\nleads to rethinking conventions and interesting findings that can guide future\nresearch on offboard 3D object detection.\n", "date": "2023-06-09", "categories": []} {"id": "2307.07944", "title": "Revisiting Domain-Adaptive 3D Object Detection by Reliable, Diverse and\n Class-balanced Pseudo-Labeling", "abstract": " Unsupervised domain adaptation (DA) with the aid of pseudo labeling\ntechniques has emerged as a crucial approach for domain-adaptive 3D object\ndetection. While effective, existing DA methods suffer from a substantial drop\nin performance when applied to a multi-class training setting, due to the\nco-existence of low-quality pseudo labels and class imbalance issues. In this\npaper, we address this challenge by proposing a novel ReDB framework tailored\nfor learning to detect all classes at once. Our approach produces Reliable,\nDiverse, and class-Balanced pseudo 3D boxes to iteratively guide the\nself-training on a distributionally different target domain. To alleviate\ndisruptions caused by the environmental discrepancy (e.g., beam numbers), the\nproposed cross-domain examination (CDE) assesses the correctness of pseudo\nlabels by copy-pasting target instances into a source environment and measuring\nthe prediction consistency. To reduce computational overhead and mitigate the\nobject shift (e.g., scales and point densities), we design an overlapped boxes\ncounting (OBC) metric that allows to uniformly downsample pseudo-labeled\nobjects across different geometric characteristics. To confront the issue of\ninter-class imbalance, we progressively augment the target point clouds with a\nclass-balanced set of pseudo-labeled target instances and source objects, which\nboosts recognition accuracies on both frequently appearing and rare classes.\nExperimental results on three benchmark datasets using both voxel-based (i.e.,\nSECOND) and point-based 3D detectors (i.e., PointRCNN) demonstrate that our\nproposed ReDB approach outperforms existing 3D domain adaptation methods by a\nlarge margin, improving 23.15% mAP on the nuScenes $\\rightarrow$ KITTI task.\nThe code is available at https://github.com/zhuoxiao-chen/ReDB-DA-3Ddet.\n", "date": "2023-07-16", "categories": []} {"id": "2308.09098", "title": "ImGeoNet: Image-induced Geometry-aware Voxel Representation for\n Multi-view 3D Object Detection", "abstract": " We propose ImGeoNet, a multi-view image-based 3D object detection framework\nthat models a 3D space by an image-induced geometry-aware voxel representation.\nUnlike previous methods which aggregate 2D features into 3D voxels without\nconsidering geometry, ImGeoNet learns to induce geometry from multi-view images\nto alleviate the confusion arising from voxels of free space, and during the\ninference phase, only images from multiple views are required. Besides, a\npowerful pre-trained 2D feature extractor can be leveraged by our\nrepresentation, leading to a more robust performance. To evaluate the\neffectiveness of ImGeoNet, we conduct quantitative and qualitative experiments\non three indoor datasets, namely ARKitScenes, ScanNetV2, and ScanNet200. The\nresults demonstrate that ImGeoNet outperforms the current state-of-the-art\nmulti-view image-based method, ImVoxelNet, on all three datasets in terms of\ndetection accuracy. In addition, ImGeoNet shows great data efficiency by\nachieving results comparable to ImVoxelNet with 100 views while utilizing only\n40 views. Furthermore, our studies indicate that our proposed image-induced\ngeometry-aware representation can enable image-based methods to attain superior\ndetection accuracy than the seminal point cloud-based method, VoteNet, in two\npractical scenarios: (1) scenarios where point clouds are sparse and noisy,\nsuch as in ARKitScenes, and (2) scenarios involve diverse object classes,\nparticularly classes of small objects, as in the case in ScanNet200.\n", "date": "2023-08-17", "categories": []} {"id": "2308.08140", "title": "GPA-3D: Geometry-aware Prototype Alignment for Unsupervised Domain\n Adaptive 3D Object Detection from Point Clouds", "abstract": " LiDAR-based 3D detection has made great progress in recent years. However,\nthe performance of 3D detectors is considerably limited when deployed in unseen\nenvironments, owing to the severe domain gap problem. Existing domain adaptive\n3D detection methods do not adequately consider the problem of the\ndistributional discrepancy in feature space, thereby hindering generalization\nof detectors across domains. In this work, we propose a novel unsupervised\ndomain adaptive \\textbf{3D} detection framework, namely \\textbf{G}eometry-aware\n\\textbf{P}rototype \\textbf{A}lignment (\\textbf{GPA-3D}), which explicitly\nleverages the intrinsic geometric relationship from point cloud objects to\nreduce the feature discrepancy, thus facilitating cross-domain transferring.\nSpecifically, GPA-3D assigns a series of tailored and learnable prototypes to\npoint cloud objects with distinct geometric structures. Each prototype aligns\nBEV (bird's-eye-view) features derived from corresponding point cloud objects\non source and target domains, reducing the distributional discrepancy and\nachieving better adaptation. The evaluation results obtained on various\nbenchmarks, including Waymo, nuScenes and KITTI, demonstrate the superiority of\nour GPA-3D over the state-of-the-art approaches for different adaptation\nscenarios. The MindSpore version code will be publicly available at\n\\url{https://github.com/Liz66666/GPA3D}.\n", "date": "2023-08-16", "categories": []} {"id": "2308.07732", "title": "UniTR: A Unified and Efficient Multi-Modal Transformer for\n Bird's-Eye-View Representation", "abstract": " Jointly processing information from multiple sensors is crucial to achieving\naccurate and robust perception for reliable autonomous driving systems.\nHowever, current 3D perception research follows a modality-specific paradigm,\nleading to additional computation overheads and inefficient collaboration\nbetween different sensor data. In this paper, we present an efficient\nmulti-modal backbone for outdoor 3D perception named UniTR, which processes a\nvariety of modalities with unified modeling and shared parameters. Unlike\nprevious works, UniTR introduces a modality-agnostic transformer encoder to\nhandle these view-discrepant sensor data for parallel modal-wise representation\nlearning and automatic cross-modal interaction without additional fusion steps.\nMore importantly, to make full use of these complementary sensor types, we\npresent a novel multi-modal integration strategy by both considering\nsemantic-abundant 2D perspective and geometry-aware 3D sparse neighborhood\nrelations. UniTR is also a fundamentally task-agnostic backbone that naturally\nsupports different 3D perception tasks. It sets a new state-of-the-art\nperformance on the nuScenes benchmark, achieving +1.1 NDS higher for 3D object\ndetection and +12.0 higher mIoU for BEV map segmentation with lower inference\nlatency. Code will be available at https://github.com/Haiyang-W/UniTR .\n", "date": "2023-08-15", "categories": []} {"id": "2308.07748", "title": "Exploiting Sparsity in Automotive Radar Object Detection Networks", "abstract": " Having precise perception of the environment is crucial for ensuring the\nsecure and reliable functioning of autonomous driving systems. Radar object\ndetection networks are one fundamental part of such systems. CNN-based object\ndetectors showed good performance in this context, but they require large\ncompute resources. This paper investigates sparse convolutional object\ndetection networks, which combine powerful grid-based detection with low\ncompute resources. We investigate radar specific challenges and propose sparse\nkernel point pillars (SKPP) and dual voxel point convolutions (DVPC) as\nremedies for the grid rendering and sparse backbone architectures. We evaluate\nour SKPP-DPVCN architecture on nuScenes, which outperforms the baseline by\n5.89% and the previous state of the art by 4.19% in Car AP4.0. Moreover,\nSKPP-DPVCN reduces the average scale error (ASE) by 21.41% over the baseline.\n", "date": "2023-08-15", "categories": []} {"id": "2304.13031", "title": "DQS3D: Densely-matched Quantization-aware Semi-supervised 3D Detection", "abstract": " In this paper, we study the problem of semi-supervised 3D object detection,\nwhich is of great importance considering the high annotation cost for cluttered\n3D indoor scenes. We resort to the robust and principled framework of\nselfteaching, which has triggered notable progress for semisupervised learning\nrecently. While this paradigm is natural for image-level or pixel-level\nprediction, adapting it to the detection problem is challenged by the issue of\nproposal matching. Prior methods are based upon two-stage pipelines, matching\nheuristically selected proposals generated in the first stage and resulting in\nspatially sparse training signals. In contrast, we propose the first\nsemisupervised 3D detection algorithm that works in the singlestage manner and\nallows spatially dense training signals. A fundamental issue of this new design\nis the quantization error caused by point-to-voxel discretization, which\ninevitably leads to misalignment between two transformed views in the voxel\ndomain. To this end, we derive and implement closed-form rules that compensate\nthis misalignment onthe-fly. Our results are significant, e.g., promoting\nScanNet mAP@0.5 from 35.2% to 48.5% using 20% annotation. Codes and data will\nbe publicly available.\n", "date": "2023-04-25", "categories": []} {"id": "2308.05478", "title": "Reviewing 3D Object Detectors in the Context of High-Resolution 3+1D\n Radar", "abstract": " Recent developments and the beginning market introduction of high-resolution\nimaging 4D (3+1D) radar sensors have initialized deep learning-based radar\nperception research. We investigate deep learning-based models operating on\nradar point clouds for 3D object detection. 3D object detection on lidar point\ncloud data is a mature area of 3D vision. Many different architectures have\nbeen proposed, each with strengths and weaknesses. Due to similarities between\n3D lidar point clouds and 3+1D radar point clouds, those existing 3D object\ndetectors are a natural basis to start deep learning-based 3D object detection\non radar data. Thus, the first step is to analyze the detection performance of\nthe existing models on the new data modality and evaluate them in depth. In\norder to apply existing 3D point cloud object detectors developed for lidar\npoint clouds to the radar domain, they need to be adapted first. While some\ndetectors, such as PointPillars, have already been adapted to be applicable to\nradar data, we have adapted others, e.g., Voxel R-CNN, SECOND, PointRCNN, and\nPV-RCNN. To this end, we conduct a cross-model validation (evaluating a set of\nmodels on one particular data set) as well as a cross-data set validation\n(evaluating all models in the model set on several data sets). The\nhigh-resolution radar data used are the View-of-Delft and Astyx data sets.\nFinally, we evaluate several adaptations of the models and their training\nprocedures. We also discuss major factors influencing the detection performance\non radar data and propose possible solutions indicating potential future\nresearch avenues.\n", "date": "2023-08-10", "categories": []} {"id": "2307.08209", "title": "Ada3D : Exploiting the Spatial Redundancy with Adaptive Inference for\n Efficient 3D Object Detection", "abstract": " Voxel-based methods have achieved state-of-the-art performance for 3D object\ndetection in autonomous driving. However, their significant computational and\nmemory costs pose a challenge for their application to resource-constrained\nvehicles. One reason for this high resource consumption is the presence of a\nlarge number of redundant background points in Lidar point clouds, resulting in\nspatial redundancy in both 3D voxel and dense BEV map representations. To\naddress this issue, we propose an adaptive inference framework called Ada3D,\nwhich focuses on exploiting the input-level spatial redundancy. Ada3D\nadaptively filters the redundant input, guided by a lightweight importance\npredictor and the unique properties of the Lidar point cloud. Additionally, we\nutilize the BEV features' intrinsic sparsity by introducing the Sparsity\nPreserving Batch Normalization. With Ada3D, we achieve 40% reduction for 3D\nvoxels and decrease the density of 2D BEV feature maps from 100% to 20% without\nsacrificing accuracy. Ada3D reduces the model computational and memory cost by\n5x, and achieves 1.52x/1.45x end-to-end GPU latency and 1.5x/4.5x GPU peak\nmemory optimization for the 3D and 2D backbone respectively.\n", "date": "2023-07-17", "categories": []} {"id": "2308.04515", "title": "Toward unlabeled multi-view 3D pedestrian detection by generalizable AI:\n techniques and performance analysis", "abstract": " We unveil how generalizable AI can be used to improve multi-view 3D\npedestrian detection in unlabeled target scenes. One way to increase\ngeneralization to new scenes is to automatically label target data, which can\nthen be used for training a detector model. In this context, we investigate two\napproaches for automatically labeling target data: pseudo-labeling using a\nsupervised detector and automatic labeling using an untrained detector (that\ncan be applied out of the box without any training). We adopt a training\nframework for optimizing detector models using automatic labeling procedures.\nThis framework encompasses different training sets/modes and multi-round\nautomatic labeling strategies. We conduct our analyses on the\npublicly-available WILDTRACK and MultiviewX datasets. We show that, by using\nthe automatic labeling approach based on an untrained detector, we can obtain\nsuperior results than directly using the untrained detector or a detector\ntrained with an existing labeled source dataset. It achieved a MODA about 4%\nand 1% better than the best existing unlabeled method when using WILDTRACK and\nMultiviewX as target datasets, respectively.\n", "date": "2023-08-08", "categories": []} {"id": "2308.04556", "title": "FocalFormer3D : Focusing on Hard Instance for 3D Object Detection", "abstract": " False negatives (FN) in 3D object detection, {\\em e.g.}, missing predictions\nof pedestrians, vehicles, or other obstacles, can lead to potentially dangerous\nsituations in autonomous driving. While being fatal, this issue is understudied\nin many current 3D detection methods. In this work, we propose Hard Instance\nProbing (HIP), a general pipeline that identifies \\textit{FN} in a multi-stage\nmanner and guides the models to focus on excavating difficult instances. For 3D\nobject detection, we instantiate this method as FocalFormer3D, a simple yet\neffective detector that excels at excavating difficult objects and improving\nprediction recall. FocalFormer3D features a multi-stage query generation to\ndiscover hard objects and a box-level transformer decoder to efficiently\ndistinguish objects from massive object candidates. Experimental results on the\nnuScenes and Waymo datasets validate the superior performance of FocalFormer3D.\nThe advantage leads to strong performance on both detection and tracking, in\nboth LiDAR and multi-modal settings. Notably, FocalFormer3D achieves a 70.5 mAP\nand 73.9 NDS on nuScenes detection benchmark, while the nuScenes tracking\nbenchmark shows 72.1 AMOTA, both ranking 1st place on the nuScenes LiDAR\nleaderboard. Our code is available at\n\\url{https://github.com/NVlabs/FocalFormer3D}.\n", "date": "2023-08-08", "categories": []} {"id": "2308.04054", "title": "An Empirical Analysis of Range for 3D Object Detection", "abstract": " LiDAR-based 3D detection plays a vital role in autonomous navigation.\nSurprisingly, although autonomous vehicles (AVs) must detect both near-field\nobjects (for collision avoidance) and far-field objects (for longer-term\nplanning), contemporary benchmarks focus only on near-field 3D detection.\nHowever, AVs must detect far-field objects for safe navigation. In this paper,\nwe present an empirical analysis of far-field 3D detection using the long-range\ndetection dataset Argoverse 2.0 to better understand the problem, and share the\nfollowing insight: near-field LiDAR measurements are dense and optimally\nencoded by small voxels, while far-field measurements are sparse and are better\nencoded with large voxels. We exploit this observation to build a collection of\nrange experts tuned for near-vs-far field detection, and propose simple\ntechniques to efficiently ensemble models for long-range detection that improve\nefficiency by 33% and boost accuracy by 3.2% CDS.\n", "date": "2023-08-08", "categories": []} {"id": "2308.04409", "title": "V-DETR: DETR with Vertex Relative Position Encoding for 3D Object\n Detection", "abstract": " We introduce a highly performant 3D object detector for point clouds using\nthe DETR framework. The prior attempts all end up with suboptimal results\nbecause they fail to learn accurate inductive biases from the limited scale of\ntraining data. In particular, the queries often attend to points that are far\naway from the target objects, violating the locality principle in object\ndetection. To address the limitation, we introduce a novel 3D Vertex Relative\nPosition Encoding (3DV-RPE) method which computes position encoding for each\npoint based on its relative position to the 3D boxes predicted by the queries\nin each decoder layer, thus providing clear information to guide the model to\nfocus on points near the objects, in accordance with the principle of locality.\nIn addition, we systematically improve the pipeline from various aspects such\nas data normalization based on our understanding of the task. We show\nexceptional results on the challenging ScanNetV2 benchmark, achieving\nsignificant improvements over the previous 3DETR in\n$\\rm{AP}_{25}$/$\\rm{AP}_{50}$ from 65.0\\%/47.0\\% to 77.8\\%/66.0\\%,\nrespectively. In addition, our method sets a new record on ScanNetV2 and SUN\nRGB-D datasets.Code will be released at http://github.com/yichaoshen-MS/V-DETR.\n", "date": "2023-08-08", "categories": []} {"id": "2210.10180", "title": "Domain Adaptation in 3D Object Detection with Gradual Batch Alternation\n Training", "abstract": " We consider the problem of domain adaptation in LiDAR-based 3D object\ndetection. Towards this, we propose a simple yet effective training strategy\ncalled Gradual Batch Alternation that can adapt from a large labeled source\ndomain to an insufficiently labeled target domain. The idea is to initiate the\ntraining with the batch of samples from the source and target domain data in an\nalternate fashion, but then gradually reduce the amount of the source domain\ndata over time as the training progresses. This way the model slowly shifts\ntowards the target domain and eventually better adapt to it. The domain\nadaptation experiments for 3D object detection on four benchmark autonomous\ndriving datasets, namely ONCE, PandaSet, Waymo, and nuScenes, demonstrate\nsignificant performance gains over prior arts and strong baselines.\n", "date": "2022-10-18", "categories": []} {"id": "2308.03755", "title": "FSD V2: Improving Fully Sparse 3D Object Detection with Virtual Voxels", "abstract": " LiDAR-based fully sparse architecture has garnered increasing attention.\nFSDv1 stands out as a representative work, achieving impressive efficacy and\nefficiency, albeit with intricate structures and handcrafted designs. In this\npaper, we present FSDv2, an evolution that aims to simplify the previous FSDv1\nwhile eliminating the inductive bias introduced by its handcrafted\ninstance-level representation, thus promoting better general applicability. To\nthis end, we introduce the concept of \\textbf{virtual voxels}, which takes over\nthe clustering-based instance segmentation in FSDv1. Virtual voxels not only\naddress the notorious issue of the Center Feature Missing problem in fully\nsparse detectors but also endow the framework with a more elegant and\nstreamlined approach. Consequently, we develop a suite of components to\ncomplement the virtual voxel concept, including a virtual voxel encoder, a\nvirtual voxel mixer, and a virtual voxel assignment strategy. Through empirical\nvalidation, we demonstrate that the virtual voxel mechanism is functionally\nsimilar to the handcrafted clustering in FSDv1 while being more general. We\nconduct experiments on three large-scale datasets: Waymo Open Dataset,\nArgoverse 2 dataset, and nuScenes dataset. Our results showcase\nstate-of-the-art performance on all three datasets, highlighting the\nsuperiority of FSDv2 in long-range scenarios and its general applicability to\nachieve competitive performance across diverse scenarios. Moreover, we provide\ncomprehensive experimental analysis to elucidate the workings of FSDv2. To\nfoster reproducibility and further research, we have open-sourced FSDv2 at\nhttps://github.com/tusen-ai/SST.\n", "date": "2023-08-07", "categories": []} {"id": "2308.01000", "title": "MDT3D: Multi-Dataset Training for LiDAR 3D Object Detection\n Generalization", "abstract": " Supervised 3D Object Detection models have been displaying increasingly\nbetter performance in single-domain cases where the training data comes from\nthe same environment and sensor as the testing data. However, in real-world\nscenarios data from the target domain may not be available for finetuning or\nfor domain adaptation methods. Indeed, 3D object detection models trained on a\nsource dataset with a specific point distribution have shown difficulties in\ngeneralizing to unseen datasets. Therefore, we decided to leverage the\ninformation available from several annotated source datasets with our\nMulti-Dataset Training for 3D Object Detection (MDT3D) method to increase the\nrobustness of 3D object detection models when tested in a new environment with\na different sensor configuration. To tackle the labelling gap between datasets,\nwe used a new label mapping based on coarse labels. Furthermore, we show how we\nmanaged the mix of datasets during training and finally introduce a new\ncross-dataset augmentation method: cross-dataset object injection. We\ndemonstrate that this training paradigm shows improvements for different types\nof 3D object detection models. The source code and additional results for this\nresearch project will be publicly available on GitHub for interested parties to\naccess and utilize: https://github.com/LouisSF/MDT3D\n", "date": "2023-08-02", "categories": []} {"id": "2308.00596", "title": "MonoNext: A 3D Monocular Object Detection with ConvNext", "abstract": " Autonomous driving perception tasks rely heavily on cameras as the primary\nsensor for Object Detection, Semantic Segmentation, Instance Segmentation, and\nObject Tracking. However, RGB images captured by cameras lack depth\ninformation, which poses a significant challenge in 3D detection tasks. To\nsupplement this missing data, mapping sensors such as LIDAR and RADAR are used\nfor accurate 3D Object Detection. Despite their significant accuracy, the\nmulti-sensor models are expensive and require a high computational demand. In\ncontrast, Monocular 3D Object Detection models are becoming increasingly\npopular, offering a faster, cheaper, and easier-to-implement solution for 3D\ndetections. This paper introduces a different Multi-Tasking Learning approach\ncalled MonoNext that utilizes a spatial grid to map objects in the scene.\nMonoNext employs a straightforward approach based on the ConvNext network and\nrequires only 3D bounding box annotated data. In our experiments with the KITTI\ndataset, MonoNext achieved high precision and competitive performance\ncomparable with state-of-the-art approaches. Furthermore, by adding more\ntraining data, MonoNext surpassed itself and achieved higher accuracies.\n", "date": "2023-08-01", "categories": []} {"id": "2307.16617", "title": "FULLER: Unified Multi-modality Multi-task 3D Perception via Multi-level\n Gradient Calibration", "abstract": " Multi-modality fusion and multi-task learning are becoming trendy in 3D\nautonomous driving scenario, considering robust prediction and computation\nbudget. However, naively extending the existing framework to the domain of\nmulti-modality multi-task learning remains ineffective and even poisonous due\nto the notorious modality bias and task conflict. Previous works manually\ncoordinate the learning framework with empirical knowledge, which may lead to\nsub-optima. To mitigate the issue, we propose a novel yet simple multi-level\ngradient calibration learning framework across tasks and modalities during\noptimization. Specifically, the gradients, produced by the task heads and used\nto update the shared backbone, will be calibrated at the backbone's last layer\nto alleviate the task conflict. Before the calibrated gradients are further\npropagated to the modality branches of the backbone, their magnitudes will be\ncalibrated again to the same level, ensuring the downstream tasks pay balanced\nattention to different modalities. Experiments on large-scale benchmark\nnuScenes demonstrate the effectiveness of the proposed method, eg, an absolute\n14.4% mIoU improvement on map segmentation and 1.4% mAP improvement on 3D\ndetection, advancing the application of 3D autonomous driving in the domain of\nmulti-modality fusion and multi-task learning. We also discuss the links\nbetween modalities and tasks.\n", "date": "2023-07-31", "categories": []} {"id": "2211.14710", "title": "3DPPE: 3D Point Positional Encoding for Multi-Camera 3D Object Detection\n Transformers", "abstract": " Transformer-based methods have swept the benchmarks on 2D and 3D detection on\nimages. Because tokenization before the attention mechanism drops the spatial\ninformation, positional encoding becomes critical for those methods. Recent\nworks found that encodings based on samples of the 3D viewing rays can\nsignificantly improve the quality of multi-camera 3D object detection. We\nhypothesize that 3D point locations can provide more information than rays.\nTherefore, we introduce 3D point positional encoding, 3DPPE, to the 3D\ndetection Transformer decoder. Although 3D measurements are not available at\nthe inference time of monocular 3D object detection, 3DPPE uses predicted depth\nto approximate the real point positions. Our hybriddepth module combines direct\nand categorical depth to estimate the refined depth of each pixel. Despite the\napproximation, 3DPPE achieves 46.0 mAP and 51.4 NDS on the competitive nuScenes\ndataset, significantly outperforming encodings based on ray samples. We make\nthe codes available at https://github.com/drilistbox/3DPPE.\n", "date": "2022-11-27", "categories": []} {"id": "2302.08231", "title": "3M3D: Multi-view, Multi-path, Multi-representation for 3D Object\n Detection", "abstract": " 3D visual perception tasks based on multi-camera images are essential for\nautonomous driving systems. Latest work in this field performs 3D object\ndetection by leveraging multi-view images as an input and iteratively enhancing\nobject queries (object proposals) by cross-attending multi-view features.\nHowever, individual backbone features are not updated with multi-view features\nand it stays as a mere collection of the output of the single-image backbone\nnetwork. Therefore we propose 3M3D: A Multi-view, Multi-path,\nMulti-representation for 3D Object Detection where we update both multi-view\nfeatures and query features to enhance the representation of the scene in both\nfine panoramic view and coarse global view. Firstly, we update multi-view\nfeatures by multi-view axis self-attention. It will incorporate panoramic\ninformation in the multi-view features and enhance understanding of the global\nscene. Secondly, we update multi-view features by self-attention of the ROI\n(Region of Interest) windows which encodes local finer details in the features.\nIt will help exchange the information not only along the multi-view axis but\nalso along the other spatial dimension. Lastly, we leverage the fact of\nmulti-representation of queries in different domains to further boost the\nperformance. Here we use sparse floating queries along with dense BEV (Bird's\nEye View) queries, which are later post-processed to filter duplicate\ndetections. Moreover, we show performance improvements on nuScenes benchmark\ndataset on top of our baselines.\n", "date": "2023-02-16", "categories": []} {"id": "2307.14620", "title": "NeRF-Det: Learning Geometry-Aware Volumetric Representation for\n Multi-View 3D Object Detection", "abstract": " We present NeRF-Det, a novel method for indoor 3D detection with posed RGB\nimages as input. Unlike existing indoor 3D detection methods that struggle to\nmodel scene geometry, our method makes novel use of NeRF in an end-to-end\nmanner to explicitly estimate 3D geometry, thereby improving 3D detection\nperformance. Specifically, to avoid the significant extra latency associated\nwith per-scene optimization of NeRF, we introduce sufficient geometry priors to\nenhance the generalizability of NeRF-MLP. Furthermore, we subtly connect the\ndetection and NeRF branches through a shared MLP, enabling an efficient\nadaptation of NeRF to detection and yielding geometry-aware volumetric\nrepresentations for 3D detection. Our method outperforms state-of-the-arts by\n3.9 mAP and 3.1 mAP on the ScanNet and ARKITScenes benchmarks, respectively. We\nprovide extensive analysis to shed light on how NeRF-Det works. As a result of\nour joint-training design, NeRF-Det is able to generalize well to unseen scenes\nfor object detection, view synthesis, and depth estimation tasks without\nrequiring per-scene optimization. Code is available at\n\\url{https://github.com/facebookresearch/NeRF-Det}.\n", "date": "2023-07-27", "categories": []} {"id": "2307.12637", "title": "PG-RCNN: Semantic Surface Point Generation for 3D Object Detection", "abstract": " One of the main challenges in LiDAR-based 3D object detection is that the\nsensors often fail to capture the complete spatial information about the\nobjects due to long distance and occlusion. Two-stage detectors with point\ncloud completion approaches tackle this problem by adding more points to the\nregions of interest (RoIs) with a pre-trained network. However, these methods\ngenerate dense point clouds of objects for all region proposals, assuming that\nobjects always exist in the RoIs. This leads to the indiscriminate point\ngeneration for incorrect proposals as well. Motivated by this, we propose Point\nGeneration R-CNN (PG-RCNN), a novel end-to-end detector that generates semantic\nsurface points of foreground objects for accurate detection. Our method uses a\njointly trained RoI point generation module to process the contextual\ninformation of RoIs and estimate the complete shape and displacement of\nforeground objects. For every generated point, PG-RCNN assigns a semantic\nfeature that indicates the estimated foreground probability. Extensive\nexperiments show that the point clouds generated by our method provide\ngeometrically and semantically rich information for refining false positive and\nmisaligned proposals. PG-RCNN achieves competitive performance on the KITTI\nbenchmark, with significantly fewer parameters than state-of-the-art models.\nThe code is available at https://github.com/quotation2520/PG-RCNN.\n", "date": "2023-07-24", "categories": []} {"id": "2307.12972", "title": "DFA3D: 3D Deformable Attention For 2D-to-3D Feature Lifting", "abstract": " In this paper, we propose a new operator, called 3D DeFormable Attention\n(DFA3D), for 2D-to-3D feature lifting, which transforms multi-view 2D image\nfeatures into a unified 3D space for 3D object detection. Existing feature\nlifting approaches, such as Lift-Splat-based and 2D attention-based, either use\nestimated depth to get pseudo LiDAR features and then splat them to a 3D space,\nwhich is a one-pass operation without feature refinement, or ignore depth and\nlift features by 2D attention mechanisms, which achieve finer semantics while\nsuffering from a depth ambiguity problem. In contrast, our DFA3D-based method\nfirst leverages the estimated depth to expand each view's 2D feature map to 3D\nand then utilizes DFA3D to aggregate features from the expanded 3D feature\nmaps. With the help of DFA3D, the depth ambiguity problem can be effectively\nalleviated from the root, and the lifted features can be progressively refined\nlayer by layer, thanks to the Transformer-like architecture. In addition, we\npropose a mathematically equivalent implementation of DFA3D which can\nsignificantly improve its memory efficiency and computational speed. We\nintegrate DFA3D into several methods that use 2D attention-based feature\nlifting with only a few modifications in code and evaluate on the nuScenes\ndataset. The experiment results show a consistent improvement of +1.41\\% mAP on\naverage, and up to +15.1\\% mAP improvement when high-quality depth information\nis available, demonstrating the superiority, applicability, and huge potential\nof DFA3D. The code is available at\nhttps://github.com/IDEA-Research/3D-deformable-attention.git.\n", "date": "2023-07-24", "categories": []} {"id": "2307.11323", "title": "HVDetFusion: A Simple and Robust Camera-Radar Fusion Framework", "abstract": " In the field of autonomous driving, 3D object detection is a very important\nperception module. Although the current SOTA algorithm combines Camera and\nLidar sensors, limited by the high price of Lidar, the current mainstream\nlanding schemes are pure Camera sensors or Camera+Radar sensors. In this study,\nwe propose a new detection algorithm called HVDetFusion, which is a multi-modal\ndetection algorithm that not only supports pure camera data as input for\ndetection, but also can perform fusion input of radar data and camera data. The\ncamera stream does not depend on the input of Radar data, thus addressing the\ndownside of previous methods. In the pure camera stream, we modify the\nframework of Bevdet4D for better perception and more efficient inference, and\nthis stream has the whole 3D detection output. Further, to incorporate the\nbenefits of Radar signals, we use the prior information of different object\npositions to filter the false positive information of the original radar data,\naccording to the positioning information and radial velocity information\nrecorded by the radar sensors to supplement and fuse the BEV features generated\nby the original camera data, and the effect is further improved in the process\nof fusion training. Finally, HVDetFusion achieves the new state-of-the-art\n67.4\\% NDS on the challenging nuScenes test set among all camera-radar 3D\nobject detectors. The code is available at\nhttps://github.com/HVXLab/HVDetFusion\n", "date": "2023-07-21", "categories": []} {"id": "2307.11477", "title": "SA-BEV: Generating Semantic-Aware Bird's-Eye-View Feature for Multi-view\n 3D Object Detection", "abstract": " Recently, the pure camera-based Bird's-Eye-View (BEV) perception provides a\nfeasible solution for economical autonomous driving. However, the existing\nBEV-based multi-view 3D detectors generally transform all image features into\nBEV features, without considering the problem that the large proportion of\nbackground information may submerge the object information. In this paper, we\npropose Semantic-Aware BEV Pooling (SA-BEVPool), which can filter out\nbackground information according to the semantic segmentation of image features\nand transform image features into semantic-aware BEV features. Accordingly, we\npropose BEV-Paste, an effective data augmentation strategy that closely matches\nwith semantic-aware BEV feature. In addition, we design a Multi-Scale\nCross-Task (MSCT) head, which combines task-specific and cross-task information\nto predict depth distribution and semantic segmentation more accurately,\nfurther improving the quality of semantic-aware BEV feature. Finally, we\nintegrate the above modules into a novel multi-view 3D object detection\nframework, namely SA-BEV. Experiments on nuScenes show that SA-BEV achieves\nstate-of-the-art performance. Code has been available at\nhttps://github.com/mengtan00/SA-BEV.git.\n", "date": "2023-07-21", "categories": []} {"id": "2307.09155", "title": "MLF-DET: Multi-Level Fusion for Cross-Modal 3D Object Detection", "abstract": " In this paper, we propose a novel and effective Multi-Level Fusion network,\nnamed as MLF-DET, for high-performance cross-modal 3D object DETection, which\nintegrates both the feature-level fusion and decision-level fusion to fully\nutilize the information in the image. For the feature-level fusion, we present\nthe Multi-scale Voxel Image fusion (MVI) module, which densely aligns\nmulti-scale voxel features with image features. For the decision-level fusion,\nwe propose the lightweight Feature-cued Confidence Rectification (FCR) module\nwhich further exploits image semantics to rectify the confidence of detection\ncandidates. Besides, we design an effective data augmentation strategy termed\nOcclusion-aware GT Sampling (OGS) to reserve more sampled objects in the\ntraining scenes, so as to reduce overfitting. Extensive experiments on the\nKITTI dataset demonstrate the effectiveness of our method. Notably, on the\nextremely competitive KITTI car 3D object detection benchmark, our method\nreaches 82.89% moderate AP and achieves state-of-the-art performance without\nbells and whistles.\n", "date": "2023-07-18", "categories": []} {"id": "2209.12068", "title": "NeRF-Loc: Transformer-Based Object Localization Within Neural Radiance\n Fields", "abstract": " Neural Radiance Fields (NeRFs) have become a widely-applied scene\nrepresentation technique in recent years, showing advantages for robot\nnavigation and manipulation tasks. To further advance the utility of NeRFs for\nrobotics, we propose a transformer-based framework, NeRF-Loc, to extract 3D\nbounding boxes of objects in NeRF scenes. NeRF-Loc takes a pre-trained NeRF\nmodel and camera view as input and produces labeled, oriented 3D bounding boxes\nof objects as output. Using current NeRF training tools, a robot can train a\nNeRF environment model in real-time and, using our algorithm, identify 3D\nbounding boxes of objects of interest within the NeRF for downstream navigation\nor manipulation tasks. Concretely, we design a pair of paralleled transformer\nencoder branches, namely the coarse stream and the fine stream, to encode both\nthe context and details of target objects. The encoded features are then fused\ntogether with attention layers to alleviate ambiguities for accurate object\nlocalization. We have compared our method with conventional RGB(-D) based\nmethods that take rendered RGB images and depths from NeRFs as inputs. Our\nmethod is better than the baselines.\n", "date": "2022-09-24", "categories": []} {"id": "2307.08233", "title": "ROFusion: Efficient Object Detection using Hybrid Point-wise\n Radar-Optical Fusion", "abstract": " Radars, due to their robustness to adverse weather conditions and ability to\nmeasure object motions, have served in autonomous driving and intelligent\nagents for years. However, Radar-based perception suffers from its unintuitive\nsensing data, which lack of semantic and structural information of scenes. To\ntackle this problem, camera and Radar sensor fusion has been investigated as a\ntrending strategy with low cost, high reliability and strong maintenance. While\nmost recent works explore how to explore Radar point clouds and images, rich\ncontextual information within Radar observation are discarded. In this paper,\nwe propose a hybrid point-wise Radar-Optical fusion approach for object\ndetection in autonomous driving scenarios. The framework benefits from dense\ncontextual information from both the range-doppler spectrum and images which\nare integrated to learn a multi-modal feature representation. Furthermore, we\npropose a novel local coordinate formulation, tackling the object detection\ntask in an object-centric coordinate. Extensive results show that with the\ninformation gained from optical images, we could achieve leading performance in\nobject detection (97.69\\% recall) compared to recent state-of-the-art methods\nFFT-RadNet (82.86\\% recall). Ablation studies verify the key design choices and\npracticability of our approach given machine generated imperfect detections.\nThe code will be available at https://github.com/LiuLiu-55/ROFusion.\n", "date": "2023-07-17", "categories": []} {"id": "2307.08415", "title": "Monocular 3D Object Detection with LiDAR Guided Semi Supervised Active\n Learning", "abstract": " We propose a novel semi-supervised active learning (SSAL) framework for\nmonocular 3D object detection with LiDAR guidance (MonoLiG), which leverages\nall modalities of collected data during model development. We utilize LiDAR to\nguide the data selection and training of monocular 3D detectors without\nintroducing any overhead in the inference phase. During training, we leverage\nthe LiDAR teacher, monocular student cross-modal framework from semi-supervised\nlearning to distill information from unlabeled data as pseudo-labels. To handle\nthe differences in sensor characteristics, we propose a data noise-based\nweighting mechanism to reduce the effect of propagating noise from LiDAR\nmodality to monocular. For selecting which samples to label to improve the\nmodel performance, we propose a sensor consistency-based selection score that\nis also coherent with the training objective. Extensive experimental results on\nKITTI and Waymo datasets verify the effectiveness of our proposed framework. In\nparticular, our selection strategy consistently outperforms state-of-the-art\nactive learning baselines, yielding up to 17% better saving rate in labeling\ncosts. Our training strategy attains the top place in KITTI 3D and\nbirds-eye-view (BEV) monocular object detection official benchmarks by\nimproving the BEV Average Precision (AP) by 2.02.\n", "date": "2023-07-17", "categories": []} {"id": "2307.04106", "title": "Parametric Depth Based Feature Representation Learning for Object\n Detection and Segmentation in Bird's Eye View", "abstract": " Recent vision-only perception models for autonomous driving achieved\npromising results by encoding multi-view image features into Bird's-Eye-View\n(BEV) space. A critical step and the main bottleneck of these methods is\ntransforming image features into the BEV coordinate frame. This paper focuses\non leveraging geometry information, such as depth, to model such feature\ntransformation. Existing works rely on non-parametric depth distribution\nmodeling leading to significant memory consumption, or ignore the geometry\ninformation to address this problem. In contrast, we propose to use parametric\ndepth distribution modeling for feature transformation. We first lift the 2D\nimage features to the 3D space defined for the ego vehicle via a predicted\nparametric depth distribution for each pixel in each view. Then, we aggregate\nthe 3D feature volume based on the 3D space occupancy derived from depth to the\nBEV frame. Finally, we use the transformed features for downstream tasks such\nas object detection and semantic segmentation. Existing semantic segmentation\nmethods do also suffer from an hallucination problem as they do not take\nvisibility information into account. This hallucination can be particularly\nproblematic for subsequent modules such as control and planning. To mitigate\nthe issue, our method provides depth uncertainty and reliable visibility-aware\nestimations. We further leverage our parametric depth modeling to present a\nnovel visibility-aware evaluation metric that, when taken into account, can\nmitigate the hallucination problem. Extensive experiments on object detection\nand semantic segmentation on the nuScenes datasets demonstrate that our method\noutperforms existing methods on both tasks.\n", "date": "2023-07-09", "categories": []} {"id": "2307.01807", "title": "SUIT: Learning Significance-guided Information for 3D Temporal Detection", "abstract": " 3D object detection from LiDAR point cloud is of critical importance for\nautonomous driving and robotics. While sequential point cloud has the potential\nto enhance 3D perception through temporal information, utilizing these temporal\nfeatures effectively and efficiently remains a challenging problem. Based on\nthe observation that the foreground information is sparsely distributed in\nLiDAR scenes, we believe sufficient knowledge can be provided by sparse format\nrather than dense maps. To this end, we propose to learn Significance-gUided\nInformation for 3D Temporal detection (SUIT), which simplifies temporal\ninformation as sparse features for information fusion across frames.\nSpecifically, we first introduce a significant sampling mechanism that extracts\ninformation-rich yet sparse features based on predicted object centroids. On\ntop of that, we present an explicit geometric transformation learning\ntechnique, which learns the object-centric transformations among sparse\nfeatures across frames. We evaluate our method on large-scale nuScenes and\nWaymo dataset, where our SUIT not only significantly reduces the memory and\ncomputation cost of temporal fusion, but also performs well over the\nstate-of-the-art baselines.\n", "date": "2023-07-04", "categories": []} {"id": "2307.02270", "title": "SVDM: Single-View Diffusion Model for Pseudo-Stereo 3D Object Detection", "abstract": " One of the key problems in 3D object detection is to reduce the accuracy gap\nbetween methods based on LiDAR sensors and those based on monocular cameras. A\nrecently proposed framework for monocular 3D detection based on Pseudo-Stereo\nhas received considerable attention in the community. However, so far these two\nproblems are discovered in existing practices, including (1) monocular depth\nestimation and Pseudo-Stereo detector must be trained separately, (2) Difficult\nto be compatible with different stereo detectors and (3) the overall\ncalculation is large, which affects the reasoning speed. In this work, we\npropose an end-to-end, efficient pseudo-stereo 3D detection framework by\nintroducing a Single-View Diffusion Model (SVDM) that uses a few iterations to\ngradually deliver right informative pixels to the left image. SVDM allows the\nentire pseudo-stereo 3D detection pipeline to be trained end-to-end and can\nbenefit from the training of stereo detectors. Afterwards, we further explore\nthe application of SVDM in depth-free stereo 3D detection, and the final\nframework is compatible with most stereo detectors. Among multiple benchmarks\non the KITTI dataset, we achieve new state-of-the-art performance.\n", "date": "2023-07-05", "categories": []} {"id": "2206.14451", "title": "SRCN3D: Sparse R-CNN 3D for Compact Convolutional Multi-View 3D Object\n Detection and Tracking", "abstract": " Detection and tracking of moving objects is an essential component in\nenvironmental perception for autonomous driving. In the flourishing field of\nmulti-view 3D camera-based detectors, different transformer-based pipelines are\ndesigned to learn queries in 3D space from 2D feature maps of perspective\nviews, but the dominant dense BEV query mechanism is computationally\ninefficient. This paper proposes Sparse R-CNN 3D (SRCN3D), a novel two-stage\nfully-sparse detector that incorporates sparse queries, sparse attention with\nbox-wise sampling, and sparse prediction. SRCN3D adopts a cascade structure\nwith the twin-track update of both a fixed number of query boxes and latent\nquery features. Our novel sparse feature sampling module only utilizes local 2D\nregion of interest (RoI) features calculated by the projection of 3D query\nboxes for further box refinement, leading to a fully-convolutional and\ndeployment-friendly pipeline. For multi-object tracking, motion features, query\nfeatures and RoI features are comprehensively utilized in multi-hypotheses data\nassociation. Extensive experiments on nuScenes dataset demonstrate that SRCN3D\nachieves competitive performance in both 3D object detection and multi-object\ntracking tasks, while also exhibiting superior efficiency compared to\ntransformer-based methods. Code and models are available at\nhttps://github.com/synsin0/SRCN3D.\n", "date": "2022-06-29", "categories": []} {"id": "2211.16779", "title": "Attention-Based Depth Distillation with 3D-Aware Positional Encoding for\n Monocular 3D Object Detection", "abstract": " Monocular 3D object detection is a low-cost but challenging task, as it\nrequires generating accurate 3D localization solely from a single image input.\nRecent developed depth-assisted methods show promising results by using\nexplicit depth maps as intermediate features, which are either precomputed by\nmonocular depth estimation networks or jointly evaluated with 3D object\ndetection. However, inevitable errors from estimated depth priors may lead to\nmisaligned semantic information and 3D localization, hence resulting in feature\nsmearing and suboptimal predictions. To mitigate this issue, we propose ADD, an\nAttention-based Depth knowledge Distillation framework with 3D-aware positional\nencoding. Unlike previous knowledge distillation frameworks that adopt stereo-\nor LiDAR-based teachers, we build up our teacher with identical architecture as\nthe student but with extra ground-truth depth as input. Credit to our teacher\ndesign, our framework is seamless, domain-gap free, easily implementable, and\nis compatible with object-wise ground-truth depth. Specifically, we leverage\nintermediate features and responses for knowledge distillation. Considering\nlong-range 3D dependencies, we propose \\emph{3D-aware self-attention} and\n\\emph{target-aware cross-attention} modules for student adaptation. Extensive\nexperiments are performed to verify the effectiveness of our framework on the\nchallenging KITTI 3D object detection benchmark. We implement our framework on\nthree representative monocular detectors, and we achieve state-of-the-art\nperformance with no additional inference computational cost relative to\nbaseline models. Our code is available at https://github.com/rockywind/ADD.\n", "date": "2022-11-30", "categories": []} {"id": "2306.17450", "title": "GMM: Delving into Gradient Aware and Model Perceive Depth Mining for\n Monocular 3D Detection", "abstract": " Depth perception is a crucial component of monoc-ular 3D detection tasks that\ntypically involve ill-posed problems. In light of the success of sample mining\ntechniques in 2D object detection, we propose a simple yet effective mining\nstrategy for improving depth perception in 3D object detection. Concretely, we\nintroduce a plain metric to evaluate the quality of depth predictions, which\nchooses the mined sample for the model. Moreover, we propose a Gradient-aware\nand Model-perceive Mining strategy (GMM) for depth learning, which exploits the\npredicted depth quality for better depth learning through easy mining. GMM is a\ngeneral strategy that can be readily applied to several state-of-the-art\nmonocular 3D detectors, improving the accuracy of depth prediction. Extensive\nexperiments on the nuScenes dataset demonstrate that the proposed methods\nsignificantly improve the performance of 3D object detection while\noutperforming other state-of-the-art sample mining techniques by a considerable\nmargin. On the nuScenes benchmark, GMM achieved the state-of-the-art (42.1% mAP\nand 47.3% NDS) performance in monocular object detection.\n", "date": "2023-06-30", "categories": []} {"id": "2207.00865", "title": "ORA3D: Overlap Region Aware Multi-view 3D Object Detection", "abstract": " Current multi-view 3D object detection methods often fail to detect objects\nin the overlap region properly, and the networks' understanding of the scene is\noften limited to that of a monocular detection network. Moreover, objects in\nthe overlap region are often largely occluded or suffer from deformation due to\ncamera distortion, causing a domain shift. To mitigate this issue, we propose\nusing the following two main modules: (1) Stereo Disparity Estimation for Weak\nDepth Supervision and (2) Adversarial Overlap Region Discriminator. The former\nutilizes the traditional stereo disparity estimation method to obtain reliable\ndisparity information from the overlap region. Given the disparity estimates as\nsupervision, we propose regularizing the network to fully utilize the geometric\npotential of binocular images and improve the overall detection accuracy\naccordingly. Further, the latter module minimizes the representational gap\nbetween non-overlap and overlapping regions. We demonstrate the effectiveness\nof the proposed method with the nuScenes large-scale multi-view 3D object\ndetection data. Our experiments show that our proposed method outperforms\ncurrent state-of-the-art models, i.e., DETR3D and BEVDet.\n", "date": "2022-07-02", "categories": []} {"id": "2304.13390", "title": "Group Equivariant BEV for 3D Object Detection", "abstract": " Recently, 3D object detection has attracted significant attention and\nachieved continuous improvement in real road scenarios. The environmental\ninformation is collected from a single sensor or multi-sensor fusion to detect\ninterested objects. However, most of the current 3D object detection approaches\nfocus on developing advanced network architectures to improve the detection\nprecision of the object rather than considering the dynamic driving scenes,\nwhere data collected from sensors equipped in the vehicle contain various\nperturbation features. As a result, existing work cannot still tackle the\nperturbation issue. In order to solve this problem, we propose a group\nequivariant bird's eye view network (GeqBevNet) based on the group equivariant\ntheory, which introduces the concept of group equivariant into the BEV fusion\nobject detection network. The group equivariant network is embedded into the\nfused BEV feature map to facilitate the BEV-level rotational equivariant\nfeature extraction, thus leading to lower average orientation error. In order\nto demonstrate the effectiveness of the GeqBevNet, the network is verified on\nthe nuScenes validation dataset in which mAOE can be decreased to 0.325.\nExperimental results demonstrate that GeqBevNet can extract more rotational\nequivariant features in the 3D object detection of the actual road scene and\nimprove the performance of object orientation prediction.\n", "date": "2023-04-26", "categories": []} {"id": "2306.15919", "title": "Fine-grained 3D object recognition: an approach and experiments", "abstract": " Three-dimensional (3D) object recognition technology is being used as a core\ntechnology in advanced technologies such as autonomous driving of automobiles.\nThere are two sets of approaches for 3D object recognition: (i) hand-crafted\napproaches like Global Orthographic Object Descriptor (GOOD), and (ii) deep\nlearning-based approaches such as MobileNet and VGG. However, it is needed to\nknow which of these approaches works better in an open-ended domain where the\nnumber of known categories increases over time, and the system should learn\nabout new object categories using few training examples. In this paper, we\nfirst implemented an offline 3D object recognition system that takes an object\nview as input and generates category labels as output. In the offline stage,\ninstance-based learning (IBL) is used to form a new category and we use K-fold\ncross-validation to evaluate the obtained object recognition performance. We\nthen test the proposed approach in an online fashion by integrating the code\ninto a simulated teacher test. As a result, we concluded that the approach\nusing deep learning features is more suitable for open-ended fashion. Moreover,\nwe observed that concatenating the hand-crafted and deep learning features\nincreases the classification accuracy.\n", "date": "2023-06-28", "categories": []} {"id": "2212.05265", "title": "Multi-Sem Fusion: Multimodal Semantic Fusion for 3D Object Detection", "abstract": " LiDAR and camera fusion techniques are promising for achieving 3D object\ndetection in autonomous driving. Most multi-modal 3D object detection\nframeworks integrate semantic knowledge from 2D images into 3D LiDAR point\nclouds to enhance detection accuracy. Nevertheless, the restricted resolution\nof 2D feature maps impedes accurate re-projection and often induces a\npronounced boundary-blurring effect, which is primarily attributed to erroneous\nsemantic segmentation. To well handle this limitation, we propose a general\nmulti-modal fusion framework Multi-Sem Fusion (MSF) to fuse the semantic\ninformation from both the 2D image and 3D points scene parsing results.\nSpecifically, we employ 2D/3D semantic segmentation methods to generate the\nparsing results for 2D images and 3D point clouds. The 2D semantic information\nis further reprojected into the 3D point clouds with calibration parameters. To\nhandle the misalignment between the 2D and 3D parsing results, we propose an\nAdaptive Attention-based Fusion (AAF) module to fuse them by learning an\nadaptive fusion score. Then the point cloud with the fused semantic label is\nsent to the following 3D object detectors. Furthermore, we propose a Deep\nFeature Fusion (DFF) module to aggregate deep features at different levels to\nboost the final detection performance. The effectiveness of the framework has\nbeen verified on two public large-scale 3D object detection benchmarks by\ncomparing them with different baselines. The experimental results show that the\nproposed fusion strategies can significantly improve the detection performance\ncompared to the methods using only point clouds and the methods using only 2D\nsemantic information. Most importantly, the proposed approach significantly\noutperforms other approaches and sets state-of-the-art results on the nuScenes\ntesting benchmark.\n", "date": "2022-12-10", "categories": []} {"id": "2306.10699", "title": "Frame Fusion with Vehicle Motion Prediction for 3D Object Detection", "abstract": " In LiDAR-based 3D detection, history point clouds contain rich temporal\ninformation helpful for future prediction. In the same way, history detections\nshould contribute to future detections. In this paper, we propose a detection\nenhancement method, namely FrameFusion, which improves 3D object detection\nresults by fusing history frames. In FrameFusion, we ''forward'' history frames\nto the current frame and apply weighted Non-Maximum-Suppression on dense\nbounding boxes to obtain a fused frame with merged boxes. To ''forward''\nframes, we use vehicle motion models to estimate the future pose of the\nbounding boxes. However, the commonly used constant velocity model fails\nnaturally on turning vehicles, so we explore two vehicle motion models to\naddress this issue. On Waymo Open Dataset, our FrameFusion method consistently\nimproves the performance of various 3D detectors by about $2$ vehicle level 2\nAPH with negligible latency and slightly enhances the performance of the\ntemporal fusion method MPPNet. We also conduct extensive experiments on motion\nmodel selection.\n", "date": "2023-06-19", "categories": []} {"id": "2306.10921", "title": "Understanding Depth Map Progressively: Adaptive Distance Interval\n Separation for Monocular 3d Object Detection", "abstract": " Monocular 3D object detection aims to locate objects in different scenes with\njust a single image. Due to the absence of depth information, several monocular\n3D detection techniques have emerged that rely on auxiliary depth maps from the\ndepth estimation task. There are multiple approaches to understanding the\nrepresentation of depth maps, including treating them as pseudo-LiDAR point\nclouds, leveraging implicit end-to-end learning of depth information, or\nconsidering them as an image input. However, these methods have certain\ndrawbacks, such as their reliance on the accuracy of estimated depth maps and\nsuboptimal utilization of depth maps due to their image-based nature. While\nLiDAR-based methods and convolutional neural networks (CNNs) can be utilized\nfor pseudo point clouds and depth maps, respectively, it is always an\nalternative. In this paper, we propose a framework named the Adaptive Distance\nInterval Separation Network (ADISN) that adopts a novel perspective on\nunderstanding depth maps, as a form that lies between LiDAR and images. We\nutilize an adaptive separation approach that partitions the depth map into\nvarious subgraphs based on distance and treats each of these subgraphs as an\nindividual image for feature extraction. After adaptive separations, each\nsubgraph solely contains pixels within a learned interval range. If there is a\ntruncated object within this range, an evident curved edge will appear, which\nwe can leverage for texture extraction using CNNs to obtain rich depth\ninformation in pixels. Meanwhile, to mitigate the inaccuracy of depth\nestimation, we designed an uncertainty module. To take advantage of both images\nand depth maps, we use different branches to learn localization detection tasks\nand appearance tasks separately.\n", "date": "2023-06-19", "categories": []} {"id": "2306.07835", "title": "LMD: Light-weight Prediction Quality Estimation for Object Detection in\n Lidar Point Clouds", "abstract": " Object detection on Lidar point cloud data is a promising technology for\nautonomous driving and robotics which has seen a significant rise in\nperformance and accuracy during recent years. Particularly uncertainty\nestimation is a crucial component for down-stream tasks and deep neural\nnetworks remain error-prone even for predictions with high confidence.\nPreviously proposed methods for quantifying prediction uncertainty tend to\nalter the training scheme of the detector or rely on prediction sampling which\nresults in vastly increased inference time. In order to address these two\nissues, we propose LidarMetaDetect (LMD), a light-weight post-processing scheme\nfor prediction quality estimation. Our method can easily be added to any\npre-trained Lidar object detector without altering anything about the base\nmodel and is purely based on post-processing, therefore, only leading to a\nnegligible computational overhead. Our experiments show a significant increase\nof statistical reliability in separating true from false predictions. We\npropose and evaluate an additional application of our method leading to the\ndetection of annotation errors. Explicit samples and a conservative count of\nannotation error proposals indicates the viability of our method for\nlarge-scale datasets like KITTI and nuScenes. On the widely-used nuScenes test\ndataset, 43 out of the top 100 proposals of our method indicate, in fact,\nerroneous annotations.\n", "date": "2023-06-13", "categories": []} {"id": "2306.09266", "title": "A9 Intersection Dataset: All You Need for Urban 3D Camera-LiDAR Roadside\n Perception", "abstract": " Intelligent Transportation Systems (ITS) allow a drastic expansion of the\nvisibility range and decrease occlusions for autonomous driving. To obtain\naccurate detections, detailed labeled sensor data for training is required.\nUnfortunately, high-quality 3D labels of LiDAR point clouds from the\ninfrastructure perspective of an intersection are still rare. Therefore, we\nprovide the A9 Intersection Dataset, which consists of labeled LiDAR point\nclouds and synchronized camera images. Here, we recorded the sensor output from\ntwo roadside cameras and LiDARs mounted on intersection gantry bridges. The\npoint clouds were labeled in 3D by experienced annotators. Furthermore, we\nprovide calibration data between all sensors, which allow the projection of the\n3D labels into the camera images and an accurate data fusion. Our dataset\nconsists of 4.8k images and point clouds with more than 57.4k manually labeled\n3D boxes. With ten object classes, it has a high diversity of road users in\ncomplex driving maneuvers, such as left and right turns, overtaking, and\nU-turns. In experiments, we provided multiple baselines for the perception\ntasks. Overall, our dataset is a valuable contribution to the scientific\ncommunity to perform complex 3D camera-LiDAR roadside perception tasks. Find\ndata, code, and more information at https://a9-dataset.com.\n", "date": "2023-06-15", "categories": []} {"id": "2306.07344", "title": "Towards a Robust Sensor Fusion Step for 3D Object Detection on Corrupted\n Data", "abstract": " Multimodal sensor fusion methods for 3D object detection have been\nrevolutionizing the autonomous driving research field. Nevertheless, most of\nthese methods heavily rely on dense LiDAR data and accurately calibrated\nsensors which is often not the case in real-world scenarios. Data from LiDAR\nand cameras often come misaligned due to the miscalibration, decalibration, or\ndifferent frequencies of the sensors. Additionally, some parts of the LiDAR\ndata may be occluded and parts of the data may be missing due to hardware\nmalfunction or weather conditions. This work presents a novel fusion step that\naddresses data corruptions and makes sensor fusion for 3D object detection more\nrobust. Through extensive experiments, we demonstrate that our method performs\non par with state-of-the-art approaches on normal data and outperforms them on\nmisaligned data.\n", "date": "2023-06-12", "categories": []} {"id": "2306.05663", "title": "Improving LiDAR 3D Object Detection via Range-based Point Cloud Density\n Optimization", "abstract": " In recent years, much progress has been made in LiDAR-based 3D object\ndetection mainly due to advances in detector architecture designs and\navailability of large-scale LiDAR datasets. Existing 3D object detectors tend\nto perform well on the point cloud regions closer to the LiDAR sensor as\nopposed to on regions that are farther away. In this paper, we investigate this\nproblem from the data perspective instead of detector architecture design. We\nobserve that there is a learning bias in detection models towards the dense\nobjects near the sensor and show that the detection performance can be improved\nby simply manipulating the input point cloud density at different distance\nranges without modifying the detector architecture and without data\naugmentation. We propose a model-free point cloud density adjustment\npre-processing mechanism that uses iterative MCMC optimization to estimate\noptimal parameters for altering the point density at different distance ranges.\nWe conduct experiments using four state-of-the-art LiDAR 3D object detectors on\ntwo public LiDAR datasets, namely Waymo and ONCE. Our results demonstrate that\nour range-based point cloud density manipulation technique can improve the\nperformance of the existing detectors, which in turn could potentially inspire\nfuture detector designs.\n", "date": "2023-06-09", "categories": []} {"id": "2303.11926", "title": "Exploring Object-Centric Temporal Modeling for Efficient Multi-View 3D\n Object Detection", "abstract": " In this paper, we propose a long-sequence modeling framework, named\nStreamPETR, for multi-view 3D object detection. Built upon the sparse query\ndesign in the PETR series, we systematically develop an object-centric temporal\nmechanism. The model is performed in an online manner and the long-term\nhistorical information is propagated through object queries frame by frame.\nBesides, we introduce a motion-aware layer normalization to model the movement\nof the objects. StreamPETR achieves significant performance improvements only\nwith negligible computation cost, compared to the single-frame baseline. On the\nstandard nuScenes benchmark, it is the first online multi-view method that\nachieves comparable performance (67.6% NDS & 65.3% AMOTA) with lidar-based\nmethods. The lightweight version realizes 45.0% mAP and 31.7 FPS, outperforming\nthe state-of-the-art method (SOLOFusion) by 2.3% mAP and 1.8x faster FPS. Code\nhas been available at https://github.com/exiawsh/StreamPETR.git.\n", "date": "2023-03-21", "categories": []} {"id": "2306.03206", "title": "MoDAR: Using Motion Forecasting for 3D Object Detection in Point Cloud\n Sequences", "abstract": " Occluded and long-range objects are ubiquitous and challenging for 3D object\ndetection. Point cloud sequence data provide unique opportunities to improve\nsuch cases, as an occluded or distant object can be observed from different\nviewpoints or gets better visibility over time. However, the efficiency and\neffectiveness in encoding long-term sequence data can still be improved. In\nthis work, we propose MoDAR, using motion forecasting outputs as a type of\nvirtual modality, to augment LiDAR point clouds. The MoDAR modality propagates\nobject information from temporal contexts to a target frame, represented as a\nset of virtual points, one for each object from a waypoint on a forecasted\ntrajectory. A fused point cloud of both raw sensor points and the virtual\npoints can then be fed to any off-the-shelf point-cloud based 3D object\ndetector. Evaluated on the Waymo Open Dataset, our method significantly\nimproves prior art detectors by using motion forecasting from extra-long\nsequences (e.g. 18 seconds), achieving new state of the arts, while not adding\nmuch computation overhead.\n", "date": "2023-06-05", "categories": []} {"id": "2302.13002", "title": "Introducing Depth into Transformer-based 3D Object Detection", "abstract": " In this paper, we present DAT, a Depth-Aware Transformer framework designed\nfor camera-based 3D detection. Our model is based on observing two major issues\nin existing methods: large depth translation errors and duplicate predictions\nalong depth axes. To mitigate these issues, we propose two key solutions within\nDAT. To address the first issue, we introduce a Depth-Aware Spatial\nCross-Attention (DA-SCA) module that incorporates depth information into\nspatial cross-attention when lifting image features to 3D space. To address the\nsecond issue, we introduce an auxiliary learning task called Depth-aware\nNegative Suppression loss. First, based on their reference points, we organize\nfeatures as a Bird's-Eye-View (BEV) feature map. Then, we sample positive and\nnegative features along each object ray that connects an object and a camera\nand train the model to distinguish between them. The proposed DA-SCA and DNS\nmethods effectively alleviate these two problems. We show that DAT is a\nversatile method that enhances the performance of all three popular models,\nBEVFormer, DETR3D, and PETR. Our evaluation on BEVFormer demonstrates that DAT\nachieves a significant improvement of +2.8 NDS on nuScenes val under the same\nsettings. Moreover, when using pre-trained VoVNet-99 as the backbone, DAT\nachieves strong results of 60.0 NDS and 51.5 mAP on nuScenes test. Our code\nwill be soon.\n", "date": "2023-02-25", "categories": []} {"id": "2304.14446", "title": "HyperMODEST: Self-Supervised 3D Object Detection with Confidence Score\n Filtering", "abstract": " Current LiDAR-based 3D object detectors for autonomous driving are almost\nentirely trained on human-annotated data collected in specific geographical\ndomains with specific sensor setups, making it difficult to adapt to a\ndifferent domain. MODEST is the first work to train 3D object detectors without\nany labels. Our work, HyperMODEST, proposes a universal method implemented on\ntop of MODEST that can largely accelerate the self-training process and does\nnot require tuning on a specific dataset. We filter intermediate pseudo-labels\nused for data augmentation with low confidence scores. On the nuScenes dataset,\nwe observe a significant improvement of 1.6% in AP BEV in 0-80m range at\nIoU=0.25 and an improvement of 1.7% in AP BEV in 0-80m range at IoU=0.5 while\nonly using one-fifth of the training time in the original approach by MODEST.\nOn the Lyft dataset, we also observe an improvement over the baseline during\nthe first round of iterative self-training. We explore the trade-off between\nhigh precision and high recall in the early stage of the self-training process\nby comparing our proposed method with two other score filtering methods:\nconfidence score filtering for pseudo-labels with and without static label\nretention. The code and models of this work are available at\nhttps://github.com/TRAILab/HyperMODEST\n", "date": "2023-04-27", "categories": []} {"id": "2306.01438", "title": "Bi-LRFusion: Bi-Directional LiDAR-Radar Fusion for 3D Dynamic Object\n Detection", "abstract": " LiDAR and Radar are two complementary sensing approaches in that LiDAR\nspecializes in capturing an object's 3D shape while Radar provides longer\ndetection ranges as well as velocity hints. Though seemingly natural, how to\nefficiently combine them for improved feature representation is still unclear.\nThe main challenge arises from that Radar data are extremely sparse and lack\nheight information. Therefore, directly integrating Radar features into\nLiDAR-centric detection networks is not optimal. In this work, we introduce a\nbi-directional LiDAR-Radar fusion framework, termed Bi-LRFusion, to tackle the\nchallenges and improve 3D detection for dynamic objects. Technically,\nBi-LRFusion involves two steps: first, it enriches Radar's local features by\nlearning important details from the LiDAR branch to alleviate the problems\ncaused by the absence of height information and extreme sparsity; second, it\ncombines LiDAR features with the enhanced Radar features in a unified\nbird's-eye-view representation. We conduct extensive experiments on nuScenes\nand ORR datasets, and show that our Bi-LRFusion achieves state-of-the-art\nperformance for detecting dynamic objects. Notably, Radar data in these two\ndatasets have different formats, which demonstrates the generalizability of our\nmethod. Codes are available at https://github.com/JessieW0806/BiLRFusion.\n", "date": "2023-06-02", "categories": []} {"id": "2306.01738", "title": "OCBEV: Object-Centric BEV Transformer for Multi-View 3D Object Detection", "abstract": " Multi-view 3D object detection is becoming popular in autonomous driving due\nto its high effectiveness and low cost. Most of the current state-of-the-art\ndetectors follow the query-based bird's-eye-view (BEV) paradigm, which benefits\nfrom both BEV's strong perception power and end-to-end pipeline. Despite\nachieving substantial progress, existing works model objects via globally\nleveraging temporal and spatial information of BEV features, resulting in\nproblems when handling the challenging complex and dynamic autonomous driving\nscenarios. In this paper, we proposed an Object-Centric query-BEV detector\nOCBEV, which can carve the temporal and spatial cues of moving targets more\neffectively. OCBEV comprises three designs: Object Aligned Temporal Fusion\naligns the BEV feature based on ego-motion and estimated current locations of\nmoving objects, leading to a precise instance-level feature fusion. Object\nFocused Multi-View Sampling samples more 3D features from an adaptive local\nheight ranges of objects for each scene to enrich foreground information.\nObject Informed Query Enhancement replaces part of pre-defined decoder queries\nin common DETR-style decoders with positional features of objects on\nhigh-confidence locations, introducing more direct object positional priors.\nExtensive experimental evaluations are conducted on the challenging nuScenes\ndataset. Our approach achieves a state-of-the-art result, surpassing the\ntraditional BEVFormer by 1.5 NDS points. Moreover, we have a faster convergence\nspeed and only need half of the training iterations to get comparable\nperformance, which further demonstrates its effectiveness.\n", "date": "2023-06-02", "categories": []} {"id": "2305.17318", "title": "Radar Enlighten the Dark: Enhancing Low-Visibility Perception for\n Automated Vehicles with Camera-Radar Fusion", "abstract": " Sensor fusion is a crucial augmentation technique for improving the accuracy\nand reliability of perception systems for automated vehicles under diverse\ndriving conditions. However, adverse weather and low-light conditions remain\nchallenging, where sensor performance degrades significantly, exposing vehicle\nsafety to potential risks. Advanced sensors such as LiDARs can help mitigate\nthe issue but with extremely high marginal costs. In this paper, we propose a\nnovel transformer-based 3D object detection model \"REDFormer\" to tackle low\nvisibility conditions, exploiting the power of a more practical and\ncost-effective solution by leveraging bird's-eye-view camera-radar fusion.\nUsing the nuScenes dataset with multi-radar point clouds, weather information,\nand time-of-day data, our model outperforms state-of-the-art (SOTA) models on\nclassification and detection accuracy. Finally, we provide extensive ablation\nstudies of each model component on their contributions to address the\nabove-mentioned challenges. Particularly, it is shown in the experiments that\nour model achieves a significant performance improvement over the baseline\nmodel in low-visibility scenarios, specifically exhibiting a 31.31% increase in\nrainy scenes and a 46.99% enhancement in nighttime scenes.The source code of\nthis study is publicly available.\n", "date": "2023-05-27", "categories": []} {"id": "2305.17763", "title": "NeurOCS: Neural NOCS Supervision for Monocular 3D Object Localization", "abstract": " Monocular 3D object localization in driving scenes is a crucial task, but\nchallenging due to its ill-posed nature. Estimating 3D coordinates for each\npixel on the object surface holds great potential as it provides dense 2D-3D\ngeometric constraints for the underlying PnP problem. However, high-quality\nground truth supervision is not available in driving scenes due to sparsity and\nvarious artifacts of Lidar data, as well as the practical infeasibility of\ncollecting per-instance CAD models. In this work, we present NeurOCS, a\nframework that uses instance masks and 3D boxes as input to learn 3D object\nshapes by means of differentiable rendering, which further serves as\nsupervision for learning dense object coordinates. Our approach rests on\ninsights in learning a category-level shape prior directly from real driving\nscenes, while properly handling single-view ambiguities. Furthermore, we study\nand make critical design choices to learn object coordinates more effectively\nfrom an object-centric view. Altogether, our framework leads to new\nstate-of-the-art in monocular 3D localization that ranks 1st on the\nKITTI-Object benchmark among published monocular methods.\n", "date": "2023-05-28", "categories": []} {"id": "2305.17972", "title": "View-to-Label: Multi-View Consistency for Self-Supervised 3D Object\n Detection", "abstract": " For autonomous vehicles, driving safely is highly dependent on the capability\nto correctly perceive the environment in 3D space, hence the task of 3D object\ndetection represents a fundamental aspect of perception. While 3D sensors\ndeliver accurate metric perception, monocular approaches enjoy cost and\navailability advantages that are valuable in a wide range of applications.\nUnfortunately, training monocular methods requires a vast amount of annotated\ndata. Interestingly, self-supervised approaches have recently been successfully\napplied to ease the training process and unlock access to widely available\nunlabelled data. While related research leverages different priors including\nLIDAR scans and stereo images, such priors again limit usability. Therefore, in\nthis work, we propose a novel approach to self-supervise 3D object detection\npurely from RGB sequences alone, leveraging multi-view constraints and weak\nlabels. Our experiments on KITTI 3D dataset demonstrate performance on par with\nstate-of-the-art self-supervised methods using LIDAR scans or stereo images.\n", "date": "2023-05-29", "categories": []} {"id": "2305.10974", "title": "MonoTDP: Twin Depth Perception for Monocular 3D Object Detection in\n Adverse Scenes", "abstract": " 3D object detection plays a crucial role in numerous intelligent vision\nsystems. Detection in the open world inevitably encounters various adverse\nscenes, such as dense fog, heavy rain, and low light conditions. Although\nexisting efforts primarily focus on diversifying network architecture or\ntraining schemes, resulting in significant progress in 3D object detection,\nmost of these learnable modules fail in adverse scenes, thereby hindering\ndetection performance. To address this issue, this paper proposes a monocular\n3D detection model designed to perceive twin depth in adverse scenes, termed\nMonoTDP, which effectively mitigates the degradation of detection performance\nin various harsh environments. Specifically, we first introduce an adaptive\nlearning strategy to aid the model in handling uncontrollable weather\nconditions, significantly resisting degradation caused by various degrading\nfactors. Then, to address the depth/content loss in adverse regions, we propose\na novel twin depth perception module that simultaneously estimates scene and\nobject depth, enabling the integration of scene-level features and object-level\nfeatures. Additionally, we assemble a new adverse 3D object detection dataset\nencompassing a wide range of challenging scenes, including rainy, foggy, and\nlow light weather conditions, with each type of scene containing 7,481 images.\nExperimental results demonstrate that our proposed method outperforms current\nstate-of-the-art approaches by an average of 3.12% in terms of AP_R40 for car\ncategory across various adverse environments.\n", "date": "2023-05-18", "categories": []} {"id": "2305.15694", "title": "Learning Occupancy for Monocular 3D Object Detection", "abstract": " Monocular 3D detection is a challenging task due to the lack of accurate 3D\ninformation. Existing approaches typically rely on geometry constraints and\ndense depth estimates to facilitate the learning, but often fail to fully\nexploit the benefits of three-dimensional feature extraction in frustum and 3D\nspace. In this paper, we propose \\textbf{OccupancyM3D}, a method of learning\noccupancy for monocular 3D detection. It directly learns occupancy in frustum\nand 3D space, leading to more discriminative and informative 3D features and\nrepresentations. Specifically, by using synchronized raw sparse LiDAR point\nclouds, we define the space status and generate voxel-based occupancy labels.\nWe formulate occupancy prediction as a simple classification problem and design\nassociated occupancy losses. Resulting occupancy estimates are employed to\nenhance original frustum/3D features. As a result, experiments on KITTI and\nWaymo open datasets demonstrate that the proposed method achieves a new state\nof the art and surpasses other methods by a significant margin. Codes and\npre-trained models will be available at\n\\url{https://github.com/SPengLiang/OccupancyM3D}.\n", "date": "2023-05-25", "categories": []} {"id": "2305.15765", "title": "Language-Guided 3D Object Detection in Point Cloud for Autonomous\n Driving", "abstract": " This paper addresses the problem of 3D referring expression comprehension\n(REC) in autonomous driving scenario, which aims to ground a natural language\nto the targeted region in LiDAR point clouds. Previous approaches for REC\nusually focus on the 2D or 3D-indoor domain, which is not suitable for\naccurately predicting the location of the queried 3D region in an autonomous\ndriving scene. In addition, the upper-bound limitation and the heavy\ncomputation cost motivate us to explore a better solution. In this work, we\npropose a new multi-modal visual grounding task, termed LiDAR Grounding. Then\nwe devise a Multi-modal Single Shot Grounding (MSSG) approach with an effective\ntoken fusion strategy. It jointly learns the LiDAR-based object detector with\nthe language features and predicts the targeted region directly from the\ndetector without any post-processing. Moreover, the image feature can be\nflexibly integrated into our approach to provide rich texture and color\ninformation. The cross-modal learning enforces the detector to concentrate on\nimportant regions in the point cloud by considering the informative language\nexpressions, thus leading to much better accuracy and efficiency. Extensive\nexperiments on the Talk2Car dataset demonstrate the effectiveness of the\nproposed methods. Our work offers a deeper insight into the LiDAR-based\ngrounding task and we expect it presents a promising direction for the\nautonomous driving community.\n", "date": "2023-05-25", "categories": []} {"id": "2305.14018", "title": "Sparse4D v2: Recurrent Temporal Fusion with Sparse Model", "abstract": " Sparse algorithms offer great flexibility for multi-view temporal perception\ntasks. In this paper, we present an enhanced version of Sparse4D, in which we\nimprove the temporal fusion module by implementing a recursive form of\nmulti-frame feature sampling. By effectively decoupling image features and\nstructured anchor features, Sparse4D enables a highly efficient transformation\nof temporal features, thereby facilitating temporal fusion solely through the\nframe-by-frame transmission of sparse features. The recurrent temporal fusion\napproach provides two main benefits. Firstly, it reduces the computational\ncomplexity of temporal fusion from $O(T)$ to $O(1)$, resulting in significant\nimprovements in inference speed and memory usage. Secondly, it enables the\nfusion of long-term information, leading to more pronounced performance\nimprovements due to temporal fusion. Our proposed approach, Sparse4Dv2, further\nenhances the performance of the sparse perception algorithm and achieves\nstate-of-the-art results on the nuScenes 3D detection benchmark. Code will be\navailable at \\url{https://github.com/linxuewu/Sparse4D}.\n", "date": "2023-05-23", "categories": []} {"id": "2305.15219", "title": "DynStatF: An Efficient Feature Fusion Strategy for LiDAR 3D Object\n Detection", "abstract": " Augmenting LiDAR input with multiple previous frames provides richer semantic\ninformation and thus boosts performance in 3D object detection, However,\ncrowded point clouds in multi-frames can hurt the precise position information\ndue to the motion blur and inaccurate point projection. In this work, we\npropose a novel feature fusion strategy, DynStaF (Dynamic-Static Fusion), which\nenhances the rich semantic information provided by the multi-frame (dynamic\nbranch) with the accurate location information from the current single-frame\n(static branch). To effectively extract and aggregate complimentary features,\nDynStaF contains two modules, Neighborhood Cross Attention (NCA) and\nDynamic-Static Interaction (DSI), operating through a dual pathway\narchitecture. NCA takes the features in the static branch as queries and the\nfeatures in the dynamic branch as keys (values). When computing the attention,\nwe address the sparsity of point clouds and take only neighborhood positions\ninto consideration. NCA fuses two features at different feature map scales,\nfollowed by DSI providing the comprehensive interaction. To analyze our\nproposed strategy DynStaF, we conduct extensive experiments on the nuScenes\ndataset. On the test set, DynStaF increases the performance of PointPillars in\nNDS by a large margin from 57.7% to 61.6%. When combined with CenterPoint, our\nframework achieves 61.0% mAP and 67.7% NDS, leading to state-of-the-art\nperformance without bells and whistles.\n", "date": "2023-05-24", "categories": []} {"id": "2211.08691", "title": "Towards Long-Tailed 3D Detection", "abstract": " Contemporary autonomous vehicle (AV) benchmarks have advanced techniques for\ntraining 3D detectors, particularly on large-scale lidar data. Surprisingly,\nalthough semantic class labels naturally follow a long-tailed distribution,\ncontemporary benchmarks focus on only a few common classes (e.g., pedestrian\nand car) and neglect many rare classes in-the-tail (e.g., debris and stroller).\nHowever, AVs must still detect rare classes to ensure safe operation. Moreover,\nsemantic classes are often organized within a hierarchy, e.g., tail classes\nsuch as child and construction-worker are arguably subclasses of pedestrian.\nHowever, such hierarchical relationships are often ignored, which may lead to\nmisleading estimates of performance and missed opportunities for algorithmic\ninnovation. We address these challenges by formally studying the problem of\nLong-Tailed 3D Detection (LT3D), which evaluates on all classes, including\nthose in-the-tail. We evaluate and innovate upon popular 3D detection\ncodebases, such as CenterPoint and PointPillars, adapting them for LT3D. We\ndevelop hierarchical losses that promote feature sharing across common-vs-rare\nclasses, as well as improved detection metrics that award partial credit to\n\"reasonable\" mistakes respecting the hierarchy (e.g., mistaking a child for an\nadult). Finally, we point out that fine-grained tail class accuracy is\nparticularly improved via multimodal fusion of RGB images with LiDAR; simply\nput, small fine-grained classes are challenging to identify from sparse (lidar)\ngeometry alone, suggesting that multimodal cues are crucial to long-tailed 3D\ndetection. Our modifications improve accuracy by 5% AP on average for all\nclasses, and dramatically improve AP for rare classes (e.g., stroller AP\nimproves from 3.6 to 31.6)! Our code is available at\nhttps://github.com/neeharperi/LT3D\n", "date": "2022-11-16", "categories": []} {"id": "2305.12853", "title": "Real-Aug: Realistic Scene Synthesis for LiDAR Augmentation in 3D Object\n Detection", "abstract": " Data and model are the undoubtable two supporting pillars for LiDAR object\ndetection. However, data-centric works have fallen far behind compared with the\never-growing list of fancy new models. In this work, we systematically study\nthe synthesis-based LiDAR data augmentation approach (so-called GT-Aug) which\noffers maxium controllability over generated data samples. We pinpoint the main\nshortcoming of existing works is introducing unrealistic LiDAR scan patterns\nduring GT-Aug. In light of this finding, we propose Real-Aug, a synthesis-based\naugmentation method which prioritizes on generating realistic LiDAR scans. Our\nmethod consists a reality-conforming scene composition module which handles the\ndetails of the composition and a real-synthesis mixing up training strategy\nwhich gradually adapts the data distribution from synthetic data to the real\none. To verify the effectiveness of our methods, we conduct extensive ablation\nstudies and validate the proposed Real-Aug on a wide combination of detectors\nand datasets. We achieve a state-of-the-art 0.744 NDS and 0.702 mAP on nuScenes\ntest set. The code shall be released soon.\n", "date": "2023-05-22", "categories": []} {"id": "2304.00788", "title": "Open-Vocabulary Point-Cloud Object Detection without 3D Annotation", "abstract": " The goal of open-vocabulary detection is to identify novel objects based on\narbitrary textual descriptions. In this paper, we address open-vocabulary 3D\npoint-cloud detection by a dividing-and-conquering strategy, which involves: 1)\ndeveloping a point-cloud detector that can learn a general representation for\nlocalizing various objects, and 2) connecting textual and point-cloud\nrepresentations to enable the detector to classify novel object categories\nbased on text prompting. Specifically, we resort to rich image pre-trained\nmodels, by which the point-cloud detector learns localizing objects under the\nsupervision of predicted 2D bounding boxes from 2D pre-trained detectors.\nMoreover, we propose a novel de-biased triplet cross-modal contrastive learning\nto connect the modalities of image, point-cloud and text, thereby enabling the\npoint-cloud detector to benefit from vision-language pre-trained\nmodels,i.e.,CLIP. The novel use of image and vision-language pre-trained models\nfor point-cloud detectors allows for open-vocabulary 3D object detection\nwithout the need for 3D annotations. Experiments demonstrate that the proposed\nmethod improves at least 3.03 points and 7.47 points over a wide range of\nbaselines on the ScanNet and SUN RGB-D datasets, respectively. Furthermore, we\nprovide a comprehensive analysis to explain why our approach works.\n", "date": "2023-04-03", "categories": []} {"id": "2305.07713", "title": "Multi-Modal 3D Object Detection by Box Matching", "abstract": " Multi-modal 3D object detection has received growing attention as the\ninformation from different sensors like LiDAR and cameras are complementary.\nMost fusion methods for 3D detection rely on an accurate alignment and\ncalibration between 3D point clouds and RGB images. However, such an assumption\nis not reliable in a real-world self-driving system, as the alignment between\ndifferent modalities is easily affected by asynchronous sensors and disturbed\nsensor placement. We propose a novel {F}usion network by {B}ox {M}atching\n(FBMNet) for multi-modal 3D detection, which provides an alternative way for\ncross-modal feature alignment by learning the correspondence at the bounding\nbox level to free up the dependency of calibration during inference. With the\nlearned assignments between 3D and 2D object proposals, the fusion for\ndetection can be effectively performed by combing their ROI features. Extensive\nexperiments on the nuScenes dataset demonstrate that our method is much more\nstable in dealing with challenging cases such as asynchronous sensors,\nmisaligned sensor placement, and degenerated camera images than existing fusion\nmethods. We hope that our FBMNet could provide an available solution to dealing\nwith these challenging cases for safety in real autonomous driving scenarios.\nCodes will be publicly available at https://github.com/happinesslz/FBMNet.\n", "date": "2023-05-12", "categories": []} {"id": "2305.06621", "title": "PVT-SSD: Single-Stage 3D Object Detector with Point-Voxel Transformer", "abstract": " Recent Transformer-based 3D object detectors learn point cloud features\neither from point- or voxel-based representations. However, the former requires\ntime-consuming sampling while the latter introduces quantization errors. In\nthis paper, we present a novel Point-Voxel Transformer for single-stage 3D\ndetection (PVT-SSD) that takes advantage of these two representations.\nSpecifically, we first use voxel-based sparse convolutions for efficient\nfeature encoding. Then, we propose a Point-Voxel Transformer (PVT) module that\nobtains long-range contexts in a cheap manner from voxels while attaining\naccurate positions from points. The key to associating the two different\nrepresentations is our introduced input-dependent Query Initialization module,\nwhich could efficiently generate reference points and content queries. Then,\nPVT adaptively fuses long-range contextual and local geometric information\naround reference points into content queries. Further, to quickly find the\nneighboring points of reference points, we design the Virtual Range Image\nmodule, which generalizes the native range image to multi-sensor and\nmulti-frame. The experiments on several autonomous driving benchmarks verify\nthe effectiveness and efficiency of the proposed method. Code will be available\nat https://github.com/Nightmare-n/PVT-SSD.\n", "date": "2023-05-11", "categories": []} {"id": "2304.02431", "title": "MS3D: Leveraging Multiple Detectors for Unsupervised Domain Adaptation\n in 3D Object Detection", "abstract": " We introduce Multi-Source 3D (MS3D), a new self-training pipeline for\nunsupervised domain adaptation in 3D object detection. Despite the remarkable\naccuracy of 3D detectors, they often overfit to specific domain biases, leading\nto suboptimal performance in various sensor setups and environments. Existing\nmethods typically focus on adapting a single detector to the target domain,\noverlooking the fact that different detectors possess distinct expertise on\ndifferent unseen domains. MS3D leverages this by combining different\npre-trained detectors from multiple source domains and incorporating temporal\ninformation to produce high-quality pseudo-labels for fine-tuning. Our proposed\nKernel-Density Estimation (KDE) Box Fusion method fuses box proposals from\nmultiple domains to obtain pseudo-labels that surpass the performance of the\nbest source domain detectors. MS3D exhibits greater robustness to domain shift\nand produces accurate pseudo-labels over greater distances, making it\nwell-suited for high-to-low beam domain adaptation and vice versa. Our method\nachieved state-of-the-art performance on all evaluated datasets, and we\ndemonstrate that the pre-trained detector's source dataset has minimal impact\non the fine-tuned result, making MS3D suitable for real-world applications.\n", "date": "2023-04-05", "categories": []} {"id": "2305.04925", "title": "PillarNeXt: Rethinking Network Designs for 3D Object Detection in LiDAR\n Point Clouds", "abstract": " In order to deal with the sparse and unstructured raw point clouds, LiDAR\nbased 3D object detection research mostly focuses on designing dedicated local\npoint aggregators for fine-grained geometrical modeling. In this paper, we\nrevisit the local point aggregators from the perspective of allocating\ncomputational resources. We find that the simplest pillar based models perform\nsurprisingly well considering both accuracy and latency. Additionally, we show\nthat minimal adaptions from the success of 2D object detection, such as\nenlarging receptive field, significantly boost the performance. Extensive\nexperiments reveal that our pillar based networks with modernized designs in\nterms of architecture and training render the state-of-the-art performance on\nthe two popular benchmarks: Waymo Open Dataset and nuScenes. Our results\nchallenge the common intuition that the detailed geometry modeling is essential\nto achieve high performance for 3D object detection.\n", "date": "2023-05-08", "categories": []} {"id": "2305.02909", "title": "Aligning Bird-Eye View Representation of Point Cloud Sequences using\n Scene Flow", "abstract": " Low-resolution point clouds are challenging for object detection methods due\nto their sparsity. Densifying the present point cloud by concatenating it with\nits predecessors is a popular solution to this challenge. Such concatenation is\npossible thanks to the removal of ego vehicle motion using its odometry. This\nmethod is called Ego Motion Compensation (EMC). Thanks to the added points, EMC\nsignificantly improves the performance of single-frame detectors. However, it\nsuffers from the shadow effect that manifests in dynamic objects' points\nscattering along their trajectories. This effect results in a misalignment\nbetween feature maps and objects' locations, thus limiting performance\nimprovement to stationary and slow-moving objects only. Scene flow allows\naligning point clouds in 3D space, thus naturally resolving the misalignment in\nfeature spaces. By observing that scene flow computation shares several\ncomponents with 3D object detection pipelines, we develop a plug-in module that\nenables single-frame detectors to compute scene flow to rectify their Bird-Eye\nView representation. Experiments on the NuScenes dataset show that our module\nleads to a significant increase (up to 16%) in the Average Precision of large\nvehicles, which interestingly demonstrates the most severe shadow effect. The\ncode is published at https://github.com/quan-dao/pc-corrector.\n", "date": "2023-05-04", "categories": []} {"id": "2305.00314", "title": "InfraDet3D: Multi-Modal 3D Object Detection based on Roadside\n Infrastructure Camera and LiDAR Sensors", "abstract": " Current multi-modal object detection approaches focus on the vehicle domain\nand are limited in the perception range and the processing capabilities.\nRoadside sensor units (RSUs) introduce a new domain for perception systems and\nleverage altitude to observe traffic. Cameras and LiDARs mounted on gantry\nbridges increase the perception range and produce a full digital twin of the\ntraffic. In this work, we introduce InfraDet3D, a multi-modal 3D object\ndetector for roadside infrastructure sensors. We fuse two LiDARs using early\nfusion and further incorporate detections from monocular cameras to increase\nthe robustness and to detect small objects. Our monocular 3D detection module\nuses HD maps to ground object yaw hypotheses, improving the final perception\nresults. The perception framework is deployed on a real-world intersection that\nis part of the A9 Test Stretch in Munich, Germany. We perform several ablation\nstudies and experiments and show that fusing two LiDARs with two cameras leads\nto an improvement of +1.90 mAP compared to a camera-only solution. We evaluate\nour results on the A9 infrastructure dataset and achieve 68.48 mAP on the test\nset. The dataset and code will be available at https://a9-dataset.com to allow\nthe research community to further improve the perception results and make\nautonomous driving safer.\n", "date": "2023-04-29", "categories": []} {"id": "2305.00397", "title": "TransCAR: Transformer-based Camera-And-Radar Fusion for 3D Object\n Detection", "abstract": " Despite radar's popularity in the automotive industry, for fusion-based 3D\nobject detection, most existing works focus on LiDAR and camera fusion. In this\npaper, we propose TransCAR, a Transformer-based Camera-And-Radar fusion\nsolution for 3D object detection. Our TransCAR consists of two modules. The\nfirst module learns 2D features from surround-view camera images and then uses\na sparse set of 3D object queries to index into these 2D features. The\nvision-updated queries then interact with each other via transformer\nself-attention layer. The second module learns radar features from multiple\nradar scans and then applies transformer decoder to learn the interactions\nbetween radar features and vision-updated queries. The cross-attention layer\nwithin the transformer decoder can adaptively learn the soft-association\nbetween the radar features and vision-updated queries instead of\nhard-association based on sensor calibration only. Finally, our model estimates\na bounding box per query using set-to-set Hungarian loss, which enables the\nmethod to avoid non-maximum suppression. TransCAR improves the velocity\nestimation using the radar scans without temporal information. The superior\nexperimental results of our TransCAR on the challenging nuScenes datasets\nillustrate that our TransCAR outperforms state-of-the-art Camera-Radar\nfusion-based 3D object detection approaches.\n", "date": "2023-04-30", "categories": []} {"id": "2303.06880", "title": "Uni3D: A Unified Baseline for Multi-dataset 3D Object Detection", "abstract": " Current 3D object detection models follow a single dataset-specific training\nand testing paradigm, which often faces a serious detection accuracy drop when\nthey are directly deployed in another dataset. In this paper, we study the task\nof training a unified 3D detector from multiple datasets. We observe that this\nappears to be a challenging task, which is mainly due to that these datasets\npresent substantial data-level differences and taxonomy-level variations caused\nby different LiDAR types and data acquisition standards. Inspired by such\nobservation, we present a Uni3D which leverages a simple data-level correction\noperation and a designed semantic-level coupling-and-recoupling module to\nalleviate the unavoidable data-level and taxonomy-level differences,\nrespectively. Our method is simple and easily combined with many 3D object\ndetection baselines such as PV-RCNN and Voxel-RCNN, enabling them to\neffectively learn from multiple off-the-shelf 3D datasets to obtain more\ndiscriminative and generalizable representations. Experiments are conducted on\nmany dataset consolidation settings including Waymo-nuScenes, nuScenes-KITTI,\nWaymo-KITTI, and Waymo-nuScenes-KITTI consolidations. Their results demonstrate\nthat Uni3D exceeds a series of individual detectors trained on a single\ndataset, with a 1.04x parameter increase over a selected baseline detector. We\nexpect this work will inspire the research of 3D generalization since it will\npush the limits of perceptual performance.\n", "date": "2023-03-13", "categories": []} {"id": "2209.12880", "title": "Center Feature Fusion: Selective Multi-Sensor Fusion of Center-based\n Objects", "abstract": " Leveraging multi-modal fusion, especially between camera and LiDAR, has\nbecome essential for building accurate and robust 3D object detection systems\nfor autonomous vehicles. Until recently, point decorating approaches, in which\npoint clouds are augmented with camera features, have been the dominant\napproach in the field. However, these approaches fail to utilize the higher\nresolution images from cameras. Recent works projecting camera features to the\nbird's-eye-view (BEV) space for fusion have also been proposed, however they\nrequire projecting millions of pixels, most of which only contain background\ninformation. In this work, we propose a novel approach Center Feature Fusion\n(CFF), in which we leverage center-based detection networks in both the camera\nand LiDAR streams to identify relevant object locations. We then use the\ncenter-based detection to identify the locations of pixel features relevant to\nobject locations, a small fraction of the total number in the image. These are\nthen projected and fused in the BEV frame. On the nuScenes dataset, we\noutperform the LiDAR-only baseline by 4.9% mAP while fusing up to 100x fewer\nfeatures than other fusion methods.\n", "date": "2022-09-26", "categories": []} {"id": "2304.14340", "title": "SparseFusion: Fusing Multi-Modal Sparse Representations for Multi-Sensor\n 3D Object Detection", "abstract": " By identifying four important components of existing LiDAR-camera 3D object\ndetection methods (LiDAR and camera candidates, transformation, and fusion\noutputs), we observe that all existing methods either find dense candidates or\nyield dense representations of scenes. However, given that objects occupy only\na small part of a scene, finding dense candidates and generating dense\nrepresentations is noisy and inefficient. We propose SparseFusion, a novel\nmulti-sensor 3D detection method that exclusively uses sparse candidates and\nsparse representations. Specifically, SparseFusion utilizes the outputs of\nparallel detectors in the LiDAR and camera modalities as sparse candidates for\nfusion. We transform the camera candidates into the LiDAR coordinate space by\ndisentangling the object representations. Then, we can fuse the multi-modality\ncandidates in a unified 3D space by a lightweight self-attention module. To\nmitigate negative transfer between modalities, we propose novel semantic and\ngeometric cross-modality transfer modules that are applied prior to the\nmodality-specific detectors. SparseFusion achieves state-of-the-art performance\non the nuScenes benchmark while also running at the fastest speed, even\noutperforming methods with stronger backbones. We perform extensive experiments\nto demonstrate the effectiveness and efficiency of our modules and overall\nmethod pipeline. Our code will be made publicly available at\nhttps://github.com/yichen928/SparseFusion.\n", "date": "2023-04-27", "categories": []} {"id": "2304.12315", "title": "Once Detected, Never Lost: Surpassing Human Performance in Offline LiDAR\n based 3D Object Detection", "abstract": " This paper aims for high-performance offline LiDAR-based 3D object detection.\nWe first observe that experienced human annotators annotate objects from a\ntrack-centric perspective. They first label the objects with clear shapes in a\ntrack, and then leverage the temporal coherence to infer the annotations of\nobscure objects. Drawing inspiration from this, we propose a high-performance\noffline detector in a track-centric perspective instead of the conventional\nobject-centric perspective. Our method features a bidirectional tracking module\nand a track-centric learning module. Such a design allows our detector to infer\nand refine a complete track once the object is detected at a certain moment. We\nrefer to this characteristic as \"onCe detecTed, neveR Lost\" and name the\nproposed system CTRL. Extensive experiments demonstrate the remarkable\nperformance of our method, surpassing the human-level annotating accuracy and\nthe previous state-of-the-art methods in the highly competitive Waymo Open\nDataset without model ensemble. The code will be made publicly available at\nhttps://github.com/tusen-ai/SST.\n", "date": "2023-04-24", "categories": []} {"id": "2304.10628", "title": "HM-ViT: Hetero-modal Vehicle-to-Vehicle Cooperative perception with\n vision transformer", "abstract": " Vehicle-to-Vehicle technologies have enabled autonomous vehicles to share\ninformation to see through occlusions, greatly enhancing perception\nperformance. Nevertheless, existing works all focused on homogeneous traffic\nwhere vehicles are equipped with the same type of sensors, which significantly\nhampers the scale of collaboration and benefit of cross-modality interactions.\nIn this paper, we investigate the multi-agent hetero-modal cooperative\nperception problem where agents may have distinct sensor modalities. We present\nHM-ViT, the first unified multi-agent hetero-modal cooperative perception\nframework that can collaboratively predict 3D objects for highly dynamic\nvehicle-to-vehicle (V2V) collaborations with varying numbers and types of\nagents. To effectively fuse features from multi-view images and LiDAR point\nclouds, we design a novel heterogeneous 3D graph transformer to jointly reason\ninter-agent and intra-agent interactions. The extensive experiments on the V2V\nperception dataset OPV2V demonstrate that the HM-ViT outperforms SOTA\ncooperative perception methods for V2V hetero-modal cooperative perception. We\nwill release codes to facilitate future research.\n", "date": "2023-04-20", "categories": []} {"id": "2303.16628", "title": "DORT: Modeling Dynamic Objects in Recurrent for Multi-Camera 3D Object\n Detection and Tracking", "abstract": " Recent multi-camera 3D object detectors usually leverage temporal information\nto construct multi-view stereo that alleviates the ill-posed depth estimation.\nHowever, they typically assume all the objects are static and directly\naggregate features across frames. This work begins with a theoretical and\nempirical analysis to reveal that ignoring the motion of moving objects can\nresult in serious localization bias. Therefore, we propose to model Dynamic\nObjects in RecurrenT (DORT) to tackle this problem. In contrast to previous\nglobal Bird-Eye-View (BEV) methods, DORT extracts object-wise local volumes for\nmotion estimation that also alleviates the heavy computational burden. By\niteratively refining the estimated object motion and location, the preceding\nfeatures can be precisely aggregated to the current frame to mitigate the\naforementioned adverse effects. The simple framework has two significant\nappealing properties. It is flexible and practical that can be plugged into\nmost camera-based 3D object detectors. As there are predictions of object\nmotion in the loop, it can easily track objects across frames according to\ntheir nearest center distances. Without bells and whistles, DORT outperforms\nall the previous methods on the nuScenes detection and tracking benchmarks with\n62.5\\% NDS and 57.6\\% AMOTA, respectively. The source code will be released.\n", "date": "2023-03-29", "categories": []} {"id": "2304.09446", "title": "Density-Insensitive Unsupervised Domain Adaption on 3D Object Detection", "abstract": " 3D object detection from point clouds is crucial in safety-critical\nautonomous driving. Although many works have made great efforts and achieved\nsignificant progress on this task, most of them suffer from expensive\nannotation cost and poor transferability to unknown data due to the domain gap.\nRecently, few works attempt to tackle the domain gap in objects, but still fail\nto adapt to the gap of varying beam-densities between two domains, which is\ncritical to mitigate the characteristic differences of the LiDAR collectors. To\nthis end, we make the attempt to propose a density-insensitive domain adaption\nframework to address the density-induced domain gap. In particular, we first\nintroduce Random Beam Re-Sampling (RBRS) to enhance the robustness of 3D\ndetectors trained on the source domain to the varying beam-density. Then, we\ntake this pre-trained detector as the backbone model, and feed the unlabeled\ntarget domain data into our newly designed task-specific teacher-student\nframework for predicting its high-quality pseudo labels. To further adapt the\nproperty of density-insensitivity into the target domain, we feed the teacher\nand student branches with the same sample of different densities, and propose\nan Object Graph Alignment (OGA) module to construct two object-graphs between\nthe two branches for enforcing the consistency in both the attribute and\nrelation of cross-density objects. Experimental results on three widely adopted\n3D object detection datasets demonstrate that our proposed domain adaption\nmethod outperforms the state-of-the-art methods, especially over\nvarying-density data. Code is available at\nhttps://github.com/WoodwindHu/DTS}{https://github.com/WoodwindHu/DTS.\n", "date": "2023-04-19", "categories": []} {"id": "2304.09609", "title": "MMDR: A Result Feature Fusion Object Detection Approach for Autonomous\n System", "abstract": " Object detection has been extensively utilized in autonomous systems in\nrecent years, encompassing both 2D and 3D object detection. Recent research in\nthis field has primarily centered around multimodal approaches for addressing\nthis issue.In this paper, a multimodal fusion approach based on result\nfeature-level fusion is proposed. This method utilizes the outcome features\ngenerated from single modality sources, and fuses them for downstream\ntasks.Based on this method, a new post-fusing network is proposed for\nmultimodal object detection, which leverages the single modality outcomes as\nfeatures. The proposed approach, called Multi-Modal Detector based on Result\nfeatures (MMDR), is designed to work for both 2D and 3D object detection tasks.\nCompared to previous multimodal models, the proposed approach in this paper\nperforms feature fusion at a later stage, enabling better representation of the\ndeep-level features of single modality sources. Additionally, the MMDR model\nincorporates shallow global features during the feature fusion stage, endowing\nthe model with the ability to perceive background information and the overall\ninput, thereby avoiding issues such as missed detections.\n", "date": "2023-04-19", "categories": []} {"id": "2304.09694", "title": "CrossFusion: Interleaving Cross-modal Complementation for\n Noise-resistant 3D Object Detection", "abstract": " The combination of LiDAR and camera modalities is proven to be necessary and\ntypical for 3D object detection according to recent studies. Existing fusion\nstrategies tend to overly rely on the LiDAR modal in essence, which exploits\nthe abundant semantics from the camera sensor insufficiently. However, existing\nmethods cannot rely on information from other modalities because the corruption\nof LiDAR features results in a large domain gap. Following this, we propose\nCrossFusion, a more robust and noise-resistant scheme that makes full use of\nthe camera and LiDAR features with the designed cross-modal complementation\nstrategy. Extensive experiments we conducted show that our method not only\noutperforms the state-of-the-art methods under the setting without introducing\nan extra depth estimation network but also demonstrates our model's noise\nresistance without re-training for the specific malfunction scenarios by\nincreasing 5.2\\% mAP and 2.4\\% NDS.\n", "date": "2023-04-19", "categories": []} {"id": "2303.03583", "title": "Calibration-free BEV Representation for Infrastructure Perception", "abstract": " Effective BEV object detection on infrastructure can greatly improve traffic\nscenes understanding and vehicle-toinfrastructure (V2I) cooperative perception.\nHowever, cameras installed on infrastructure have various postures, and\nprevious BEV detection methods rely on accurate calibration, which is difficult\nfor practical applications due to inevitable natural factors (e.g., wind and\nsnow). In this paper, we propose a Calibration-free BEV Representation (CBR)\nnetwork, which achieves 3D detection based on BEV representation without\ncalibration parameters and additional depth supervision. Specifically, we\nutilize two multi-layer perceptrons for decoupling the features from\nperspective view to front view and birdeye view under boxes-induced foreground\nsupervision. Then, a cross-view feature fusion module matches features from\northogonal views according to similarity and conducts BEV feature enhancement\nwith front view features. Experimental results on DAIR-V2X demonstrate that CBR\nachieves acceptable performance without any camera parameters and is naturally\nnot affected by calibration noises. We hope CBR can serve as a baseline for\nfuture research addressing practical challenges of infrastructure perception.\n", "date": "2023-03-07", "categories": []} {"id": "2303.08498", "title": "BEVHeight: A Robust Framework for Vision-based Roadside 3D Object\n Detection", "abstract": " While most recent autonomous driving system focuses on developing perception\nmethods on ego-vehicle sensors, people tend to overlook an alternative approach\nto leverage intelligent roadside cameras to extend the perception ability\nbeyond the visual range. We discover that the state-of-the-art vision-centric\nbird's eye view detection methods have inferior performances on roadside\ncameras. This is because these methods mainly focus on recovering the depth\nregarding the camera center, where the depth difference between the car and the\nground quickly shrinks while the distance increases. In this paper, we propose\na simple yet effective approach, dubbed BEVHeight, to address this issue. In\nessence, instead of predicting the pixel-wise depth, we regress the height to\nthe ground to achieve a distance-agnostic formulation to ease the optimization\nprocess of camera-only perception methods. On popular 3D detection benchmarks\nof roadside cameras, our method surpasses all previous vision-centric methods\nby a significant margin. The code is available at\n{\\url{https://github.com/ADLab-AutoDrive/BEVHeight}}.\n", "date": "2023-03-15", "categories": []} {"id": "2304.04179", "title": "Sparse Dense Fusion for 3D Object Detection", "abstract": " With the prevalence of multimodal learning, camera-LiDAR fusion has gained\npopularity in 3D object detection. Although multiple fusion approaches have\nbeen proposed, they can be classified into either sparse-only or dense-only\nfashion based on the feature representation in the fusion module. In this\npaper, we analyze them in a common taxonomy and thereafter observe two\nchallenges: 1) sparse-only solutions preserve 3D geometric prior and yet lose\nrich semantic information from the camera, and 2) dense-only alternatives\nretain the semantic continuity but miss the accurate geometric information from\nLiDAR. By analyzing these two formulations, we conclude that the information\nloss is inevitable due to their design scheme. To compensate for the\ninformation loss in either manner, we propose Sparse Dense Fusion (SDF), a\ncomplementary framework that incorporates both sparse-fusion and dense-fusion\nmodules via the Transformer architecture. Such a simple yet effective\nsparse-dense fusion structure enriches semantic texture and exploits spatial\nstructure information simultaneously. Through our SDF strategy, we assemble two\npopular methods with moderate performance and outperform baseline by 4.3% in\nmAP and 2.5% in NDS, ranking first on the nuScenes benchmark. Extensive\nablations demonstrate the effectiveness of our method and empirically align our\nanalysis.\n", "date": "2023-04-09", "categories": []} {"id": "2304.04185", "title": "BEVStereo++: Accurate Depth Estimation in Multi-view 3D Object Detection\n via Dynamic Temporal Stereo", "abstract": " Bounded by the inherent ambiguity of depth perception, contemporary\nmulti-view 3D object detection methods fall into the performance bottleneck.\nIntuitively, leveraging temporal multi-view stereo (MVS) technology is the\nnatural knowledge for tackling this ambiguity. However, traditional attempts of\nMVS has two limitations when applying to 3D object detection scenes: 1) The\naffinity measurement among all views suffers expensive computational cost; 2)\nIt is difficult to deal with outdoor scenarios where objects are often mobile.\nTo this end, we propose BEVStereo++: by introducing a dynamic temporal stereo\nstrategy, BEVStereo++ is able to cut down the harm that is brought by\nintroducing temporal stereo when dealing with those two scenarios. Going one\nstep further, we apply Motion Compensation Module and long sequence Frame\nFusion to BEVStereo++, which shows further performance boosting and error\nreduction. Without bells and whistles, BEVStereo++ achieves\nstate-of-the-art(SOTA) on both Waymo and nuScenes dataset.\n", "date": "2023-04-09", "categories": []} {"id": "2304.04248", "title": "Curricular Object Manipulation in LiDAR-based Object Detection", "abstract": " This paper explores the potential of curriculum learning in LiDAR-based 3D\nobject detection by proposing a curricular object manipulation (COM) framework.\nThe framework embeds the curricular training strategy into both the loss design\nand the augmentation process. For the loss design, we propose the COMLoss to\ndynamically predict object-level difficulties and emphasize objects of\ndifferent difficulties based on training stages. On top of the widely-used\naugmentation technique called GT-Aug in LiDAR detection tasks, we propose a\nnovel COMAug strategy which first clusters objects in ground-truth database\nbased on well-designed heuristics. Group-level difficulties rather than\nindividual ones are then predicted and updated during training for stable\nresults. Model performance and generalization capabilities can be improved by\nsampling and augmenting progressively more difficult objects into the training\nsamples. Extensive experiments and ablation studies reveal the superior and\ngenerality of the proposed framework. The code is available at\nhttps://github.com/ZZY816/COM.\n", "date": "2023-04-09", "categories": []} {"id": "2303.14548", "title": "Viewpoint Equivariance for Multi-View 3D Object Detection", "abstract": " 3D object detection from visual sensors is a cornerstone capability of\nrobotic systems. State-of-the-art methods focus on reasoning and decoding\nobject bounding boxes from multi-view camera input. In this work we gain\nintuition from the integral role of multi-view consistency in 3D scene\nunderstanding and geometric learning. To this end, we introduce VEDet, a novel\n3D object detection framework that exploits 3D multi-view geometry to improve\nlocalization through viewpoint awareness and equivariance. VEDet leverages a\nquery-based transformer architecture and encodes the 3D scene by augmenting\nimage features with positional encodings from their 3D perspective geometry. We\ndesign view-conditioned queries at the output level, which enables the\ngeneration of multiple virtual frames during training to learn viewpoint\nequivariance by enforcing multi-view consistency. The multi-view geometry\ninjected at the input level as positional encodings and regularized at the loss\nlevel provides rich geometric cues for 3D object detection, leading to\nstate-of-the-art performance on the nuScenes benchmark. The code and model are\nmade available at https://github.com/TRI-ML/VEDet.\n", "date": "2023-03-25", "categories": []} {"id": "2304.03105", "title": "Geometric-aware Pretraining for Vision-centric 3D Object Detection", "abstract": " Multi-camera 3D object detection for autonomous driving is a challenging\nproblem that has garnered notable attention from both academia and industry. An\nobstacle encountered in vision-based techniques involves the precise extraction\nof geometry-conscious features from RGB images. Recent approaches have utilized\ngeometric-aware image backbones pretrained on depth-relevant tasks to acquire\nspatial information. However, these approaches overlook the critical aspect of\nview transformation, resulting in inadequate performance due to the\nmisalignment of spatial knowledge between the image backbone and view\ntransformation. To address this issue, we propose a novel geometric-aware\npretraining framework called GAPretrain. Our approach incorporates spatial and\nstructural cues to camera networks by employing the geometric-rich modality as\nguidance during the pretraining phase. The transference of modal-specific\nattributes across different modalities is non-trivial, but we bridge this gap\nby using a unified bird's-eye-view (BEV) representation and structural hints\nderived from LiDAR point clouds to facilitate the pretraining process.\nGAPretrain serves as a plug-and-play solution that can be flexibly applied to\nmultiple state-of-the-art detectors. Our experiments demonstrate the\neffectiveness and generalization ability of the proposed method. We achieve\n46.2 mAP and 55.5 NDS on the nuScenes val set using the BEVFormer method, with\na gain of 2.7 and 2.1 points, respectively. We also conduct experiments on\nvarious image backbones and view transformations to validate the efficacy of\nour approach. Code will be released at\nhttps://github.com/OpenDriveLab/BEVPerception-Survey-Recipe.\n", "date": "2023-04-06", "categories": []} {"id": "2207.10909", "title": "DBQ-SSD: Dynamic Ball Query for Efficient 3D Object Detection", "abstract": " Many point-based 3D detectors adopt point-feature sampling strategies to drop\nsome points for efficient inference. These strategies are typically based on\nfixed and handcrafted rules, making it difficult to handle complicated scenes.\nDifferent from them, we propose a Dynamic Ball Query (DBQ) network to\nadaptively select a subset of input points according to the input features, and\nassign the feature transform with a suitable receptive field for each selected\npoint. It can be embedded into some state-of-the-art 3D detectors and trained\nin an end-to-end manner, which significantly reduces the computational cost.\nExtensive experiments demonstrate that our method can increase the inference\nspeed by 30%-100% on KITTI, Waymo, and ONCE datasets. Specifically, the\ninference speed of our detector can reach 162 FPS on KITTI scene, and 30 FPS on\nWaymo and ONCE scenes without performance degradation. Due to skipping the\nredundant points, some evaluation metrics show significant improvements. Codes\nwill be released at https://github.com/yancie-yjr/DBQ-SSD.\n", "date": "2022-07-22", "categories": []} {"id": "2206.09474", "title": "3D Object Detection for Autonomous Driving: A Comprehensive Survey", "abstract": " Autonomous driving, in recent years, has been receiving increasing attention\nfor its potential to relieve drivers' burdens and improve the safety of\ndriving. In modern autonomous driving pipelines, the perception system is an\nindispensable component, aiming to accurately estimate the status of\nsurrounding environments and provide reliable observations for prediction and\nplanning. 3D object detection, which intelligently predicts the locations,\nsizes, and categories of the critical 3D objects near an autonomous vehicle, is\nan important part of a perception system. This paper reviews the advances in 3D\nobject detection for autonomous driving. First, we introduce the background of\n3D object detection and discuss the challenges in this task. Second, we conduct\na comprehensive survey of the progress in 3D object detection from the aspects\nof models and sensory inputs, including LiDAR-based, camera-based, and\nmulti-modal detection approaches. We also provide an in-depth analysis of the\npotentials and challenges in each category of methods. Additionally, we\nsystematically investigate the applications of 3D object detection in driving\nsystems. Finally, we conduct a performance analysis of the 3D object detection\napproaches, and we further summarize the research trends over the years and\nprospect the future directions of this area.\n", "date": "2022-06-19", "categories": []} {"id": "2211.12501", "title": "AeDet: Azimuth-invariant Multi-view 3D Object Detection", "abstract": " Recent LSS-based multi-view 3D object detection has made tremendous progress,\nby processing the features in Brid-Eye-View (BEV) via the convolutional\ndetector. However, the typical convolution ignores the radial symmetry of the\nBEV features and increases the difficulty of the detector optimization. To\npreserve the inherent property of the BEV features and ease the optimization,\nwe propose an azimuth-equivariant convolution (AeConv) and an\nazimuth-equivariant anchor. The sampling grid of AeConv is always in the radial\ndirection, thus it can learn azimuth-invariant BEV features. The proposed\nanchor enables the detection head to learn predicting azimuth-irrelevant\ntargets. In addition, we introduce a camera-decoupled virtual depth to unify\nthe depth prediction for the images with different camera intrinsic parameters.\nThe resultant detector is dubbed Azimuth-equivariant Detector (AeDet).\nExtensive experiments are conducted on nuScenes, and AeDet achieves a 62.0%\nNDS, surpassing the recent multi-view 3D object detectors such as PETRv2 and\nBEVDepth by a large margin. Project page: https://fcjian.github.io/aedet.\n", "date": "2022-11-22", "categories": []} {"id": "2304.01289", "title": "Monocular 3D Object Detection with Bounding Box Denoising in 3D by\n Perceiver", "abstract": " The main challenge of monocular 3D object detection is the accurate\nlocalization of 3D center. Motivated by a new and strong observation that this\nchallenge can be remedied by a 3D-space local-grid search scheme in an ideal\ncase, we propose a stage-wise approach, which combines the information flow\nfrom 2D-to-3D (3D bounding box proposal generation with a single 2D image) and\n3D-to-2D (proposal verification by denoising with 3D-to-2D contexts) in a\ntop-down manner. Specifically, we first obtain initial proposals from\noff-the-shelf backbone monocular 3D detectors. Then, we generate a 3D anchor\nspace by local-grid sampling from the initial proposals. Finally, we perform 3D\nbounding box denoising at the 3D-to-2D proposal verification stage. To\neffectively learn discriminative features for denoising highly overlapped\nproposals, this paper presents a method of using the Perceiver I/O model to\nfuse the 3D-to-2D geometric information and the 2D appearance information. With\nthe encoded latent representation of a proposal, the verification head is\nimplemented with a self-attention module. Our method, named as MonoXiver, is\ngeneric and can be easily adapted to any backbone monocular 3D detectors.\nExperimental results on the well-established KITTI dataset and the challenging\nlarge-scale Waymo dataset show that MonoXiver consistently achieves improvement\nwith limited computation overhead.\n", "date": "2023-04-03", "categories": []} {"id": "2304.01464", "title": "Hierarchical Supervision and Shuffle Data Augmentation for 3D\n Semi-Supervised Object Detection", "abstract": " State-of-the-art 3D object detectors are usually trained on large-scale\ndatasets with high-quality 3D annotations. However, such 3D annotations are\noften expensive and time-consuming, which may not be practical for real\napplications. A natural remedy is to adopt semi-supervised learning (SSL) by\nleveraging a limited amount of labeled samples and abundant unlabeled samples.\nCurrent pseudolabeling-based SSL object detection methods mainly adopt a\nteacher-student framework, with a single fixed threshold strategy to generate\nsupervision signals, which inevitably brings confused supervision when guiding\nthe student network training. Besides, the data augmentation of the point cloud\nin the typical teacher-student framework is too weak, and only contains basic\ndown sampling and flip-and-shift (i.e., rotate and scaling), which hinders the\neffective learning of feature information. Hence, we address these issues by\nintroducing a novel approach of Hierarchical Supervision and Shuffle Data\nAugmentation (HSSDA), which is a simple yet effective teacher-student\nframework. The teacher network generates more reasonable supervision for the\nstudent network by designing a dynamic dual-threshold strategy. Besides, the\nshuffle data augmentation strategy is designed to strengthen the feature\nrepresentation ability of the student network. Extensive experiments show that\nHSSDA consistently outperforms the recent state-of-the-art methods on different\ndatasets. The code will be released at https://github.com/azhuantou/HSSDA.\n", "date": "2023-04-04", "categories": []} {"id": "2304.01519", "title": "LiDAR-Based 3D Object Detection via Hybrid 2D Semantic Scene Generation", "abstract": " Bird's-Eye View (BEV) features are popular intermediate scene representations\nshared by the 3D backbone and the detector head in LiDAR-based object\ndetectors. However, little research has been done to investigate how to\nincorporate additional supervision on the BEV features to improve proposal\ngeneration in the detector head, while still balancing the number of powerful\n3D layers and efficient 2D network operations. This paper proposes a novel\nscene representation that encodes both the semantics and geometry of the 3D\nenvironment in 2D, which serves as a dense supervision signal for better BEV\nfeature learning. The key idea is to use auxiliary networks to predict a\ncombination of explicit and implicit semantic probabilities by exploiting their\ncomplementary properties. Extensive experiments show that our simple yet\neffective design can be easily integrated into most state-of-the-art 3D object\ndetectors and consistently improves upon baseline models.\n", "date": "2023-04-04", "categories": []} {"id": "2304.00967", "title": "Temporal Enhanced Training of Multi-view 3D Object Detector via\n Historical Object Prediction", "abstract": " In this paper, we propose a new paradigm, named Historical Object Prediction\n(HoP) for multi-view 3D detection to leverage temporal information more\neffectively. The HoP approach is straightforward: given the current timestamp\nt, we generate a pseudo Bird's-Eye View (BEV) feature of timestamp t-k from its\nadjacent frames and utilize this feature to predict the object set at timestamp\nt-k. Our approach is motivated by the observation that enforcing the detector\nto capture both the spatial location and temporal motion of objects occurring\nat historical timestamps can lead to more accurate BEV feature learning. First,\nwe elaborately design short-term and long-term temporal decoders, which can\ngenerate the pseudo BEV feature for timestamp t-k without the involvement of\nits corresponding camera images. Second, an additional object decoder is\nflexibly attached to predict the object targets using the generated pseudo BEV\nfeature. Note that we only perform HoP during training, thus the proposed\nmethod does not introduce extra overheads during inference. As a plug-and-play\napproach, HoP can be easily incorporated into state-of-the-art BEV detection\nframeworks, including BEVFormer and BEVDet series. Furthermore, the auxiliary\nHoP approach is complementary to prevalent temporal modeling methods, leading\nto significant performance gains. Extensive experiments are conducted to\nevaluate the effectiveness of the proposed HoP on the nuScenes dataset. We\nchoose the representative methods, including BEVFormer and BEVDet4D-Depth to\nevaluate our method. Surprisingly, HoP achieves 68.5% NDS and 62.4% mAP with\nViT-L on nuScenes test, outperforming all the 3D object detectors on the\nleaderboard. Codes will be available at https://github.com/Sense-X/HoP.\n", "date": "2023-04-03", "categories": []} {"id": "2304.01054", "title": "VoxelFormer: Bird's-Eye-View Feature Generation based on Dual-view\n Attention for Multi-view 3D Object Detection", "abstract": " In recent years, transformer-based detectors have demonstrated remarkable\nperformance in 2D visual perception tasks. However, their performance in\nmulti-view 3D object detection remains inferior to the state-of-the-art (SOTA)\nof convolutional neural network based detectors. In this work, we investigate\nthis issue from the perspective of bird's-eye-view (BEV) feature generation.\nSpecifically, we examine the BEV feature generation method employed by the\ntransformer-based SOTA, BEVFormer, and identify its two limitations: (i) it\nonly generates attention weights from BEV, which precludes the use of lidar\npoints for supervision, and (ii) it aggregates camera view features to the BEV\nthrough deformable sampling, which only selects a small subset of features and\nfails to exploit all information. To overcome these limitations, we propose a\nnovel BEV feature generation method, dual-view attention, which generates\nattention weights from both the BEV and camera view. This method encodes all\ncamera features into the BEV feature. By combining dual-view attention with the\nBEVFormer architecture, we build a new detector named VoxelFormer. Extensive\nexperiments are conducted on the nuScenes benchmark to verify the superiority\nof dual-view attention and VoxelForer. We observe that even only adopting 3\nencoders and 1 historical frame during training, VoxelFormer still outperforms\nBEVFormer significantly. When trained in the same setting, VoxelFormer can\nsurpass BEVFormer by 4.9% NDS point. Code is available at:\nhttps://github.com/Lizhuoling/VoxelFormer-public.git.\n", "date": "2023-04-03", "categories": []} {"id": "2303.17921", "title": "IC-FPS: Instance-Centroid Faster Point Sampling Module for 3D Point-base\n Object Detection", "abstract": " 3D object detection is one of the most important tasks in autonomous driving\nand robotics. Our research focuses on tackling low efficiency issue of\npoint-based methods on large-scale point clouds. Existing point-based methods\nadopt farthest point sampling (FPS) strategy for downsampling, which is\ncomputationally expensive in terms of inference time and memory consumption\nwhen the number of point cloud increases. In order to improve efficiency, we\npropose a novel Instance-Centroid Faster Point Sampling Module (IC-FPS) , which\neffectively replaces the first Set Abstraction (SA) layer that is extremely\ntedious. IC-FPS module is comprised of two methods, local feature diffusion\nbased background point filter (LFDBF) and Centroid-Instance Sampling Strategy\n(CISS). LFDBF is constructed to exclude most invalid background points, while\nCISS substitutes FPS strategy by fast sampling centroids and instance points.\nIC-FPS module can be inserted to almost every point-based models. Extensive\nexperiments on multiple public benchmarks have demonstrated the superiority of\nIC-FPS. On Waymo dataset, the proposed module significantly improves\nperformance of baseline model and accelerates inference speed by 3.8 times. For\nthe first time, real-time detection of point-based models in large-scale point\ncloud scenario is realized.\n", "date": "2023-03-31", "categories": []} {"id": "2303.17099", "title": "BEVFusion4D: Learning LiDAR-Camera Fusion Under Bird's-Eye-View via\n Cross-Modality Guidance and Temporal Aggregation", "abstract": " Integrating LiDAR and Camera information into Bird's-Eye-View (BEV) has\nbecome an essential topic for 3D object detection in autonomous driving.\nExisting methods mostly adopt an independent dual-branch framework to generate\nLiDAR and camera BEV, then perform an adaptive modality fusion. Since point\nclouds provide more accurate localization and geometry information, they could\nserve as a reliable spatial prior to acquiring relevant semantic information\nfrom the images. Therefore, we design a LiDAR-Guided View Transformer (LGVT) to\neffectively obtain the camera representation in BEV space and thus benefit the\nwhole dual-branch fusion system. LGVT takes camera BEV as the primitive\nsemantic query, repeatedly leveraging the spatial cue of LiDAR BEV for\nextracting image features across multiple camera views. Moreover, we extend our\nframework into the temporal domain with our proposed Temporal Deformable\nAlignment (TDA) module, which aims to aggregate BEV features from multiple\nhistorical frames. Including these two modules, our framework dubbed\nBEVFusion4D achieves state-of-the-art results in 3D object detection, with\n72.0% mAP and 73.5% NDS on the nuScenes validation set, and 73.3% mAP and 74.7%\nNDS on nuScenes test set, respectively.\n", "date": "2023-03-30", "categories": []} {"id": "2303.16094", "title": "LinK: Linear Kernel for LiDAR-based 3D Perception", "abstract": " Extending the success of 2D Large Kernel to 3D perception is challenging due\nto: 1. the cubically-increasing overhead in processing 3D data; 2. the\noptimization difficulties from data scarcity and sparsity. Previous work has\ntaken the first step to scale up the kernel size from 3x3x3 to 7x7x7 by\nintroducing block-shared weights. However, to reduce the feature variations\nwithin a block, it only employs modest block size and fails to achieve larger\nkernels like the 21x21x21. To address this issue, we propose a new method,\ncalled LinK, to achieve a wider-range perception receptive field in a\nconvolution-like manner with two core designs. The first is to replace the\nstatic kernel matrix with a linear kernel generator, which adaptively provides\nweights only for non-empty voxels. The second is to reuse the pre-computed\naggregation results in the overlapped blocks to reduce computation complexity.\nThe proposed method successfully enables each voxel to perceive context within\na range of 21x21x21. Extensive experiments on two basic perception tasks, 3D\nobject detection and 3D semantic segmentation, demonstrate the effectiveness of\nour method. Notably, we rank 1st on the public leaderboard of the 3D detection\nbenchmark of nuScenes (LiDAR track), by simply incorporating a LinK-based\nbackbone into the basic detector, CenterPoint. We also boost the strong\nsegmentation baseline's mIoU with 2.7% in the SemanticKITTI test set. Code is\navailable at https://github.com/MCG-NJU/LinK.\n", "date": "2023-03-28", "categories": []} {"id": "2211.11646", "title": "NeRF-RPN: A general framework for object detection in NeRFs", "abstract": " This paper presents the first significant object detection framework,\nNeRF-RPN, which directly operates on NeRF. Given a pre-trained NeRF model,\nNeRF-RPN aims to detect all bounding boxes of objects in a scene. By exploiting\na novel voxel representation that incorporates multi-scale 3D neural volumetric\nfeatures, we demonstrate it is possible to regress the 3D bounding boxes of\nobjects in NeRF directly without rendering the NeRF at any viewpoint. NeRF-RPN\nis a general framework and can be applied to detect objects without class\nlabels. We experimented NeRF-RPN with various backbone architectures, RPN head\ndesigns and loss functions. All of them can be trained in an end-to-end manner\nto estimate high quality 3D bounding boxes. To facilitate future research in\nobject detection for NeRF, we built a new benchmark dataset which consists of\nboth synthetic and real-world data with careful labeling and clean up. Code and\ndataset are available at https://github.com/lyclyc52/NeRF_RPN.\n", "date": "2022-11-21", "categories": []} {"id": "2303.14893", "title": "Context-Aware Transformer for 3D Point Cloud Automatic Annotation", "abstract": " 3D automatic annotation has received increased attention since manually\nannotating 3D point clouds is laborious. However, existing methods are usually\ncomplicated, e.g., pipelined training for 3D foreground/background\nsegmentation, cylindrical object proposals, and point completion. Furthermore,\nthey often overlook the inter-object feature relation that is particularly\ninformative to hard samples for 3D annotation. To this end, we propose a simple\nyet effective end-to-end Context-Aware Transformer (CAT) as an automated 3D-box\nlabeler to generate precise 3D box annotations from 2D boxes, trained with a\nsmall number of human annotations. We adopt the general encoder-decoder\narchitecture, where the CAT encoder consists of an intra-object encoder (local)\nand an inter-object encoder (global), performing self-attention along the\nsequence and batch dimensions, respectively. The former models intra-object\ninteractions among points, and the latter extracts feature relations among\ndifferent objects, thus boosting scene-level understanding. Via local and\nglobal encoders, CAT can generate high-quality 3D box annotations with a\nstreamlined workflow, allowing it to outperform existing state-of-the-art by up\nto 1.79% 3D AP on the hard task of the KITTI test set.\n", "date": "2023-03-27", "categories": []} {"id": "2303.15083", "title": "UniDistill: A Universal Cross-Modality Knowledge Distillation Framework\n for 3D Object Detection in Bird's-Eye View", "abstract": " In the field of 3D object detection for autonomous driving, the sensor\nportfolio including multi-modality and single-modality is diverse and complex.\nSince the multi-modal methods have system complexity while the accuracy of\nsingle-modal ones is relatively low, how to make a tradeoff between them is\ndifficult. In this work, we propose a universal cross-modality knowledge\ndistillation framework (UniDistill) to improve the performance of\nsingle-modality detectors. Specifically, during training, UniDistill projects\nthe features of both the teacher and the student detector into Bird's-Eye-View\n(BEV), which is a friendly representation for different modalities. Then, three\ndistillation losses are calculated to sparsely align the foreground features,\nhelping the student learn from the teacher without introducing additional cost\nduring inference. Taking advantage of the similar detection paradigm of\ndifferent detectors in BEV, UniDistill easily supports LiDAR-to-camera,\ncamera-to-LiDAR, fusion-to-LiDAR and fusion-to-camera distillation paths.\nFurthermore, the three distillation losses can filter the effect of misaligned\nbackground information and balance between objects of different sizes,\nimproving the distillation effectiveness. Extensive experiments on nuScenes\ndemonstrate that UniDistill effectively improves the mAP and NDS of student\ndetectors by 2.0%~3.2%.\n", "date": "2023-03-27", "categories": []} {"id": "2303.15416", "title": "3D Video Object Detection with Learnable Object-Centric Global\n Optimization", "abstract": " We explore long-term temporal visual correspondence-based optimization for 3D\nvideo object detection in this work. Visual correspondence refers to one-to-one\nmappings for pixels across multiple images. Correspondence-based optimization\nis the cornerstone for 3D scene reconstruction but is less studied in 3D video\nobject detection, because moving objects violate multi-view geometry\nconstraints and are treated as outliers during scene reconstruction. We address\nthis issue by treating objects as first-class citizens during\ncorrespondence-based optimization. In this work, we propose BA-Det, an\nend-to-end optimizable object detector with object-centric temporal\ncorrespondence learning and featuremetric object bundle adjustment.\nEmpirically, we verify the effectiveness and efficiency of BA-Det for multiple\nbaseline 3D detectors under various setups. Our BA-Det achieves SOTA\nperformance on the large-scale Waymo Open Dataset (WOD) with only marginal\ncomputation cost. Our code is available at\nhttps://github.com/jiaweihe1996/BA-Det.\n", "date": "2023-03-27", "categories": []} {"id": "2207.10660", "title": "Omni3D: A Large Benchmark and Model for 3D Object Detection in the Wild", "abstract": " Recognizing scenes and objects in 3D from a single image is a longstanding\ngoal of computer vision with applications in robotics and AR/VR. For 2D\nrecognition, large datasets and scalable solutions have led to unprecedented\nadvances. In 3D, existing benchmarks are small in size and approaches\nspecialize in few object categories and specific domains, e.g. urban driving\nscenes. Motivated by the success of 2D recognition, we revisit the task of 3D\nobject detection by introducing a large benchmark, called Omni3D. Omni3D\nre-purposes and combines existing datasets resulting in 234k images annotated\nwith more than 3 million instances and 98 categories. 3D detection at such\nscale is challenging due to variations in camera intrinsics and the rich\ndiversity of scene and object types. We propose a model, called Cube R-CNN,\ndesigned to generalize across camera and scene types with a unified approach.\nWe show that Cube R-CNN outperforms prior works on the larger Omni3D and\nexisting benchmarks. Finally, we prove that Omni3D is a powerful dataset for 3D\nobject recognition and show that it improves single-dataset performance and can\naccelerate learning on new smaller datasets via pre-training.\n", "date": "2022-07-21", "categories": []} {"id": "2301.04467", "title": "FrustumFormer: Adaptive Instance-aware Resampling for Multi-view 3D\n Detection", "abstract": " The transformation of features from 2D perspective space to 3D space is\nessential to multi-view 3D object detection. Recent approaches mainly focus on\nthe design of view transformation, either pixel-wisely lifting perspective view\nfeatures into 3D space with estimated depth or grid-wisely constructing BEV\nfeatures via 3D projection, treating all pixels or grids equally. However,\nchoosing what to transform is also important but has rarely been discussed\nbefore. The pixels of a moving car are more informative than the pixels of the\nsky. To fully utilize the information contained in images, the view\ntransformation should be able to adapt to different image regions according to\ntheir contents. In this paper, we propose a novel framework named\nFrustumFormer, which pays more attention to the features in instance regions\nvia adaptive instance-aware resampling. Specifically, the model obtains\ninstance frustums on the bird's eye view by leveraging image view object\nproposals. An adaptive occupancy mask within the instance frustum is learned to\nrefine the instance location. Moreover, the temporal frustum intersection could\nfurther reduce the localization uncertainty of objects. Comprehensive\nexperiments on the nuScenes dataset demonstrate the effectiveness of\nFrustumFormer, and we achieve a new state-of-the-art performance on the\nbenchmark. Codes and models will be made available at\nhttps://github.com/Robertwyq/Frustum.\n", "date": "2023-01-10", "categories": []} {"id": "2303.13560", "title": "Collaboration Helps Camera Overtake LiDAR in 3D Detection", "abstract": " Camera-only 3D detection provides an economical solution with a simple\nconfiguration for localizing objects in 3D space compared to LiDAR-based\ndetection systems. However, a major challenge lies in precise depth estimation\ndue to the lack of direct 3D measurements in the input. Many previous methods\nattempt to improve depth estimation through network designs, e.g., deformable\nlayers and larger receptive fields. This work proposes an orthogonal direction,\nimproving the camera-only 3D detection by introducing multi-agent\ncollaborations. Our proposed collaborative camera-only 3D detection (CoCa3D)\nenables agents to share complementary information with each other through\ncommunication. Meanwhile, we optimize communication efficiency by selecting the\nmost informative cues. The shared messages from multiple viewpoints\ndisambiguate the single-agent estimated depth and complement the occluded and\nlong-range regions in the single-agent view. We evaluate CoCa3D in one\nreal-world dataset and two new simulation datasets. Results show that CoCa3D\nimproves previous SOTA performances by 44.21% on DAIR-V2X, 30.60% on OPV2V+,\n12.59% on CoPerception-UAVs+ for AP@70. Our preliminary results show a\npotential that with sufficient collaboration, the camera might overtake LiDAR\nin some practical scenarios. We released the dataset and code at\nhttps://siheng-chen.github.io/dataset/CoPerception+ and\nhttps://github.com/MediaBrain-SJTU/CoCa3D.\n", "date": "2023-03-23", "categories": []} {"id": "2303.13561", "title": "MoGDE: Boosting Mobile Monocular 3D Object Detection with Ground Depth\n Estimation", "abstract": " Monocular 3D object detection (Mono3D) in mobile settings (e.g., on a\nvehicle, a drone, or a robot) is an important yet challenging task. Due to the\nnear-far disparity phenomenon of monocular vision and the ever-changing camera\npose, it is hard to acquire high detection accuracy, especially for far\nobjects. Inspired by the insight that the depth of an object can be well\ndetermined according to the depth of the ground where it stands, in this paper,\nwe propose a novel Mono3D framework, called MoGDE, which constantly estimates\nthe corresponding ground depth of an image and then utilizes the estimated\nground depth information to guide Mono3D. To this end, we utilize a pose\ndetection network to estimate the pose of the camera and then construct a\nfeature map portraying pixel-level ground depth according to the 3D-to-2D\nperspective geometry. Moreover, to improve Mono3D with the estimated ground\ndepth, we design an RGB-D feature fusion network based on the transformer\nstructure, where the long-range self-attention mechanism is utilized to\neffectively identify ground-contacting points and pin the corresponding ground\ndepth to the image feature map. We conduct extensive experiments on the\nreal-world KITTI dataset. The results demonstrate that MoGDE can effectively\nimprove the Mono3D accuracy and robustness for both near and far objects. MoGDE\nyields the best performance compared with the state-of-the-art methods by a\nlarge margin and is ranked number one on the KITTI 3D benchmark.\n", "date": "2023-03-23", "categories": []} {"id": "2303.13018", "title": "MonoATT: Online Monocular 3D Object Detection with Adaptive Token\n Transformer", "abstract": " Mobile monocular 3D object detection (Mono3D) (e.g., on a vehicle, a drone,\nor a robot) is an important yet challenging task. Existing transformer-based\noffline Mono3D models adopt grid-based vision tokens, which is suboptimal when\nusing coarse tokens due to the limited available computational power. In this\npaper, we propose an online Mono3D framework, called MonoATT, which leverages a\nnovel vision transformer with heterogeneous tokens of varying shapes and sizes\nto facilitate mobile Mono3D. The core idea of MonoATT is to adaptively assign\nfiner tokens to areas of more significance before utilizing a transformer to\nenhance Mono3D. To this end, we first use prior knowledge to design a scoring\nnetwork for selecting the most important areas of the image, and then propose a\ntoken clustering and merging network with an attention mechanism to gradually\nmerge tokens around the selected areas in multiple stages. Finally, a\npixel-level feature map is reconstructed from heterogeneous tokens before\nemploying a SOTA Mono3D detector as the underlying detection core. Experiment\nresults on the real-world KITTI dataset demonstrate that MonoATT can\neffectively improve the Mono3D accuracy for both near and far objects and\nguarantee low latency. MonoATT yields the best performance compared with the\nstate-of-the-art methods by a large margin and is ranked number one on the\nKITTI 3D benchmark.\n", "date": "2023-03-23", "categories": []} {"id": "2206.10555", "title": "LargeKernel3D: Scaling up Kernels in 3D Sparse CNNs", "abstract": " Recent advance in 2D CNNs has revealed that large kernels are important.\nHowever, when directly applying large convolutional kernels in 3D CNNs, severe\ndifficulties are met, where those successful module designs in 2D become\nsurprisingly ineffective on 3D networks, including the popular depth-wise\nconvolution. To address this vital challenge, we instead propose the\nspatial-wise partition convolution and its large-kernel module. As a result, it\navoids the optimization and efficiency issues of naive 3D large kernels. Our\nlarge-kernel 3D CNN network, LargeKernel3D, yields notable improvement in 3D\ntasks of semantic segmentation and object detection. It achieves 73.9% mIoU on\nthe ScanNetv2 semantic segmentation and 72.8% NDS nuScenes object detection\nbenchmarks, ranking 1st on the nuScenes LIDAR leaderboard. The performance\nfurther boosts to 74.2% NDS with a simple multi-modal fusion. In addition,\nLargeKernel3D can be scaled to 17x17x17 kernel size on Waymo 3D object\ndetection. For the first time, we show that large kernels are feasible and\nessential for 3D visual tasks.\n", "date": "2022-06-21", "categories": []} {"id": "2212.08273", "title": "Learning for Vehicle-to-Vehicle Cooperative Perception under Lossy\n Communication", "abstract": " Deep learning has been widely used in the perception (e.g., 3D object\ndetection) of intelligent vehicle driving. Due to the beneficial\nVehicle-to-Vehicle (V2V) communication, the deep learning based features from\nother agents can be shared to the ego vehicle so as to improve the perception\nof the ego vehicle. It is named as Cooperative Perception in the V2V research,\nwhose algorithms have been dramatically advanced recently. However, all the\nexisting cooperative perception algorithms assume the ideal V2V communication\nwithout considering the possible lossy shared features because of the Lossy\nCommunication (LC) which is common in the complex real-world driving scenarios.\nIn this paper, we first study the side effect (e.g., detection performance\ndrop) by the lossy communication in the V2V Cooperative Perception, and then we\npropose a novel intermediate LC-aware feature fusion method to relieve the side\neffect of lossy communication by a LC-aware Repair Network (LCRN) and enhance\nthe interaction between the ego vehicle and other vehicles by a specially\ndesigned V2V Attention Module (V2VAM) including intra-vehicle attention of ego\nvehicle and uncertainty-aware inter-vehicle attention. The extensive experiment\non the public cooperative perception dataset OPV2V (based on digital-twin CARLA\nsimulator) demonstrates that the proposed method is quite effective for the\ncooperative point cloud based 3D object detection under lossy V2V\ncommunication.\n", "date": "2022-12-16", "categories": []} {"id": "2303.12621", "title": "OcTr: Octree-based Transformer for 3D Object Detection", "abstract": " A key challenge for LiDAR-based 3D object detection is to capture sufficient\nfeatures from large scale 3D scenes especially for distant or/and occluded\nobjects. Albeit recent efforts made by Transformers with the long sequence\nmodeling capability, they fail to properly balance the accuracy and efficiency,\nsuffering from inadequate receptive fields or coarse-grained holistic\ncorrelations. In this paper, we propose an Octree-based Transformer, named\nOcTr, to address this issue. It first constructs a dynamic octree on the\nhierarchical feature pyramid through conducting self-attention on the top level\nand then recursively propagates to the level below restricted by the octants,\nwhich captures rich global context in a coarse-to-fine manner while maintaining\nthe computational complexity under control. Furthermore, for enhanced\nforeground perception, we propose a hybrid positional embedding, composed of\nthe semantic-aware positional embedding and attention mask, to fully exploit\nsemantic and geometry clues. Extensive experiments are conducted on the Waymo\nOpen Dataset and KITTI Dataset, and OcTr reaches newly state-of-the-art\nresults.\n", "date": "2023-03-22", "categories": []} {"id": "2303.10209", "title": "CAPE: Camera View Position Embedding for Multi-View 3D Object Detection", "abstract": " In this paper, we address the problem of detecting 3D objects from multi-view\nimages. Current query-based methods rely on global 3D position embeddings (PE)\nto learn the geometric correspondence between images and 3D space. We claim\nthat directly interacting 2D image features with global 3D PE could increase\nthe difficulty of learning view transformation due to the variation of camera\nextrinsics. Thus we propose a novel method based on CAmera view Position\nEmbedding, called CAPE. We form the 3D position embeddings under the local\ncamera-view coordinate system instead of the global coordinate system, such\nthat 3D position embedding is free of encoding camera extrinsic parameters.\nFurthermore, we extend our CAPE to temporal modeling by exploiting the object\nqueries of previous frames and encoding the ego-motion for boosting 3D object\ndetection. CAPE achieves state-of-the-art performance (61.0% NDS and 52.5% mAP)\namong all LiDAR-free methods on nuScenes dataset. Codes and models are\navailable on \\href{https://github.com/PaddlePaddle/Paddle3D}{Paddle3D} and\n\\href{https://github.com/kaixinbear/CAPE}{PyTorch Implementation}.\n", "date": "2023-03-17", "categories": []} {"id": "2303.10975", "title": "VIMI: Vehicle-Infrastructure Multi-view Intermediate Fusion for\n Camera-based 3D Object Detection", "abstract": " In autonomous driving, Vehicle-Infrastructure Cooperative 3D Object Detection\n(VIC3D) makes use of multi-view cameras from both vehicles and traffic\ninfrastructure, providing a global vantage point with rich semantic context of\nroad conditions beyond a single vehicle viewpoint. Two major challenges prevail\nin VIC3D: 1) inherent calibration noise when fusing multi-view images, caused\nby time asynchrony across cameras; 2) information loss when projecting 2D\nfeatures into 3D space. To address these issues, We propose a novel 3D object\ndetection framework, Vehicles-Infrastructure Multi-view Intermediate fusion\n(VIMI). First, to fully exploit the holistic perspectives from both vehicles\nand infrastructure, we propose a Multi-scale Cross Attention (MCA) module that\nfuses infrastructure and vehicle features on selective multi-scales to correct\nthe calibration noise introduced by camera asynchrony. Then, we design a\nCamera-aware Channel Masking (CCM) module that uses camera parameters as priors\nto augment the fused features. We further introduce a Feature Compression (FC)\nmodule with channel and spatial compression blocks to reduce the size of\ntransmitted features for enhanced efficiency. Experiments show that VIMI\nachieves 15.61% overall AP_3D and 21.44% AP_BEV on the new VIC3D dataset,\nDAIR-V2X-C, significantly outperforming state-of-the-art early fusion and late\nfusion methods with comparable transmission cost.\n", "date": "2023-03-20", "categories": []} {"id": "2303.11040", "title": "Benchmarking Robustness of 3D Object Detection to Common Corruptions in\n Autonomous Driving", "abstract": " 3D object detection is an important task in autonomous driving to perceive\nthe surroundings. Despite the excellent performance, the existing 3D detectors\nlack the robustness to real-world corruptions caused by adverse weathers,\nsensor noises, etc., provoking concerns about the safety and reliability of\nautonomous driving systems. To comprehensively and rigorously benchmark the\ncorruption robustness of 3D detectors, in this paper we design 27 types of\ncommon corruptions for both LiDAR and camera inputs considering real-world\ndriving scenarios. By synthesizing these corruptions on public datasets, we\nestablish three corruption robustness benchmarks -- KITTI-C, nuScenes-C, and\nWaymo-C. Then, we conduct large-scale experiments on 24 diverse 3D object\ndetection models to evaluate their corruption robustness. Based on the\nevaluation results, we draw several important findings, including: 1)\nmotion-level corruptions are the most threatening ones that lead to significant\nperformance drop of all models; 2) LiDAR-camera fusion models demonstrate\nbetter robustness; 3) camera-only models are extremely vulnerable to image\ncorruptions, showing the indispensability of LiDAR point clouds. We release the\nbenchmarks and codes at https://github.com/kkkcx/3D_Corruptions_AD. We hope\nthat our benchmarks and findings can provide insights for future research on\ndeveloping robust 3D object detection models.\n", "date": "2023-03-20", "categories": []} {"id": "2303.11243", "title": "Augment and Criticize: Exploring Informative Samples for Semi-Supervised\n Monocular 3D Object Detection", "abstract": " In this paper, we improve the challenging monocular 3D object detection\nproblem with a general semi-supervised framework. Specifically, having observed\nthat the bottleneck of this task lies in lacking reliable and informative\nsamples to train the detector, we introduce a novel, simple, yet effective\n`Augment and Criticize' framework that explores abundant informative samples\nfrom unlabeled data for learning more robust detection models. In the `Augment'\nstage, we present the Augmentation-based Prediction aGgregation (APG), which\naggregates detections from various automatically learned augmented views to\nimprove the robustness of pseudo label generation. Since not all pseudo labels\nfrom APG are beneficially informative, the subsequent `Criticize' phase is\npresented. In particular, we introduce the Critical Retraining Strategy (CRS)\nthat, unlike simply filtering pseudo labels using a fixed threshold (e.g.,\nclassification score) as in 2D semi-supervised tasks, leverages a learnable\nnetwork to evaluate the contribution of unlabeled images at different training\ntimestamps. This way, the noisy samples prohibitive to model evolution could be\neffectively suppressed. To validate our framework, we apply it to MonoDLE and\nMonoFlex. The two new detectors, dubbed 3DSeMo_DLE and 3DSeMo_FLEX, achieve\nstate-of-the-art results with remarkable improvements for over 3.5% AP_3D/BEV\n(Easy) on KITTI, showing its effectiveness and generality. Code and models will\nbe released.\n", "date": "2023-03-20", "categories": []} {"id": "2303.11301", "title": "VoxelNeXt: Fully Sparse VoxelNet for 3D Object Detection and Tracking", "abstract": " 3D object detectors usually rely on hand-crafted proxies, e.g., anchors or\ncenters, and translate well-studied 2D frameworks to 3D. Thus, sparse voxel\nfeatures need to be densified and processed by dense prediction heads, which\ninevitably costs extra computation. In this paper, we instead propose VoxelNext\nfor fully sparse 3D object detection. Our core insight is to predict objects\ndirectly based on sparse voxel features, without relying on hand-crafted\nproxies. Our strong sparse convolutional network VoxelNeXt detects and tracks\n3D objects through voxel features entirely. It is an elegant and efficient\nframework, with no need for sparse-to-dense conversion or NMS post-processing.\nOur method achieves a better speed-accuracy trade-off than other mainframe\ndetectors on the nuScenes dataset. For the first time, we show that a fully\nsparse voxel-based representation works decently for LIDAR 3D object detection\nand tracking. Extensive experiments on nuScenes, Waymo, and Argoverse2\nbenchmarks validate the effectiveness of our approach. Without bells and\nwhistles, our model outperforms all existing LIDAR methods on the nuScenes\ntracking test benchmark.\n", "date": "2023-03-20", "categories": []} {"id": "2303.09731", "title": "Exorcising ''Wraith'': Protecting LiDAR-based Object Detector in\n Automated Driving System from Appearing Attacks", "abstract": " Automated driving systems rely on 3D object detectors to recognize possible\nobstacles from LiDAR point clouds. However, recent works show the adversary can\nforge non-existent cars in the prediction results with a few fake points (i.e.,\nappearing attack). By removing statistical outliers, existing defenses are\nhowever designed for specific attacks or biased by predefined heuristic rules.\nTowards more comprehensive mitigation, we first systematically inspect the\nmechanism of recent appearing attacks: Their common weaknesses are observed in\ncrafting fake obstacles which (i) have obvious differences in the local parts\ncompared with real obstacles and (ii) violate the physical relation between\ndepth and point density. In this paper, we propose a novel plug-and-play\ndefensive module which works by side of a trained LiDAR-based object detector\nto eliminate forged obstacles where a major proportion of local parts have low\nobjectness, i.e., to what degree it belongs to a real object. At the core of\nour module is a local objectness predictor, which explicitly incorporates the\ndepth information to model the relation between depth and point density, and\npredicts each local part of an obstacle with an objectness score. Extensive\nexperiments show, our proposed defense eliminates at least 70% cars forged by\nthree known appearing attacks in most cases, while, for the best previous\ndefense, less than 30% forged cars are eliminated. Meanwhile, under the same\ncircumstance, our defense incurs less overhead for AP/precision on cars\ncompared with existing defenses. Furthermore, We validate the effectiveness of\nour proposed defense on simulation-based closed-loop control driving tests in\nthe open-source system of Baidu's Apollo.\n", "date": "2023-03-17", "categories": []} {"id": "2211.11950", "title": "UpCycling: Semi-supervised 3D Object Detection without Sharing Raw-level\n Unlabeled Scenes", "abstract": " Semi-supervised Learning (SSL) has received increasing attention in\nautonomous driving to reduce the enormous burden of 3D annotation. In this\npaper, we propose UpCycling, a novel SSL framework for 3D object detection with\nzero additional raw-level point cloud: learning from unlabeled de-identified\nintermediate features (i.e., smashed data) to preserve privacy. Since these\nintermediate features are naturally produced by the inference pipeline, no\nadditional computation is required on autonomous vehicles. However, generating\neffective consistency loss for unlabeled feature-level scene turns out to be a\ncritical challenge. The latest SSL frameworks for 3D object detection that\nenforce consistency regularization between different augmentations of an\nunlabeled raw-point scene become detrimental when applied to intermediate\nfeatures. To solve the problem, we introduce a novel combination of hybrid\npseudo labels and feature-level Ground Truth sampling (F-GT), which safely\naugments unlabeled multi-type 3D scene features and provides high-quality\nsupervision. We implement UpCycling on two representative 3D object detection\nmodels: SECOND-IoU and PV-RCNN. Experiments on widely-used datasets (Waymo,\nKITTI, and Lyft) verify that UpCycling outperforms other augmentation methods\napplied at the feature level. In addition, while preserving privacy, UpCycling\nperforms better or comparably to the state-of-the-art methods that utilize\nraw-level unlabeled data in both domain adaptation and partial-label scenarios.\n", "date": "2022-11-22", "categories": []} {"id": "2303.08316", "title": "MSF: Motion-guided Sequential Fusion for Efficient 3D Object Detection\n from Point Cloud Sequences", "abstract": " Point cloud sequences are commonly used to accurately detect 3D objects in\napplications such as autonomous driving. Current top-performing multi-frame\ndetectors mostly follow a Detect-and-Fuse framework, which extracts features\nfrom each frame of the sequence and fuses them to detect the objects in the\ncurrent frame. However, this inevitably leads to redundant computation since\nadjacent frames are highly correlated. In this paper, we propose an efficient\nMotion-guided Sequential Fusion (MSF) method, which exploits the continuity of\nobject motion to mine useful sequential contexts for object detection in the\ncurrent frame. We first generate 3D proposals on the current frame and\npropagate them to preceding frames based on the estimated velocities. The\npoints-of-interest are then pooled from the sequence and encoded as proposal\nfeatures. A novel Bidirectional Feature Aggregation (BiFA) module is further\nproposed to facilitate the interactions of proposal features across frames.\nBesides, we optimize the point cloud pooling by a voxel-based sampling\ntechnique so that millions of points can be processed in several milliseconds.\nThe proposed MSF method achieves not only better efficiency than other\nmulti-frame detectors but also leading accuracy, with 83.12% and 78.30% mAP on\nthe LEVEL1 and LEVEL2 test sets of Waymo Open Dataset, respectively. Codes can\nbe found at \\url{https://github.com/skyhehe123/MSF}.\n", "date": "2023-03-15", "categories": []} {"id": "2303.08686", "title": "Weakly Supervised Monocular 3D Object Detection using Multi-View\n Projection and Direction Consistency", "abstract": " Monocular 3D object detection has become a mainstream approach in automatic\ndriving for its easy application. A prominent advantage is that it does not\nneed LiDAR point clouds during the inference. However, most current methods\nstill rely on 3D point cloud data for labeling the ground truths used in the\ntraining phase. This inconsistency between the training and inference makes it\nhard to utilize the large-scale feedback data and increases the data collection\nexpenses. To bridge this gap, we propose a new weakly supervised monocular 3D\nobjection detection method, which can train the model with only 2D labels\nmarked on images. To be specific, we explore three types of consistency in this\ntask, i.e. the projection, multi-view and direction consistency, and design a\nweakly-supervised architecture based on these consistencies. Moreover, we\npropose a new 2D direction labeling method in this task to guide the model for\naccurate rotation direction prediction. Experiments show that our\nweakly-supervised method achieves comparable performance with some fully\nsupervised methods. When used as a pre-training method, our model can\nsignificantly outperform the corresponding fully-supervised baseline with only\n1/3 3D labels. https://github.com/weakmono3d/weakmono3d\n", "date": "2023-03-15", "categories": []} {"id": "2303.03595", "title": "LoGoNet: Towards Accurate 3D Object Detection with Local-to-Global\n Cross-Modal Fusion", "abstract": " LiDAR-camera fusion methods have shown impressive performance in 3D object\ndetection. Recent advanced multi-modal methods mainly perform global fusion,\nwhere image features and point cloud features are fused across the whole scene.\nSuch practice lacks fine-grained region-level information, yielding suboptimal\nfusion performance. In this paper, we present the novel Local-to-Global fusion\nnetwork (LoGoNet), which performs LiDAR-camera fusion at both local and global\nlevels. Concretely, the Global Fusion (GoF) of LoGoNet is built upon previous\nliterature, while we exclusively use point centroids to more precisely\nrepresent the position of voxel features, thus achieving better cross-modal\nalignment. As to the Local Fusion (LoF), we first divide each proposal into\nuniform grids and then project these grid centers to the images. The image\nfeatures around the projected grid points are sampled to be fused with\nposition-decorated point cloud features, maximally utilizing the rich\ncontextual information around the proposals. The Feature Dynamic Aggregation\n(FDA) module is further proposed to achieve information interaction between\nthese locally and globally fused features, thus producing more informative\nmulti-modal features. Extensive experiments on both Waymo Open Dataset (WOD)\nand KITTI datasets show that LoGoNet outperforms all state-of-the-art 3D\ndetection methods. Notably, LoGoNet ranks 1st on Waymo 3D object detection\nleaderboard and obtains 81.02 mAPH (L2) detection performance. It is noteworthy\nthat, for the first time, the detection performance on three classes surpasses\n80 APH (L2) simultaneously. Code will be available at\n\\url{https://github.com/sankin97/LoGoNet}.\n", "date": "2023-03-07", "categories": []} {"id": "2303.06250", "title": "ReBound: An Open-Source 3D Bounding Box Annotation Tool for Active\n Learning", "abstract": " In recent years, supervised learning has become the dominant paradigm for\ntraining deep-learning based methods for 3D object detection. Lately, the\nacademic community has studied 3D object detection in the context of autonomous\nvehicles (AVs) using publicly available datasets such as nuScenes and Argoverse\n2.0. However, these datasets may have incomplete annotations, often only\nlabeling a small subset of objects in a scene. Although commercial services\nexists for 3D bounding box annotation, these are often prohibitively expensive.\nTo address these limitations, we propose ReBound, an open-source 3D\nvisualization and dataset re-annotation tool that works across different\ndatasets. In this paper, we detail the design of our tool and present survey\nresults that highlight the usability of our software. Further, we show that\nReBound is effective for exploratory data analysis and can facilitate\nactive-learning. Our code and documentation is available at\nhttps://github.com/ajedgley/ReBound\n", "date": "2023-03-11", "categories": []} {"id": "2303.06342", "title": "Enhanced K-Radar: Optimal Density Reduction to Improve Detection\n Performance and Accessibility of 4D Radar Tensor-based Object Detection", "abstract": " Recent works have shown the superior robustness of four-dimensional (4D)\nRadar-based three-dimensional (3D) object detection in adverse weather\nconditions. However, processing 4D Radar data remains a challenge due to the\nlarge data size, which require substantial amount of memory for computing and\nstorage. In previous work, an online density reduction is performed on the 4D\nRadar Tensor (4DRT) to reduce the data size, in which the density reduction\nlevel is chosen arbitrarily. However, the impact of density reduction on the\ndetection performance and memory consumption remains largely unknown. In this\npaper, we aim to address this issue by conducting extensive hyperparamter\ntuning on the density reduction level. Experimental results show that\nincreasing the density level from 0.01% to 50% of the original 4DRT density\nlevel proportionally improves the detection performance, at a cost of memory\nconsumption. However, when the density level is increased beyond 5%, only the\nmemory consumption increases, while the detection performance oscillates below\nthe peak point. In addition to the optimized density hyperparameter, we also\nintroduce 4D Sparse Radar Tensor (4DSRT), a new representation for 4D Radar\ndata with offline density reduction, leading to a significantly reduced raw\ndata size. An optimized development kit for training the neural networks is\nalso provided, which along with the utilization of 4DSRT, improves training\nspeed by a factor of 17.1 compared to the state-of-the-art 4DRT-based neural\nnetworks. All codes are available at: https://github.com/kaist-avelab/K-Radar.\n", "date": "2023-03-11", "categories": []} {"id": "2303.05079", "title": "DDS3D: Dense Pseudo-Labels with Dynamic Threshold for Semi-Supervised 3D\n Object Detection", "abstract": " In this paper, we present a simple yet effective semi-supervised 3D object\ndetector named DDS3D. Our main contributions have two-fold. On the one hand,\ndifferent from previous works using Non-Maximal Suppression (NMS) or its\nvariants for obtaining the sparse pseudo labels, we propose a dense\npseudo-label generation strategy to get dense pseudo-labels, which can retain\nmore potential supervision information for the student network. On the other\nhand, instead of traditional fixed thresholds, we propose a dynamic threshold\nmanner to generate pseudo-labels, which can guarantee the quality and quantity\nof pseudo-labels during the whole training process. Benefiting from these two\ncomponents, our DDS3D outperforms the state-of-the-art semi-supervised 3d\nobject detection with mAP of 3.1% on the pedestrian and 2.1% on the cyclist\nunder the same configuration of 1% samples. Extensive ablation studies on the\nKITTI dataset demonstrate the effectiveness of our DDS3D. The code and models\nwill be made publicly available at https://github.com/hust-jy/DDS3D\n", "date": "2023-03-09", "categories": []} {"id": "2303.05886", "title": "Bi3D: Bi-domain Active Learning for Cross-domain 3D Object Detection", "abstract": " Unsupervised Domain Adaptation (UDA) technique has been explored in 3D\ncross-domain tasks recently. Though preliminary progress has been made, the\nperformance gap between the UDA-based 3D model and the supervised one trained\nwith fully annotated target domain is still large. This motivates us to\nconsider selecting partial-yet-important target data and labeling them at a\nminimum cost, to achieve a good trade-off between high performance and low\nannotation cost. To this end, we propose a Bi-domain active learning approach,\nnamely Bi3D, to solve the cross-domain 3D object detection task. The Bi3D first\ndevelops a domainness-aware source sampling strategy, which identifies\ntarget-domain-like samples from the source domain to avoid the model being\ninterfered by irrelevant source data. Then a diversity-based target sampling\nstrategy is developed, which selects the most informative subset of target\ndomain to improve the model adaptability to the target domain using as little\nannotation budget as possible. Experiments are conducted on typical\ncross-domain adaptation scenarios including cross-LiDAR-beam, cross-country,\nand cross-sensor, where Bi3D achieves a promising target-domain detection\naccuracy (89.63% on KITTI) compared with UDAbased work (84.29%), even\nsurpassing the detector trained on the full set of the labeled target domain\n(88.98%). Our code is available at: https://github.com/PJLabADG/3DTrans.\n", "date": "2023-03-10", "categories": []} {"id": "2302.06650", "title": "Surround-View Vision-based 3D Detection for Autonomous Driving: A Survey", "abstract": " Vision-based 3D Detection task is fundamental task for the perception of an\nautonomous driving system, which has peaked interest amongst many researchers\nand autonomous driving engineers. However achieving a rather good 3D BEV\n(Bird's Eye View) performance is not an easy task using 2D sensor input-data\nwith cameras. In this paper we provide a literature survey for the existing\nVision Based 3D detection methods, focused on autonomous driving. We have made\ndetailed analysis of over $60$ papers leveraging Vision BEV detections\napproaches and highlighted different sub-groups for detailed understanding of\ncommon trends. Moreover, we have highlighted how the literature and industry\ntrend have moved towards surround-view image based methods and note down\nthoughts on what special cases this method addresses. In conclusion, we provoke\nthoughts of 3D Vision techniques for future research based on shortcomings of\nthe current techniques including the direction of collaborative perception.\n", "date": "2023-02-13", "categories": []} {"id": "2303.02203", "title": "X$^3$KD: Knowledge Distillation Across Modalities, Tasks and Stages for\n Multi-Camera 3D Object Detection", "abstract": " Recent advances in 3D object detection (3DOD) have obtained remarkably strong\nresults for LiDAR-based models. In contrast, surround-view 3DOD models based on\nmultiple camera images underperform due to the necessary view transformation of\nfeatures from perspective view (PV) to a 3D world representation which is\nambiguous due to missing depth information. This paper introduces X$^3$KD, a\ncomprehensive knowledge distillation framework across different modalities,\ntasks, and stages for multi-camera 3DOD. Specifically, we propose cross-task\ndistillation from an instance segmentation teacher (X-IS) in the PV feature\nextraction stage providing supervision without ambiguous error backpropagation\nthrough the view transformation. After the transformation, we apply cross-modal\nfeature distillation (X-FD) and adversarial training (X-AT) to improve the 3D\nworld representation of multi-camera features through the information contained\nin a LiDAR-based 3DOD teacher. Finally, we also employ this teacher for\ncross-modal output distillation (X-OD), providing dense supervision at the\nprediction stage. We perform extensive ablations of knowledge distillation at\ndifferent stages of multi-camera 3DOD. Our final X$^3$KD model outperforms\nprevious state-of-the-art approaches on the nuScenes and Waymo datasets and\ngeneralizes to RADAR-based 3DOD. Qualitative results video at\nhttps://youtu.be/1do9DPFmr38.\n", "date": "2023-03-03", "categories": []} {"id": "2303.02314", "title": "Virtual Sparse Convolution for Multimodal 3D Object Detection", "abstract": " Recently, virtual/pseudo-point-based 3D object detection that seamlessly\nfuses RGB images and LiDAR data by depth completion has gained great attention.\nHowever, virtual points generated from an image are very dense, introducing a\nhuge amount of redundant computation during detection. Meanwhile, noises\nbrought by inaccurate depth completion significantly degrade detection\nprecision. This paper proposes a fast yet effective backbone, termed\nVirConvNet, based on a new operator VirConv (Virtual Sparse Convolution), for\nvirtual-point-based 3D object detection. VirConv consists of two key designs:\n(1) StVD (Stochastic Voxel Discard) and (2) NRConv (Noise-Resistant Submanifold\nConvolution). StVD alleviates the computation problem by discarding large\namounts of nearby redundant voxels. NRConv tackles the noise problem by\nencoding voxel features in both 2D image and 3D LiDAR space. By integrating\nVirConv, we first develop an efficient pipeline VirConv-L based on an early\nfusion design. Then, we build a high-precision pipeline VirConv-T based on a\ntransformed refinement scheme. Finally, we develop a semi-supervised pipeline\nVirConv-S based on a pseudo-label framework. On the KITTI car 3D detection test\nleaderboard, our VirConv-L achieves 85% AP with a fast running speed of 56ms.\nOur VirConv-T and VirConv-S attains a high-precision of 86.3% and 87.2% AP, and\ncurrently rank 2nd and 1st, respectively. The code is available at\nhttps://github.com/hailanyi/VirConv.\n", "date": "2023-03-04", "categories": []} {"id": "2209.03102", "title": "MSMDFusion: Fusing LiDAR and Camera at Multiple Scales with Multi-Depth\n Seeds for 3D Object Detection", "abstract": " Fusing LiDAR and camera information is essential for achieving accurate and\nreliable 3D object detection in autonomous driving systems. This is challenging\ndue to the difficulty of combining multi-granularity geometric and semantic\nfeatures from two drastically different modalities. Recent approaches aim at\nexploring the semantic densities of camera features through lifting points in\n2D camera images (referred to as seeds) into 3D space, and then incorporate 2D\nsemantics via cross-modal interaction or fusion techniques. However, depth\ninformation is under-investigated in these approaches when lifting points into\n3D space, thus 2D semantics can not be reliably fused with 3D points. Moreover,\ntheir multi-modal fusion strategy, which is implemented as concatenation or\nattention, either can not effectively fuse 2D and 3D information or is unable\nto perform fine-grained interactions in the voxel space. To this end, we\npropose a novel framework with better utilization of the depth information and\nfine-grained cross-modal interaction between LiDAR and camera, which consists\nof two important components. First, a Multi-Depth Unprojection (MDU) method\nwith depth-aware designs is used to enhance the depth quality of the lifted\npoints at each interaction level. Second, a Gated Modality-Aware Convolution\n(GMA-Conv) block is applied to modulate voxels involved with the camera\nmodality in a fine-grained manner and then aggregate multi-modal features into\na unified space. Together they provide the detection head with more\ncomprehensive features from LiDAR and camera. On the nuScenes test benchmark,\nour proposed method, abbreviated as MSMDFusion, achieves state-of-the-art 3D\nobject detection results with 71.5% mAP and 74.0% NDS, and strong tracking\nresults with 74.0% AMOTA without using test-time-augmentation and ensemble\ntechniques. The code is available at https://github.com/SxJyJay/MSMDFusion.\n", "date": "2022-09-07", "categories": []} {"id": "2303.01686", "title": "Towards Domain Generalization for Multi-view 3D Object Detection in\n Bird-Eye-View", "abstract": " Multi-view 3D object detection (MV3D-Det) in Bird-Eye-View (BEV) has drawn\nextensive attention due to its low cost and high efficiency. Although new\nalgorithms for camera-only 3D object detection have been continuously proposed,\nmost of them may risk drastic performance degradation when the domain of input\nimages differs from that of training. In this paper, we first analyze the\ncauses of the domain gap for the MV3D-Det task. Based on the covariate shift\nassumption, we find that the gap mainly attributes to the feature distribution\nof BEV, which is determined by the quality of both depth estimation and 2D\nimage's feature representation. To acquire a robust depth prediction, we\npropose to decouple the depth estimation from the intrinsic parameters of the\ncamera (i.e. the focal length) through converting the prediction of metric\ndepth to that of scale-invariant depth and perform dynamic perspective\naugmentation to increase the diversity of the extrinsic parameters (i.e. the\ncamera poses) by utilizing homography. Moreover, we modify the focal length\nvalues to create multiple pseudo-domains and construct an adversarial training\nloss to encourage the feature representation to be more domain-agnostic.\nWithout bells and whistles, our approach, namely DG-BEV, successfully\nalleviates the performance drop on the unseen target domain without impairing\nthe accuracy of the source domain. Extensive experiments on various public\ndatasets, including Waymo, nuScenes, and Lyft, demonstrate the generalization\nand effectiveness of our approach. To the best of our knowledge, this is the\nfirst systematic study to explore a domain generalization method for MV3D-Det.\n", "date": "2023-03-03", "categories": []} {"id": "2303.01899", "title": "Quantifying the LiDAR Sim-to-Real Domain Shift: A Detailed Investigation\n Using Object Detectors and Analyzing Point Clouds at Target-Level", "abstract": " LiDAR object detection algorithms based on neural networks for autonomous\ndriving require large amounts of data for training, validation, and testing. As\nreal-world data collection and labeling are time-consuming and expensive,\nsimulation-based synthetic data generation is a viable alternative. However,\nusing simulated data for the training of neural networks leads to a domain\nshift of training and testing data due to differences in scenes, scenarios, and\ndistributions. In this work, we quantify the sim-to-real domain shift by means\nof LiDAR object detectors trained with a new scenario-identical real-world and\nsimulated dataset. In addition, we answer the questions of how well the\nsimulated data resembles the real-world data and how well object detectors\ntrained on simulated data perform on real-world data. Further, we analyze point\nclouds at the target-level by comparing real-world and simulated point clouds\nwithin the 3D bounding boxes of the targets. Our experiments show that a\nsignificant sim-to-real domain shift exists even for our scenario-identical\ndatasets. This domain shift amounts to an average precision reduction of around\n14 % for object detectors trained with simulated data. Additional experiments\nreveal that this domain shift can be lowered by introducing a simple noise\nmodel in simulation. We further show that a simple downsampling method to model\nreal-world physics does not influence the performance of the object detectors.\n", "date": "2023-03-03", "categories": []} {"id": "2303.02000", "title": "BSH-Det3D: Improving 3D Object Detection with BEV Shape Heatmap", "abstract": " The progress of LiDAR-based 3D object detection has significantly enhanced\ndevelopments in autonomous driving and robotics. However, due to the\nlimitations of LiDAR sensors, object shapes suffer from deterioration in\noccluded and distant areas, which creates a fundamental challenge to 3D\nperception. Existing methods estimate specific 3D shapes and achieve remarkable\nperformance. However, these methods rely on extensive computation and memory,\ncausing imbalances between accuracy and real-time performance. To tackle this\nchallenge, we propose a novel LiDAR-based 3D object detection model named\nBSH-Det3D, which applies an effective way to enhance spatial features by\nestimating complete shapes from a bird's eye view (BEV). Specifically, we\ndesign the Pillar-based Shape Completion (PSC) module to predict the\nprobability of occupancy whether a pillar contains object shapes. The PSC\nmodule generates a BEV shape heatmap for each scene. After integrating with\nheatmaps, BSH-Det3D can provide additional information in shape deterioration\nareas and generate high-quality 3D proposals. We also design an attention-based\ndensification fusion module (ADF) to adaptively associate the sparse features\nwith heatmaps and raw points. The ADF module integrates the advantages of\npoints and shapes knowledge with negligible overheads. Extensive experiments on\nthe KITTI benchmark achieve state-of-the-art (SOTA) performance in terms of\naccuracy and speed, demonstrating the efficiency and flexibility of BSH-Det3D.\nThe source code is available on https://github.com/mystorm16/BSH-Det3D.\n", "date": "2023-03-03", "categories": []} {"id": "2207.12988", "title": "Monocular 3D Object Detection with Depth from Motion", "abstract": " Perceiving 3D objects from monocular inputs is crucial for robotic systems,\ngiven its economy compared to multi-sensor settings. It is notably difficult as\na single image can not provide any clues for predicting absolute depth values.\nMotivated by binocular methods for 3D object detection, we take advantage of\nthe strong geometry structure provided by camera ego-motion for accurate object\ndepth estimation and detection. We first make a theoretical analysis on this\ngeneral two-view case and notice two challenges: 1) Cumulative errors from\nmultiple estimations that make the direct prediction intractable; 2) Inherent\ndilemmas caused by static cameras and matching ambiguity. Accordingly, we\nestablish the stereo correspondence with a geometry-aware cost volume as the\nalternative for depth estimation and further compensate it with monocular\nunderstanding to address the second problem. Our framework, named Depth from\nMotion (DfM), then uses the established geometry to lift 2D image features to\nthe 3D space and detects 3D objects thereon. We also present a pose-free DfM to\nmake it usable when the camera pose is unavailable. Our framework outperforms\nstate-of-the-art methods by a large margin on the KITTI benchmark. Detailed\nquantitative and qualitative analyses also validate our theoretical\nconclusions. The code will be released at\nhttps://github.com/Tai-Wang/Depth-from-Motion.\n", "date": "2022-07-26", "categories": []} {"id": "2302.14522", "title": "AdaptiveShape: Solving Shape Variability for 3D Object Detection with\n Geometry Aware Anchor Distributions", "abstract": " 3D object detection with point clouds and images plays an important role in\nperception tasks such as autonomous driving. Current methods show great\nperformance on detection and pose estimation of standard-shaped vehicles but\nlack behind on more complex shapes as e.g. semi-trailer truck combinations.\nDetermining the shape and motion of those special vehicles accurately is\ncrucial in yard operation and maneuvering and industrial automation\napplications. This work introduces several new methods to improve and measure\nthe performance for such classes. State-of-the-art methods are based on\npredefined anchor grids or heatmaps for ground truth targets. However, the\nunderlying representations do not take the shape of different sized objects\ninto account. Our main contribution, AdaptiveShape, uses shape aware anchor\ndistributions and heatmaps to improve the detection capabilities. For large\nvehicles we achieve +10.9% AP in comparison to current shape agnostic methods.\nFurthermore we introduce a new fast LiDAR-camera fusion. It is based on 2D\nbounding box camera detections which are available in many processing\npipelines. This fusion method does not rely on perfectly calibrated or\ntemporally synchronized systems and is therefore applicable to a broad range of\nrobotic applications. We extend a standard point pillar network to account for\ntemporal data and improve learning of complex object movements. In addition we\nextended a ground truth augmentation to use grouped object pairs to further\nimprove truck AP by +2.2% compared to conventional augmentation.\n", "date": "2023-02-28", "categories": []} {"id": "2206.07634", "title": "Real3D-Aug: Point Cloud Augmentation by Placing Real Objects with\n Occlusion Handling for 3D Detection and Segmentation", "abstract": " Object detection and semantic segmentation with the 3D lidar point cloud data\nrequire expensive annotation. We propose a data augmentation method that takes\nadvantage of already annotated data multiple times. We propose an augmentation\nframework that reuses real data, automatically finds suitable placements in the\nscene to be augmented, and handles occlusions explicitly. Due to the usage of\nthe real data, the scan points of newly inserted objects in augmentation\nsustain the physical characteristics of the lidar, such as intensity and\nraydrop. The pipeline proves competitive in training top-performing models for\n3D object detection and semantic segmentation. The new augmentation provides a\nsignificant performance gain in rare and essential classes, notably 6.65%\naverage precision gain for \"Hard\" pedestrian class in KITTI object detection or\n2.14 mean IoU gain in the SemanticKITTI segmentation challenge over the state\nof the art.\n", "date": "2022-06-15", "categories": []} {"id": "2302.13301", "title": "Pillar R-CNN for Point Cloud 3D Object Detection", "abstract": " The performance of point cloud 3D object detection hinges on effectively\nrepresenting raw points, grid-based voxels or pillars. Recent two-stage 3D\ndetectors typically take the point-voxel-based R-CNN paradigm, i.e., the first\nstage resorts to the 3D voxel-based backbone for 3D proposal generation on\nbird-eye-view (BEV) representation and the second stage refines them via the\nintermediate point representation. Their primary mechanisms involve the\nutilization of intermediary keypoints to restore the substantial 3D structure\ncontext from the converted BEV representation. The skilled point-voxel feature\ninteraction, however, makes the entire detection pipeline more complex and\ncompute-intensive. In this paper, we take a different viewpoint -- the\npillar-based BEV representation owns sufficient capacity to preserve the 3D\nstructure. In light of the latest advances in BEV-based perception, we devise a\nconceptually simple yet effective two-stage 3D detection architecture, named\nPillar R-CNN. On top of densified BEV feature maps, Pillar R-CNN can easily\nintroduce the feature pyramid architecture to generate 3D proposals at various\nscales and take the simple 2D R-CNN style detect head for box refinement. Our\nPillar R-CNN performs favorably against state-of-the-art 3D detectors on the\nlarge-scale Waymo Open Dataset but at a small extra cost. It should be\nhighlighted that further exploration into BEV perception for applications\ninvolving autonomous driving is now possible thanks to the effective and\nelegant Pillar R-CNN architecture.\n", "date": "2023-02-26", "categories": []} {"id": "2302.13577", "title": "DuEqNet: Dual-Equivariance Network in Outdoor 3D Object Detection for\n Autonomous Driving", "abstract": " Outdoor 3D object detection has played an essential role in the environment\nperception of autonomous driving. In complicated traffic situations, precise\nobject recognition provides indispensable information for prediction and\nplanning in the dynamic system, improving self-driving safety and reliability.\nHowever, with the vehicle's veering, the constant rotation of the surrounding\nscenario makes a challenge for the perception systems. Yet most existing\nmethods have not focused on alleviating the detection accuracy impairment\nbrought by the vehicle's rotation, especially in outdoor 3D detection. In this\npaper, we propose DuEqNet, which first introduces the concept of equivariance\ninto 3D object detection network by leveraging a hierarchical embedded\nframework. The dual-equivariance of our model can extract the equivariant\nfeatures at both local and global levels, respectively. For the local feature,\nwe utilize the graph-based strategy to guarantee the equivariance of the\nfeature in point cloud pillars. In terms of the global feature, the group\nequivariant convolution layers are adopted to aggregate the local feature to\nachieve the global equivariance. In the experiment part, we evaluate our\napproach with different baselines in 3D object detection tasks and obtain\nState-Of-The-Art performance. According to the results, our model presents\nhigher accuracy on orientation and better prediction efficiency. Moreover, our\ndual-equivariance strategy exhibits the satisfied plug-and-play ability on\nvarious popular object detection frameworks to improve their performance.\n", "date": "2023-02-27", "categories": []} {"id": "2302.11481", "title": "Transformer-Based Sensor Fusion for Autonomous Driving: A Survey", "abstract": " Sensor fusion is an essential topic in many perception systems, such as\nautonomous driving and robotics. Transformers-based detection head and\nCNN-based feature encoder to extract features from raw sensor-data has emerged\nas one of the best performing sensor-fusion 3D-detection-framework, according\nto the dataset leaderboards. In this work we provide an in-depth literature\nsurvey of transformer based 3D-object detection task in the recent past,\nprimarily focusing on the sensor fusion. We also briefly go through the Vision\ntransformers (ViT) basics, so that readers can easily follow through the paper.\nMoreover, we also briefly go through few of the non-transformer based\nless-dominant methods for sensor fusion for autonomous driving. In conclusion\nwe summarize with sensor-fusion trends to follow and provoke future research.\nMore updated summary can be found at:\nhttps://github.com/ApoorvRoboticist/Transformers-Sensor-Fusion\n", "date": "2023-02-22", "categories": []} {"id": "2302.10465", "title": "A Flexible Multi-view Multi-modal Imaging System for Outdoor Scenes", "abstract": " Multi-view imaging systems enable uniform coverage of 3D space and reduce the\nimpact of occlusion, which is beneficial for 3D object detection and tracking\naccuracy. However, existing imaging systems built with multi-view cameras or\ndepth sensors are limited by the small applicable scene and complicated\ncomposition. In this paper, we propose a wireless multi-view multi-modal 3D\nimaging system generally applicable to large outdoor scenes, which consists of\na master node and several slave nodes. Multiple spatially distributed slave\nnodes equipped with cameras and LiDARs are connected to form a wireless sensor\nnetwork. While providing flexibility and scalability, the system applies\nautomatic spatio-temporal calibration techniques to obtain accurate 3D\nmulti-view multi-modal data. This system is the first imaging system that\nintegrates mutli-view RGB cameras and LiDARs in large outdoor scenes among\nexisting 3D imaging systems. We perform point clouds based 3D object detection\nand long-term tracking using the 3D imaging dataset collected by this system.\nThe experimental results show that multi-view point clouds greatly improve 3D\nobject detection and tracking accuracy regardless of complex and various\noutdoor environments.\n", "date": "2023-02-21", "categories": []} {"id": "2302.10511", "title": "MVFusion: Multi-View 3D Object Detection with Semantic-aligned Radar and\n Camera Fusion", "abstract": " Multi-view radar-camera fused 3D object detection provides a farther\ndetection range and more helpful features for autonomous driving, especially\nunder adverse weather. The current radar-camera fusion methods deliver kinds of\ndesigns to fuse radar information with camera data. However, these fusion\napproaches usually adopt the straightforward concatenation operation between\nmulti-modal features, which ignores the semantic alignment with radar features\nand sufficient correlations across modals. In this paper, we present MVFusion,\na novel Multi-View radar-camera Fusion method to achieve semantic-aligned radar\nfeatures and enhance the cross-modal information interaction. To achieve so, we\ninject the semantic alignment into the radar features via the semantic-aligned\nradar encoder (SARE) to produce image-guided radar features. Then, we propose\nthe radar-guided fusion transformer (RGFT) to fuse our radar and image features\nto strengthen the two modals' correlation from the global scope via the\ncross-attention mechanism. Extensive experiments show that MVFusion achieves\nstate-of-the-art performance (51.7% NDS and 45.3% mAP) on the nuScenes dataset.\nWe shall release our code and trained networks upon publication.\n", "date": "2023-02-21", "categories": []} {"id": "2302.10549", "title": "MonoPGC: Monocular 3D Object Detection with Pixel Geometry Contexts", "abstract": " Monocular 3D object detection reveals an economical but challenging task in\nautonomous driving. Recently center-based monocular methods have developed\nrapidly with a great trade-off between speed and accuracy, where they usually\ndepend on the object center's depth estimation via 2D features. However, the\nvisual semantic features without sufficient pixel geometry information, may\naffect the performance of clues for spatial 3D detection tasks. To alleviate\nthis, we propose MonoPGC, a novel end-to-end Monocular 3D object detection\nframework with rich Pixel Geometry Contexts. We introduce the pixel depth\nestimation as our auxiliary task and design depth cross-attention pyramid\nmodule (DCPM) to inject local and global depth geometry knowledge into visual\nfeatures. In addition, we present the depth-space-aware transformer (DSAT) to\nintegrate 3D space position and depth-aware features efficiently. Besides, we\ndesign a novel depth-gradient positional encoding (DGPE) to bring more distinct\npixel geometry contexts into the transformer for better object detection.\nExtensive experiments demonstrate that our method achieves the state-of-the-art\nperformance on the KITTI dataset.\n", "date": "2023-02-21", "categories": []} {"id": "2211.13529", "title": "3D Dual-Fusion: Dual-Domain Dual-Query Camera-LiDAR Fusion for 3D Object\n Detection", "abstract": " Fusing data from cameras and LiDAR sensors is an essential technique to\nachieve robust 3D object detection. One key challenge in camera-LiDAR fusion\ninvolves mitigating the large domain gap between the two sensors in terms of\ncoordinates and data distribution when fusing their features. In this paper, we\npropose a novel camera-LiDAR fusion architecture called, 3D Dual-Fusion, which\nis designed to mitigate the gap between the feature representations of camera\nand LiDAR data. The proposed method fuses the features of the camera-view and\n3D voxel-view domain and models their interactions through deformable\nattention. We redesign the transformer fusion encoder to aggregate the\ninformation from the two domains. Two major changes include 1) dual query-based\ndeformable attention to fuse the dual-domain features interactively and 2) 3D\nlocal self-attention to encode the voxel-domain queries prior to dual-query\ndecoding. The results of an experimental evaluation show that the proposed\ncamera-LiDAR fusion architecture achieved competitive performance on the KITTI\nand nuScenes datasets, with state-of-the-art performances in some 3D object\ndetection benchmarks categories.\n", "date": "2022-11-24", "categories": []} {"id": "2301.12519", "title": "3D Object Detection in LiDAR Point Clouds using Graph Neural Networks", "abstract": " LiDAR (Light Detection and Ranging) is an advanced active remote sensing\ntechnique working on the principle of time of travel (ToT) for capturing highly\naccurate 3D information of the surroundings. LiDAR has gained wide attention in\nresearch and development with the LiDAR industry expected to reach 2.8 billion\n$ by 2025. Although the LiDAR dataset is of rich density and high spatial\nresolution, it is challenging to process LiDAR data due to its inherent 3D\ngeometry and massive volume. But such a high-resolution dataset possesses\nimmense potential in many applications and has great potential in 3D object\ndetection and recognition. In this research we propose Graph Neural Network\n(GNN) based framework to learn and identify the objects in the 3D LiDAR point\nclouds. GNNs are class of deep learning which learns the patterns and objects\nbased on the principle of graph learning which have shown success in various 3D\ncomputer vision tasks.\n", "date": "2023-01-29", "categories": []} {"id": "2302.06643", "title": "Vision-RADAR fusion for Robotics BEV Detections: A Survey", "abstract": " Due to the trending need of building autonomous robotic perception system,\nsensor fusion has attracted a lot of attention amongst researchers and\nengineers to make best use of cross-modality information. However, in order to\nbuild a robotic platform at scale we need to emphasize on autonomous robot\nplatform bring-up cost as well. Cameras and radars, which inherently includes\ncomplementary perception information, has potential for developing autonomous\nrobotic platform at scale. However, there is a limited work around radar fused\nwith Vision, compared to LiDAR fused with vision work. In this paper, we tackle\nthis gap with a survey on Vision-Radar fusion approaches for a BEV object\ndetection system. First we go through the background information viz., object\ndetection tasks, choice of sensors, sensor setup, benchmark datasets and\nevaluation metrics for a robotic perception system. Later, we cover\nper-modality (Camera and RADAR) data representation, then we go into detail\nabout sensor fusion techniques based on sub-groups viz., early-fusion,\ndeep-fusion, and late-fusion to easily understand the pros and cons of each\nmethod. Finally, we propose possible future trends for vision-radar fusion to\nenlighten future research. Regularly updated summary can be found at:\nhttps://github.com/ApoorvRoboticist/Vision-RADAR-Fusion-BEV-Survey\n", "date": "2023-02-13", "categories": []} {"id": "2209.15215", "title": "INT: Towards Infinite-frames 3D Detection with An Efficient Framework", "abstract": " It is natural to construct a multi-frame instead of a single-frame 3D\ndetector for a continuous-time stream. Although increasing the number of frames\nmight improve performance, previous multi-frame studies only used very limited\nframes to build their systems due to the dramatically increased computational\nand memory cost. To address these issues, we propose a novel on-stream training\nand prediction framework that, in theory, can employ an infinite number of\nframes while keeping the same amount of computation as a single-frame detector.\nThis infinite framework (INT), which can be used with most existing detectors,\nis utilized, for example, on the popular CenterPoint, with significant latency\nreductions and performance improvements. We've also conducted extensive\nexperiments on two large-scale datasets, nuScenes and Waymo Open Dataset, to\ndemonstrate the scheme's effectiveness and efficiency. By employing INT on\nCenterPoint, we can get around 7% (Waymo) and 15% (nuScenes) performance boost\nwith only 2~4ms latency overhead, and currently SOTA on the Waymo 3D Detection\nleaderboard.\n", "date": "2022-09-30", "categories": []} {"id": "2212.01528", "title": "IDMS: Instance Depth for Multi-scale Monocular 3D Object Detection", "abstract": " Due to the lack of depth information of images and poor detection accuracy in\nmonocular 3D object detection, we proposed the instance depth for multi-scale\nmonocular 3D object detection method. Firstly, to enhance the model's\nprocessing ability for different scale targets, a multi-scale perception module\nbased on dilated convolution is designed, and the depth features containing\nmulti-scale information are re-refined from both spatial and channel directions\nconsidering the inconsistency between feature maps of different scales.\nFirstly, we designed a multi-scale perception module based on dilated\nconvolution to enhance the model's processing ability for different scale\ntargets. The depth features containing multi-scale information are re-refined\nfrom spatial and channel directions considering the inconsistency between\nfeature maps of different scales. Secondly, so as to make the model obtain\nbetter 3D perception, this paper proposed to use the instance depth information\nas an auxiliary learning task to enhance the spatial depth feature of the 3D\ntarget and use the sparse instance depth to supervise the auxiliary task.\nFinally, by verifying the proposed algorithm on the KITTI test set and\nevaluation set, the experimental results show that compared with the baseline\nmethod, the proposed method improves by 5.27\\% in AP40 in the car category,\neffectively improving the detection performance of the monocular 3D object\ndetection algorithm.\n", "date": "2022-12-03", "categories": []} {"id": "2302.06291", "title": "Surface-biased Multi-Level Context 3D Object Detection", "abstract": " Object detection in 3D point clouds is a crucial task in a range of computer\nvision applications including robotics, autonomous cars, and augmented reality.\nThis work addresses the object detection task in 3D point clouds using a highly\nefficient, surface-biased, feature extraction method (wang2022rbgnet), that\nalso captures contextual cues on multiple levels. We propose a 3D object\ndetector that extracts accurate feature representations of object candidates\nand leverages self-attention on point patches, object candidates, and on the\nglobal scene in 3D scene. Self-attention is proven to be effective in encoding\ncorrelation information in 3D point clouds by (xie2020mlcvnet). While other 3D\ndetectors focus on enhancing point cloud feature extraction by selectively\nobtaining more meaningful local features (wang2022rbgnet) where contextual\ninformation is overlooked. To this end, the proposed architecture uses\nray-based surface-biased feature extraction and multi-level context encoding to\noutperform the state-of-the-art 3D object detector. In this work, 3D detection\nexperiments are performed on scenes from the ScanNet dataset whereby the\nself-attention modules are introduced one after the other to isolate the effect\nof self-attention at each level.\n", "date": "2023-02-13", "categories": []} {"id": "2211.10581", "title": "Sparse4D: Multi-view 3D Object Detection with Sparse Spatial-Temporal\n Fusion", "abstract": " Bird-eye-view (BEV) based methods have made great progress recently in\nmulti-view 3D detection task. Comparing with BEV based methods, sparse based\nmethods lag behind in performance, but still have lots of non-negligible\nmerits. To push sparse 3D detection further, in this work, we introduce a novel\nmethod, named Sparse4D, which does the iterative refinement of anchor boxes via\nsparsely sampling and fusing spatial-temporal features. (1) Sparse 4D Sampling:\nfor each 3D anchor, we assign multiple 4D keypoints, which are then projected\nto multi-view/scale/timestamp image features to sample corresponding features;\n(2) Hierarchy Feature Fusion: we hierarchically fuse sampled features of\ndifferent view/scale, different timestamp and different keypoints to generate\nhigh-quality instance feature. In this way, Sparse4D can efficiently and\neffectively achieve 3D detection without relying on dense view transformation\nnor global attention, and is more friendly to edge devices deployment.\nFurthermore, we introduce an instance-level depth reweight module to alleviate\nthe ill-posed issue in 3D-to-2D projection. In experiment, our method\noutperforms all sparse based methods and most BEV based methods on detection\ntask in the nuScenes dataset.\n", "date": "2022-11-19", "categories": []} {"id": "2301.09249", "title": "Exploring Active 3D Object Detection from a Generalization Perspective", "abstract": " To alleviate the high annotation cost in LiDAR-based 3D object detection,\nactive learning is a promising solution that learns to select only a small\nportion of unlabeled data to annotate, without compromising model performance.\nOur empirical study, however, suggests that mainstream uncertainty-based and\ndiversity-based active learning policies are not effective when applied in the\n3D detection task, as they fail to balance the trade-off between point cloud\ninformativeness and box-level annotation costs. To overcome this limitation, we\njointly investigate three novel criteria in our framework Crb for point cloud\nacquisition - label conciseness}, feature representativeness and geometric\nbalance, which hierarchically filters out the point clouds of redundant 3D\nbounding box labels, latent features and geometric characteristics (e.g., point\ncloud density) from the unlabeled sample pool and greedily selects informative\nones with fewer objects to annotate. Our theoretical analysis demonstrates that\nthe proposed criteria align the marginal distributions of the selected subset\nand the prior distributions of the unseen test set, and minimizes the upper\nbound of the generalization error. To validate the effectiveness and\napplicability of Crb, we conduct extensive experiments on the two benchmark 3D\nobject detection datasets of KITTI and Waymo and examine both one-stage (i.e.,\nSecond) and two-stage 3D detectors (i.e., Pv-rcnn). Experiments evidence that\nthe proposed approach outperforms existing active learning strategies and\nachieves fully supervised performance requiring $1\\%$ and $8\\%$ annotations of\nbounding boxes and point clouds, respectively. Source code:\nhttps://github.com/Luoyadan/CRB-active-3Ddet.\n", "date": "2023-01-23", "categories": []} {"id": "2302.03914", "title": "Generalized Few-Shot 3D Object Detection of LiDAR Point Cloud for\n Autonomous Driving", "abstract": " Recent years have witnessed huge successes in 3D object detection to\nrecognize common objects for autonomous driving (e.g., vehicles and\npedestrians). However, most methods rely heavily on a large amount of\nwell-labeled training data. This limits their capability of detecting rare\nfine-grained objects (e.g., police cars and ambulances), which is important for\nspecial cases, such as emergency rescue, and so on. To achieve simultaneous\ndetection for both common and rare objects, we propose a novel task, called\ngeneralized few-shot 3D object detection, where we have a large amount of\ntraining data for common (base) objects, but only a few data for rare (novel)\nclasses. Specifically, we analyze in-depth differences between images and point\nclouds, and then present a practical principle for the few-shot setting in the\n3D LiDAR dataset. To solve this task, we propose a simple and effective\ndetection framework, including (1) an incremental fine-tuning method to extend\nexisting 3D detection models to recognize both common and rare objects, and (2)\na sample adaptive balance loss to alleviate the issue of long-tailed data\ndistribution in autonomous driving scenarios. On the nuScenes dataset, we\nconduct sufficient experiments to demonstrate that our approach can\nsuccessfully detect the rare (novel) classes that contain only a few training\ndata, while also maintaining the detection accuracy of common objects.\n", "date": "2023-02-08", "categories": []} {"id": "2209.13507", "title": "CrossDTR: Cross-view and Depth-guided Transformers for 3D Object\n Detection", "abstract": " To achieve accurate 3D object detection at a low cost for autonomous driving,\nmany multi-camera methods have been proposed and solved the occlusion problem\nof monocular approaches. However, due to the lack of accurate estimated depth,\nexisting multi-camera methods often generate multiple bounding boxes along a\nray of depth direction for difficult small objects such as pedestrians,\nresulting in an extremely low recall. Furthermore, directly applying depth\nprediction modules to existing multi-camera methods, generally composed of\nlarge network architectures, cannot meet the real-time requirements of\nself-driving applications. To address these issues, we propose Cross-view and\nDepth-guided Transformers for 3D Object Detection, CrossDTR. First, our\nlightweight depth predictor is designed to produce precise object-wise sparse\ndepth maps and low-dimensional depth embeddings without extra depth datasets\nduring supervision. Second, a cross-view depth-guided transformer is developed\nto fuse the depth embeddings as well as image features from cameras of\ndifferent views and generate 3D bounding boxes. Extensive experiments\ndemonstrated that our method hugely surpassed existing multi-camera methods by\n10 percent in pedestrian detection and about 3 percent in overall mAP and NDS\nmetrics. Also, computational analyses showed that our method is 5 times faster\nthan prior approaches. Our codes will be made publicly available at\nhttps://github.com/sty61010/CrossDTR.\n", "date": "2022-09-27", "categories": []} {"id": "2301.13592", "title": "Priors are Powerful: Improving a Transformer for Multi-camera 3D\n Detection with 2D Priors", "abstract": " Transfomer-based approaches advance the recent development of multi-camera 3D\ndetection both in academia and industry. In a vanilla transformer architecture,\nqueries are randomly initialised and optimised for the whole dataset, without\nconsidering the differences among input frames. In this work, we propose to\nleverage the predictions from an image backbone, which is often highly\noptimised for 2D tasks, as priors to the transformer part of a 3D detection\nnetwork. The method works by (1). augmenting image feature maps with 2D priors,\n(2). sampling query locations via ray-casting along 2D box centroids, as well\nas (3). initialising query features with object-level image features.\nExperimental results shows that 2D priors not only help the model converge\nfaster, but also largely improve the baseline approach by up to 12% in terms of\naverage precision.\n", "date": "2023-01-31", "categories": []} {"id": "2209.09464", "title": "Rethinking Dimensionality Reduction in Grid-based 3D Object Detection", "abstract": " Bird's eye view (BEV) is widely adopted by most of the current point cloud\ndetectors due to the applicability of well-explored 2D detection techniques.\nHowever, existing methods obtain BEV features by simply collapsing voxel or\npoint features along the height dimension, which causes the heavy loss of 3D\nspatial information. To alleviate the information loss, we propose a novel\npoint cloud detection network based on a Multi-level feature dimensionality\nreduction strategy, called MDRNet. In MDRNet, the Spatial-aware Dimensionality\nReduction (SDR) is designed to dynamically focus on the valuable parts of the\nobject during voxel-to-BEV feature transformation. Furthermore, the Multi-level\nSpatial Residuals (MSR) is proposed to fuse the multi-level spatial information\nin the BEV feature maps. Extensive experiments on nuScenes show that the\nproposed method outperforms the state-of-the-art methods. The code will be\navailable upon publication.\n", "date": "2022-09-20", "categories": []} {"id": "2301.07301", "title": "PTA-Det: Point Transformer Associating Point cloud and Image for 3D\n Object Detection", "abstract": " In autonomous driving, 3D object detection based on multi-modal data has\nbecome an indispensable approach when facing complex environments around the\nvehicle. During multi-modal detection, LiDAR and camera are simultaneously\napplied for capturing and modeling. However, due to the intrinsic discrepancies\nbetween the LiDAR point and camera image, the fusion of the data for object\ndetection encounters a series of problems. Most multi-modal detection methods\nperform even worse than LiDAR-only methods. In this investigation, we propose a\nmethod named PTA-Det to improve the performance of multi-modal detection.\nAccompanied by PTA-Det, a Pseudo Point Cloud Generation Network is proposed,\nwhich can convert image information including texture and semantic features by\npseudo points. Thereafter, through a transformer-based Point Fusion Transition\n(PFT) module, the features of LiDAR points and pseudo points from image can be\ndeeply fused under a unified point-based representation. The combination of\nthese modules can conquer the major obstacle in feature fusion across\nmodalities and realizes a complementary and discriminative representation for\nproposal generation. Extensive experiments on the KITTI dataset show the\nPTA-Det achieves a competitive result and support its effectiveness.\n", "date": "2023-01-18", "categories": []} {"id": "2206.15398", "title": "PolarFormer: Multi-camera 3D Object Detection with Polar Transformer", "abstract": " 3D object detection in autonomous driving aims to reason \"what\" and \"where\"\nthe objects of interest present in a 3D world. Following the conventional\nwisdom of previous 2D object detection, existing methods often adopt the\ncanonical Cartesian coordinate system with perpendicular axis. However, we\nconjugate that this does not fit the nature of the ego car's perspective, as\neach onboard camera perceives the world in shape of wedge intrinsic to the\nimaging geometry with radical (non-perpendicular) axis. Hence, in this paper we\nadvocate the exploitation of the Polar coordinate system and propose a new\nPolar Transformer (PolarFormer) for more accurate 3D object detection in the\nbird's-eye-view (BEV) taking as input only multi-camera 2D images.\nSpecifically, we design a cross attention based Polar detection head without\nrestriction to the shape of input structure to deal with irregular Polar grids.\nFor tackling the unconstrained object scale variations along Polar's distance\ndimension, we further introduce a multi-scalePolar representation learning\nstrategy. As a result, our model can make best use of the Polar representation\nrasterized via attending to the corresponding image observation in a\nsequence-to-sequence fashion subject to the geometric constraints. Thorough\nexperiments on the nuScenes dataset demonstrate that our PolarFormer\noutperforms significantly state-of-the-art 3D object detection alternatives.\n", "date": "2022-06-30", "categories": []} {"id": "2301.06944", "title": "DR-WLC: Dimensionality Reduction cognition for object detection and pose\n estimation by Watching, Learning and Checking", "abstract": " Object detection and pose estimation are difficult tasks in robotics and\nautonomous driving. Existing object detection and pose estimation methods\nmostly adopt the same-dimensional data for training. For example, 2D object\ndetection usually requires a large amount of 2D annotation data with high cost.\nUsing high-dimensional information to supervise lower-dimensional tasks is a\nfeasible way to reduce datasets size. In this work, the DR-WLC, a\ndimensionality reduction cognitive model, which can perform both object\ndetection and pose estimation tasks at the same time is proposed. The model\nonly requires 3D model of objects and unlabeled environment images (with or\nwithout objects) to finish the training. In addition, a bounding boxes\ngeneration strategy is also proposed to build the relationship between 3D model\nand 2D object detection task. Experiments show that our method can qualify the\nwork without any manual annotations and it is easy to deploy for practical\napplications. Source code is at https://github.com/IN2-ViAUn/DR-WLC.\n", "date": "2023-01-17", "categories": []} {"id": "2301.04613", "title": "Object Detection in 3D Point Clouds via Local Correlation-Aware Point\n Embedding", "abstract": " We present an improved approach for 3D object detection in point cloud data\nbased on the Frustum PointNet (F-PointNet). Compared to the original\nF-PointNet, our newly proposed method considers the point neighborhood when\ncomputing point features. The newly introduced local neighborhood embedding\noperation mimics the convolutional operations in 2D neural networks. Thus\nfeatures of each point are not only computed with the features of its own or of\nthe whole point cloud but also computed especially with respect to the features\nof its neighbors. Experiments show that our proposed method achieves better\nperformance than the F-Pointnet baseline on 3D object detection tasks.\n", "date": "2023-01-11", "categories": []} {"id": "2301.04058", "title": "Rethinking Voxelization and Classification for 3D Object Detection", "abstract": " The main challenge in 3D object detection from LiDAR point clouds is\nachieving real-time performance without affecting the reliability of the\nnetwork. In other words, the detecting network must be confident enough about\nits predictions. In this paper, we present a solution to improve network\ninference speed and precision at the same time by implementing a fast dynamic\nvoxelizer that works on fast pillar-based models in the same way a voxelizer\nworks on slow voxel-based models. In addition, we propose a lightweight\ndetection sub-head model for classifying predicted objects and filter out false\ndetected objects that significantly improves model precision in a negligible\ntime and computing cost. The developed code is publicly available at:\nhttps://github.com/YoushaaMurhij/RVCDet.\n", "date": "2023-01-10", "categories": []} {"id": "2301.01615", "title": "StereoDistill: Pick the Cream from LiDAR for Distilling Stereo-based 3D\n Object Detection", "abstract": " In this paper, we propose a cross-modal distillation method named\nStereoDistill to narrow the gap between the stereo and LiDAR-based approaches\nvia distilling the stereo detectors from the superior LiDAR model at the\nresponse level, which is usually overlooked in 3D object detection\ndistillation. The key designs of StereoDistill are: the X-component Guided\nDistillation~(XGD) for regression and the Cross-anchor Logit Distillation~(CLD)\nfor classification. In XGD, instead of empirically adopting a threshold to\nselect the high-quality teacher predictions as soft targets, we decompose the\npredicted 3D box into sub-components and retain the corresponding part for\ndistillation if the teacher component pilot is consistent with ground truth to\nlargely boost the number of positive predictions and alleviate the mimicking\ndifficulty of the student model. For CLD, we aggregate the probability\ndistribution of all anchors at the same position to encourage the highest\nprobability anchor rather than individually distill the distribution at the\nanchor level. Finally, our StereoDistill achieves state-of-the-art results for\nstereo-based 3D detection on the KITTI test benchmark and extensive experiments\non KITTI and Argoverse Dataset validate the effectiveness.\n", "date": "2023-01-04", "categories": []} {"id": "2301.02562", "title": "Super Sparse 3D Object Detection", "abstract": " As the perception range of LiDAR expands, LiDAR-based 3D object detection\ncontributes ever-increasingly to the long-range perception in autonomous\ndriving. Mainstream 3D object detectors often build dense feature maps, where\nthe cost is quadratic to the perception range, making them hardly scale up to\nthe long-range settings. To enable efficient long-range detection, we first\npropose a fully sparse object detector termed FSD. FSD is built upon the\ngeneral sparse voxel encoder and a novel sparse instance recognition (SIR)\nmodule. SIR groups the points into instances and applies highly-efficient\ninstance-wise feature extraction. The instance-wise grouping sidesteps the\nissue of the center feature missing, which hinders the design of the fully\nsparse architecture. To further enjoy the benefit of fully sparse\ncharacteristic, we leverage temporal information to remove data redundancy and\npropose a super sparse detector named FSD++. FSD++ first generates residual\npoints, which indicate the point changes between consecutive frames. The\nresidual points, along with a few previous foreground points, form the super\nsparse input data, greatly reducing data redundancy and computational overhead.\nWe comprehensively analyze our method on the large-scale Waymo Open Dataset,\nand state-of-the-art performance is reported. To showcase the superiority of\nour method in long-range detection, we also conduct experiments on Argoverse 2\nDataset, where the perception range ($200m$) is much larger than Waymo Open\nDataset ($75m$). Code is open-sourced at https://github.com/tusen-ai/SST.\n", "date": "2023-01-05", "categories": []} {"id": "2301.01802", "title": "MonoEdge: Monocular 3D Object Detection Using Local Perspectives", "abstract": " We propose a novel approach for monocular 3D object detection by leveraging\nlocal perspective effects of each object. While the global perspective effect\nshown as size and position variations has been exploited for monocular 3D\ndetection extensively, the local perspectives has long been overlooked. We\ndesign a local perspective module to regress a newly defined variable named\nkeyedge-ratios as the parameterization of the local shape distortion to account\nfor the local perspective, and derive the object depth and yaw angle from it.\nTheoretically, this module does not rely on the pixel-wise size or position in\nthe image of the objects, therefore independent of the camera intrinsic\nparameters. By plugging this module in existing monocular 3D object detection\nframeworks, we incorporate the local perspective distortion with global\nperspective effect for monocular 3D reasoning, and we demonstrate the\neffectiveness and superior performance over strong baseline methods in multiple\ndatasets.\n", "date": "2023-01-04", "categories": []} {"id": "2212.13979", "title": "TiG-BEV: Multi-view BEV 3D Object Detection via Target Inner-Geometry\n Learning", "abstract": " To achieve accurate and low-cost 3D object detection, existing methods\npropose to benefit camera-based multi-view detectors with spatial cues provided\nby the LiDAR modality, e.g., dense depth supervision and bird-eye-view (BEV)\nfeature distillation. However, they directly conduct point-to-point mimicking\nfrom LiDAR to camera, which neglects the inner-geometry of foreground targets\nand suffers from the modal gap between 2D-3D features. In this paper, we\npropose the learning scheme of Target Inner-Geometry from the LiDAR modality\ninto camera-based BEV detectors for both dense depth and BEV features, termed\nas TiG-BEV. First, we introduce an inner-depth supervision module to learn the\nlow-level relative depth relations between different foreground pixels. This\nenables the camera-based detector to better understand the object-wise spatial\nstructures. Second, we design an inner-feature BEV distillation module to\nimitate the high-level semantics of different keypoints within foreground\ntargets. To further alleviate the BEV feature gap between two modalities, we\nadopt both inter-channel and inter-keypoint distillation for feature-similarity\nmodeling. With our target inner-geometry distillation, TiG-BEV can effectively\nboost BEVDepth by +2.3% NDS and +2.4% mAP, along with BEVDet by +9.1% NDS and\n+10.3% mAP on nuScenes val set. Code will be available at\nhttps://github.com/ADLab3Ds/TiG-BEV.\n", "date": "2022-12-28", "categories": []} {"id": "2212.11804", "title": "Monocular 3D Object Detection using Multi-Stage Approaches with\n Attention and Slicing aided hyper inference", "abstract": " 3D object detection is vital as it would enable us to capture objects' sizes,\norientation, and position in the world. As a result, we would be able to use\nthis 3D detection in real-world applications such as Augmented Reality (AR),\nself-driving cars, and robotics which perceive the world the same way we do as\nhumans. Monocular 3D Object Detection is the task to draw 3D bounding box\naround objects in a single 2D RGB image. It is localization task but without\nany extra information like depth or other sensors or multiple images. Monocular\n3D object detection is an important yet challenging task. Beyond the\nsignificant progress in image-based 2D object detection, 3D understanding of\nreal-world objects is an open challenge that has not been explored extensively\nthus far. In addition to the most closely related studies.\n", "date": "2022-12-22", "categories": []} {"id": "2210.05593", "title": "Prototypical VoteNet for Few-Shot 3D Point Cloud Object Detection", "abstract": " Most existing 3D point cloud object detection approaches heavily rely on\nlarge amounts of labeled training data. However, the labeling process is costly\nand time-consuming. This paper considers few-shot 3D point cloud object\ndetection, where only a few annotated samples of novel classes are needed with\nabundant samples of base classes. To this end, we propose Prototypical VoteNet\nto recognize and localize novel instances, which incorporates two new modules:\nPrototypical Vote Module (PVM) and Prototypical Head Module (PHM).\nSpecifically, as the 3D basic geometric structures can be shared among\ncategories, PVM is designed to leverage class-agnostic geometric prototypes,\nwhich are learned from base classes, to refine local features of novel\ncategories.Then PHM is proposed to utilize class prototypes to enhance the\nglobal feature of each object, facilitating subsequent object localization and\nclassification, which is trained by the episodic training strategy. To evaluate\nthe model in this new setting, we contribute two new benchmark datasets,\nFS-ScanNet and FS-SUNRGBD. We conduct extensive experiments to demonstrate the\neffectiveness of Prototypical VoteNet, and our proposed method shows\nsignificant and consistent improvements compared to baselines on two benchmark\ndatasets.\n", "date": "2022-10-11", "categories": []} {"id": "2212.00442", "title": "MGTANet: Encoding Sequential LiDAR Points Using Long Short-Term\n Motion-Guided Temporal Attention for 3D Object Detection", "abstract": " Most scanning LiDAR sensors generate a sequence of point clouds in real-time.\nWhile conventional 3D object detectors use a set of unordered LiDAR points\nacquired over a fixed time interval, recent studies have revealed that\nsubstantial performance improvement can be achieved by exploiting the\nspatio-temporal context present in a sequence of LiDAR point sets. In this\npaper, we propose a novel 3D object detection architecture, which can encode\nLiDAR point cloud sequences acquired by multiple successive scans. The encoding\nprocess of the point cloud sequence is performed on two different time scales.\nWe first design a short-term motion-aware voxel encoding that captures the\nshort-term temporal changes of point clouds driven by the motion of objects in\neach voxel. We also propose long-term motion-guided bird's eye view (BEV)\nfeature enhancement that adaptively aligns and aggregates the BEV feature maps\nobtained by the short-term voxel encoding by utilizing the dynamic motion\ncontext inferred from the sequence of the feature maps. The experiments\nconducted on the public nuScenes benchmark demonstrate that the proposed 3D\nobject detector offers significant improvements in performance compared to the\nbaseline methods and that it sets a state-of-the-art performance for certain 3D\nobject detection categories. Code is available at\nhttps://github.com/HYjhkoh/MGTANet.git\n", "date": "2022-12-01", "categories": []} {"id": "2212.07560", "title": "Multi-level and multi-modal feature fusion for accurate 3D object\n detection in Connected and Automated Vehicles", "abstract": " Aiming at highly accurate object detection for connected and automated\nvehicles (CAVs), this paper presents a Deep Neural Network based 3D object\ndetection model that leverages a three-stage feature extractor by developing a\nnovel LIDAR-Camera fusion scheme. The proposed feature extractor extracts\nhigh-level features from two input sensory modalities and recovers the\nimportant features discarded during the convolutional process. The novel fusion\nscheme effectively fuses features across sensory modalities and convolutional\nlayers to find the best representative global features. The fused features are\nshared by a two-stage network: the region proposal network (RPN) and the\ndetection head (DH). The RPN generates high-recall proposals, and the DH\nproduces final detection results. The experimental results show the proposed\nmodel outperforms more recent research on the KITTI 2D and 3D detection\nbenchmark, particularly for distant and highly occluded instances.\n", "date": "2022-12-15", "categories": []} {"id": "2212.09273", "title": "Learning Object-level Point Augmentor for Semi-supervised 3D Object\n Detection", "abstract": " Semi-supervised object detection is important for 3D scene understanding\nbecause obtaining large-scale 3D bounding box annotations on point clouds is\ntime-consuming and labor-intensive. Existing semi-supervised methods usually\nemploy teacher-student knowledge distillation together with an augmentation\nstrategy to leverage unlabeled point clouds. However, these methods adopt\nglobal augmentation with scene-level transformations and hence are sub-optimal\nfor instance-level object detection. In this work, we propose an object-level\npoint augmentor (OPA) that performs local transformations for semi-supervised\n3D object detection. In this way, the resultant augmentor is derived to\nemphasize object instances rather than irrelevant backgrounds, making the\naugmented data more useful for object detector training. Extensive experiments\non the ScanNet and SUN RGB-D datasets show that the proposed OPA performs\nfavorably against the state-of-the-art methods under various experimental\nsettings. The source code will be available at https://github.com/nomiaro/OPA.\n", "date": "2022-12-19", "categories": []} {"id": "2212.07849", "title": "DETR4D: Direct Multi-View 3D Object Detection with Sparse Attention", "abstract": " 3D object detection with surround-view images is an essential task for\nautonomous driving. In this work, we propose DETR4D, a Transformer-based\nframework that explores sparse attention and direct feature query for 3D object\ndetection in multi-view images. We design a novel projective cross-attention\nmechanism for query-image interaction to address the limitations of existing\nmethods in terms of geometric cue exploitation and information loss for\ncross-view objects. In addition, we introduce a heatmap generation technique\nthat bridges 3D and 2D spaces efficiently via query initialization.\nFurthermore, unlike the common practice of fusing intermediate spatial features\nfor temporal aggregation, we provide a new perspective by introducing a novel\nhybrid approach that performs cross-frame fusion over past object queries and\nimage features, enabling efficient and robust modeling of temporal information.\nExtensive experiments on the nuScenes dataset demonstrate the effectiveness and\nefficiency of the proposed DETR4D.\n", "date": "2022-12-15", "categories": []} {"id": "2212.07289", "title": "ConQueR: Query Contrast Voxel-DETR for 3D Object Detection", "abstract": " Although DETR-based 3D detectors can simplify the detection pipeline and\nachieve direct sparse predictions, their performance still lags behind dense\ndetectors with post-processing for 3D object detection from point clouds. DETRs\nusually adopt a larger number of queries than GTs (e.g., 300 queries v.s. 40\nobjects in Waymo) in a scene, which inevitably incur many false positives\nduring inference. In this paper, we propose a simple yet effective sparse 3D\ndetector, named Query Contrast Voxel-DETR (ConQueR), to eliminate the\nchallenging false positives, and achieve more accurate and sparser predictions.\nWe observe that most false positives are highly overlapping in local regions,\ncaused by the lack of explicit supervision to discriminate locally similar\nqueries. We thus propose a Query Contrast mechanism to explicitly enhance\nqueries towards their best-matched GTs over all unmatched query predictions.\nThis is achieved by the construction of positive and negative GT-query pairs\nfor each GT, and a contrastive loss to enhance positive GT-query pairs against\nnegative ones based on feature similarities. ConQueR closes the gap of sparse\nand dense 3D detectors, and reduces up to ~60% false positives. Our\nsingle-frame ConQueR achieves new state-of-the-art (sota) 71.6 mAPH/L2 on the\nchallenging Waymo Open Dataset validation set, outperforming previous sota\nmethods (e.g., PV-RCNN++) by over 2.0 mAPH/L2.\n", "date": "2022-12-14", "categories": []} {"id": "2212.05505", "title": "Focal-PETR: Embracing Foreground for Efficient Multi-Camera 3D Object\n Detection", "abstract": " The dominant multi-camera 3D detection paradigm is based on explicit 3D\nfeature construction, which requires complicated indexing of local image-view\nfeatures via 3D-to-2D projection. Other methods implicitly introduce geometric\npositional encoding and perform global attention (e.g., PETR) to build the\nrelationship between image tokens and 3D objects. The 3D-to-2D perspective\ninconsistency and global attention lead to a weak correlation between\nforeground tokens and queries, resulting in slow convergence. We propose\nFocal-PETR with instance-guided supervision and spatial alignment module to\nadaptively focus object queries on discriminative foreground regions.\nFocal-PETR additionally introduces a down-sampling strategy to reduce the\nconsumption of global attention. Due to the highly parallelized implementation\nand down-sampling strategy, our model, without depth supervision, achieves\nleading performance on the large-scale nuScenes benchmark and a superior speed\nof 30 FPS on a single RTX3090 GPU. Extensive experiments show that our method\noutperforms PETR while consuming 3x fewer training hours. The code will be made\npublicly available.\n", "date": "2022-12-11", "categories": []} {"id": "2207.09869", "title": "A Novel Neural Network Training Method for Autonomous Driving Using\n Semi-Pseudo-Labels and 3D Data Augmentations", "abstract": " Training neural networks to perform 3D object detection for autonomous\ndriving requires a large amount of diverse annotated data. However, obtaining\ntraining data with sufficient quality and quantity is expensive and sometimes\nimpossible due to human and sensor constraints. Therefore, a novel solution is\nneeded for extending current training methods to overcome this limitation and\nenable accurate 3D object detection. Our solution for the above-mentioned\nproblem combines semi-pseudo-labeling and novel 3D augmentations. For\ndemonstrating the applicability of the proposed method, we have designed a\nconvolutional neural network for 3D object detection which can significantly\nincrease the detection range in comparison with the training data distribution.\n", "date": "2022-07-20", "categories": []} {"id": "2208.11112", "title": "DeepInteraction: 3D Object Detection via Modality Interaction", "abstract": " Existing top-performance 3D object detectors typically rely on the\nmulti-modal fusion strategy. This design is however fundamentally restricted\ndue to overlooking the modality-specific useful information and finally\nhampering the model performance. To address this limitation, in this work we\nintroduce a novel modality interaction strategy where individual per-modality\nrepresentations are learned and maintained throughout for enabling their unique\ncharacteristics to be exploited during object detection. To realize this\nproposed strategy, we design a DeepInteraction architecture characterized by a\nmulti-modal representational interaction encoder and a multi-modal predictive\ninteraction decoder. Experiments on the large-scale nuScenes dataset show that\nour proposed method surpasses all prior arts often by a large margin.\nCrucially, our method is ranked at the first position at the highly competitive\nnuScenes object detection leaderboard.\n", "date": "2022-08-23", "categories": []} {"id": "2212.02755", "title": "Objects as Spatio-Temporal 2.5D points", "abstract": " Determining accurate bird's eye view (BEV) positions of objects and tracks in\na scene is vital for various perception tasks including object interactions\nmapping, scenario extraction etc., however, the level of supervision required\nto accomplish that is extremely challenging to procure. We propose a\nlight-weight, weakly supervised method to estimate 3D position of objects by\njointly learning to regress the 2D object detections and scene's depth\nprediction in a single feed-forward pass of a network. Our proposed method\nextends a center-point based single-shot object detector, and introduces a\nnovel object representation where each object is modeled as a BEV point\nspatio-temporally, without the need of any 3D or BEV annotations for training\nand LiDAR data at query time. The approach leverages readily available 2D\nobject supervision along with LiDAR point clouds (used only during training) to\njointly train a single network, that learns to predict 2D object detection\nalongside the whole scene's depth, to spatio-temporally model object tracks as\npoints in BEV. The proposed method is computationally over $\\sim$10x efficient\ncompared to recent SOTA approaches while achieving comparable accuracies on\nKITTI tracking benchmark.\n", "date": "2022-12-06", "categories": []} {"id": "2212.02845", "title": "SSDA3D: Semi-supervised Domain Adaptation for 3D Object Detection from\n Point Cloud", "abstract": " LiDAR-based 3D object detection is an indispensable task in advanced\nautonomous driving systems. Though impressive detection results have been\nachieved by superior 3D detectors, they suffer from significant performance\ndegeneration when facing unseen domains, such as different LiDAR\nconfigurations, different cities, and weather conditions. The mainstream\napproaches tend to solve these challenges by leveraging unsupervised domain\nadaptation (UDA) techniques. However, these UDA solutions just yield\nunsatisfactory 3D detection results when there is a severe domain shift, e.g.,\nfrom Waymo (64-beam) to nuScenes (32-beam). To address this, we present a novel\nSemi-Supervised Domain Adaptation method for 3D object detection (SSDA3D),\nwhere only a few labeled target data is available, yet can significantly\nimprove the adaptation performance. In particular, our SSDA3D includes an\nInter-domain Adaptation stage and an Intra-domain Generalization stage. In the\nfirst stage, an Inter-domain Point-CutMix module is presented to efficiently\nalign the point cloud distribution across domains. The Point-CutMix generates\nmixed samples of an intermediate domain, thus encouraging to learn\ndomain-invariant knowledge. Then, in the second stage, we further enhance the\nmodel for better generalization on the unlabeled target set. This is achieved\nby exploring Intra-domain Point-MixUp in semi-supervised learning, which\nessentially regularizes the pseudo label distribution. Experiments from Waymo\nto nuScenes show that, with only 10% labeled target data, our SSDA3D can\nsurpass the fully-supervised oracle model with 100% target label. Our code is\navailable at https://github.com/yinjunbo/SSDA3D.\n", "date": "2022-12-06", "categories": []} {"id": "2208.09446", "title": "MonoSIM: Simulating Learning Behaviors of Heterogeneous Point Cloud\n Object Detectors for Monocular 3D Object Detection", "abstract": " Monocular 3D object detection is a fundamental but very important task to\nmany applications including autonomous driving, robotic grasping and augmented\nreality. Existing leading methods tend to estimate the depth of the input image\nfirst, and detect the 3D object based on point cloud. This routine suffers from\nthe inherent gap between depth estimation and object detection. Besides, the\nprediction error accumulation would also affect the performance. In this paper,\na novel method named MonoSIM is proposed. The insight behind introducing\nMonoSIM is that we propose to simulate the feature learning behaviors of a\npoint cloud based detector for monocular detector during the training period.\nHence, during inference period, the learned features and prediction would be\nsimilar to the point cloud based detector as possible. To achieve it, we\npropose one scene-level simulation module, one RoI-level simulation module and\none response-level simulation module, which are progressively used for the\ndetector's full feature learning and prediction pipeline. We apply our method\nto the famous M3D-RPN detector and CaDDN detector, conducting extensive\nexperiments on KITTI and Waymo Open datasets. Results show that our method\nconsistently improves the performance of different monocular detectors for a\nlarge margin without changing their network architectures. Our codes will be\npublicly available at\nhttps://github.com/sunh18/MonoSIM}{https://github.com/sunh18/MonoSIM.\n", "date": "2022-08-19", "categories": []} {"id": "2212.02057", "title": "DA-CIL: Towards Domain Adaptive Class-Incremental 3D Object Detection", "abstract": " Deep learning has achieved notable success in 3D object detection with the\nadvent of large-scale point cloud datasets. However, severe performance\ndegradation in the past trained classes, i.e., catastrophic forgetting, still\nremains a critical issue for real-world deployment when the number of classes\nis unknown or may vary. Moreover, existing 3D class-incremental detection\nmethods are developed for the single-domain scenario, which fail when\nencountering domain shift caused by different datasets, varying environments,\netc. In this paper, we identify the unexplored yet valuable scenario, i.e.,\nclass-incremental learning under domain shift, and propose a novel 3D domain\nadaptive class-incremental object detection framework, DA-CIL, in which we\ndesign a novel dual-domain copy-paste augmentation method to construct multiple\naugmented domains for diversifying training distributions, thereby facilitating\ngradual domain adaptation. Then, multi-level consistency is explored to\nfacilitate dual-teacher knowledge distillation from different domains for\ndomain adaptive class-incremental learning. Extensive experiments on various\ndatasets demonstrate the effectiveness of the proposed method over baselines in\nthe domain adaptive class-incremental learning scenario.\n", "date": "2022-12-05", "categories": []} {"id": "2212.01231", "title": "BEV-SAN: Accurate BEV 3D Object Detection via Slice Attention Networks", "abstract": " Bird's-Eye-View (BEV) 3D Object Detection is a crucial multi-view technique\nfor autonomous driving systems. Recently, plenty of works are proposed,\nfollowing a similar paradigm consisting of three essential components, i.e.,\ncamera feature extraction, BEV feature construction, and task heads. Among the\nthree components, BEV feature construction is BEV-specific compared with 2D\ntasks. Existing methods aggregate the multi-view camera features to the\nflattened grid in order to construct the BEV feature. However, flattening the\nBEV space along the height dimension fails to emphasize the informative\nfeatures of different heights. For example, the barrier is located at a low\nheight while the truck is located at a high height. In this paper, we propose a\nnovel method named BEV Slice Attention Network (BEV-SAN) for exploiting the\nintrinsic characteristics of different heights. Instead of flattening the BEV\nspace, we first sample along the height dimension to build the global and local\nBEV slices. Then, the features of BEV slices are aggregated from the camera\nfeatures and merged by the attention mechanism. Finally, we fuse the merged\nlocal and global BEV features by a transformer to generate the final feature\nmap for task heads. The purpose of local BEV slices is to emphasize informative\nheights. In order to find them, we further propose a LiDAR-guided sampling\nstrategy to leverage the statistical distribution of LiDAR to determine the\nheights of local slices. Compared with uniform sampling, LiDAR-guided sampling\ncan determine more informative heights. We conduct detailed experiments to\ndemonstrate the effectiveness of BEV-SAN. Code will be released.\n", "date": "2022-12-02", "categories": []} {"id": "2212.01247", "title": "CC-3DT: Panoramic 3D Object Tracking via Cross-Camera Fusion", "abstract": " To track the 3D locations and trajectories of the other traffic participants\nat any given time, modern autonomous vehicles are equipped with multiple\ncameras that cover the vehicle's full surroundings. Yet, camera-based 3D object\ntracking methods prioritize optimizing the single-camera setup and resort to\npost-hoc fusion in a multi-camera setup. In this paper, we propose a method for\npanoramic 3D object tracking, called CC-3DT, that associates and models object\ntrajectories both temporally and across views, and improves the overall\ntracking consistency. In particular, our method fuses 3D detections from\nmultiple cameras before association, reducing identity switches significantly\nand improving motion modeling. Our experiments on large-scale driving datasets\nshow that fusion before association leads to a large margin of improvement over\npost-hoc fusion. We set a new state-of-the-art with 12.6% improvement in\naverage multi-object tracking accuracy (AMOTA) among all camera-based methods\non the competitive NuScenes 3D tracking benchmark, outperforming previously\npublished methods by 6.5% in AMOTA with the same 3D detector.\n", "date": "2022-12-02", "categories": []} {"id": "2211.11962", "title": "Transformation-Equivariant 3D Object Detection for Autonomous Driving", "abstract": " 3D object detection received increasing attention in autonomous driving\nrecently. Objects in 3D scenes are distributed with diverse orientations.\nOrdinary detectors do not explicitly model the variations of rotation and\nreflection transformations. Consequently, large networks and extensive data\naugmentation are required for robust detection. Recent equivariant networks\nexplicitly model the transformation variations by applying shared networks on\nmultiple transformed point clouds, showing great potential in object geometry\nmodeling. However, it is difficult to apply such networks to 3D object\ndetection in autonomous driving due to its large computation cost and slow\nreasoning speed. In this work, we present TED, an efficient\nTransformation-Equivariant 3D Detector to overcome the computation cost and\nspeed issues. TED first applies a sparse convolution backbone to extract\nmulti-channel transformation-equivariant voxel features; and then aligns and\naggregates these equivariant features into lightweight and compact\nrepresentations for high-performance 3D object detection. On the highly\ncompetitive KITTI 3D car detection leaderboard, TED ranked 1st among all\nsubmissions with competitive efficiency.\n", "date": "2022-11-22", "categories": []} {"id": "2212.00244", "title": "CL3D: Unsupervised Domain Adaptation for Cross-LiDAR 3D Detection", "abstract": " Domain adaptation for Cross-LiDAR 3D detection is challenging due to the\nlarge gap on the raw data representation with disparate point densities and\npoint arrangements. By exploring domain-invariant 3D geometric characteristics\nand motion patterns, we present an unsupervised domain adaptation method that\novercomes above difficulties. First, we propose the Spatial Geometry Alignment\nmodule to extract similar 3D shape geometric features of the same object class\nto align two domains, while eliminating the effect of distinct point\ndistributions. Second, we present Temporal Motion Alignment module to utilize\nmotion features in sequential frames of data to match two domains. Prototypes\ngenerated from two modules are incorporated into the pseudo-label reweighting\nprocedure and contribute to our effective self-training framework for the\ntarget domain. Extensive experiments show that our method achieves\nstate-of-the-art performance on cross-device datasets, especially for the\ndatasets with large gaps captured by mechanical scanning LiDARs and solid-state\nLiDARs in various scenes. Project homepage is at\nhttps://github.com/4DVLab/CL3D.git\n", "date": "2022-12-01", "categories": []} {"id": "2212.00623", "title": "BEV-LGKD: A Unified LiDAR-Guided Knowledge Distillation Framework for\n BEV 3D Object Detection", "abstract": " Recently, Bird's-Eye-View (BEV) representation has gained increasing\nattention in multi-view 3D object detection, which has demonstrated promising\napplications in autonomous driving. Although multi-view camera systems can be\ndeployed at low cost, the lack of depth information makes current approaches\nadopt large models for good performance. Therefore, it is essential to improve\nthe efficiency of BEV 3D object detection. Knowledge Distillation (KD) is one\nof the most practical techniques to train efficient yet accurate models.\nHowever, BEV KD is still under-explored to the best of our knowledge. Different\nfrom image classification tasks, BEV 3D object detection approaches are more\ncomplicated and consist of several components. In this paper, we propose a\nunified framework named BEV-LGKD to transfer the knowledge in the\nteacher-student manner. However, directly applying the teacher-student paradigm\nto BEV features fails to achieve satisfying results due to heavy background\ninformation in RGB cameras. To solve this problem, we propose to leverage the\nlocalization advantage of LiDAR points. Specifically, we transform the LiDAR\npoints to BEV space and generate the foreground mask and view-dependent mask\nfor the teacher-student paradigm. It is to be noted that our method only uses\nLiDAR points to guide the KD between RGB models. As the quality of depth\nestimation is crucial for BEV perception, we further introduce depth\ndistillation to our framework. Our unified framework is simple yet effective\nand achieves a significant performance boost. Code will be released.\n", "date": "2022-12-01", "categories": []} {"id": "2206.10092", "title": "BEVDepth: Acquisition of Reliable Depth for Multi-view 3D Object\n Detection", "abstract": " In this research, we propose a new 3D object detector with a trustworthy\ndepth estimation, dubbed BEVDepth, for camera-based Bird's-Eye-View (BEV) 3D\nobject detection. Our work is based on a key observation -- depth estimation in\nrecent approaches is surprisingly inadequate given the fact that depth is\nessential to camera 3D detection. Our BEVDepth resolves this by leveraging\nexplicit depth supervision. A camera-awareness depth estimation module is also\nintroduced to facilitate the depth predicting capability. Besides, we design a\nnovel Depth Refinement Module to counter the side effects carried by imprecise\nfeature unprojection. Aided by customized Efficient Voxel Pooling and\nmulti-frame mechanism, BEVDepth achieves the new state-of-the-art 60.9% NDS on\nthe challenging nuScenes test set while maintaining high efficiency. For the\nfirst time, the NDS score of a camera model reaches 60%.\n", "date": "2022-06-21", "categories": []} {"id": "2211.17111", "title": "BEVPoolv2: A Cutting-edge Implementation of BEVDet Toward Deployment", "abstract": " We release a new codebase version of the BEVDet, dubbed branch dev2.0. With\ndev2.0, we propose BEVPoolv2 upgrade the view transformation process from the\nperspective of engineering optimization, making it free from a huge burden in\nboth calculation and storage aspects. It achieves this by omitting the\ncalculation and preprocessing of the large frustum feature. As a result, it can\nbe processed within 0.82 ms even with a large input resolution of 640x1600,\nwhich is 15.1 times the previous fastest implementation. Besides, it is also\nless cache consumptive when compared with the previous implementation,\nnaturally as it no longer needs to store the large frustum feature. Last but\nnot least, this also makes the deployment to the other backend handy. We offer\nan example of deployment to the TensorRT backend in branch dev2.0 and show how\nfast the BEVDet paradigm can be processed on it. Other than BEVPoolv2, we also\nselect and integrate some substantial progress that was proposed in the past\nyear. As an example configuration, BEVDet4D-R50-Depth-CBGS scores 52.3 NDS on\nthe NuScenes validation set and can be processed at a speed of 16.4 FPS with\nthe PyTorch backend. The code has been released to facilitate the study on\nhttps://github.com/HuangJunJie2017/BEVDet/tree/dev2.0.\n", "date": "2022-11-30", "categories": []} {"id": "2207.07933", "title": "Consistency of Implicit and Explicit Features Matters for Monocular 3D\n Object Detection", "abstract": " Low-cost autonomous agents including autonomous driving vehicles chiefly\nadopt monocular 3D object detection to perceive surrounding environment. This\npaper studies 3D intermediate representation methods which generate\nintermediate 3D features for subsequent tasks. For example, the 3D features can\nbe taken as input for not only detection, but also end-to-end prediction and/or\nplanning that require a bird's-eye-view feature representation. In the study,\nwe found that in generating 3D representation previous methods do not maintain\nthe consistency between the objects' implicit poses in the latent space,\nespecially orientations, and the explicitly observed poses in the Euclidean\nspace, which can substantially hurt model performance. To tackle this problem,\nwe present a novel monocular detection method, the first one being aware of the\nposes to purposefully guarantee that they are consistent between the implicit\nand explicit features. Additionally, we introduce a local ray attention\nmechanism to efficiently transform image features to voxels at accurate 3D\nlocations. Thirdly, we propose a handcrafted Gaussian positional encoding\nfunction, which outperforms the sinusoidal encoding function while retaining\nthe benefit of being continuous. Results show that our method improves the\nstate-of-the-art 3D intermediate representation method by 3.15%. We are ranked\n1st among all the reported monocular methods on both 3D and BEV detection\nbenchmark on KITTI leaderboard as of th result's submission time.\n", "date": "2022-07-16", "categories": []} {"id": "2207.12659", "title": "Graph Neural Network and Spatiotemporal Transformer Attention for 3D\n Video Object Detection from Point Clouds", "abstract": " Previous works for LiDAR-based 3D object detection mainly focus on the\nsingle-frame paradigm. In this paper, we propose to detect 3D objects by\nexploiting temporal information in multiple frames, i.e., the point cloud\nvideos. We empirically categorize the temporal information into short-term and\nlong-term patterns. To encode the short-term data, we present a Grid Message\nPassing Network (GMPNet), which considers each grid (i.e., the grouped points)\nas a node and constructs a k-NN graph with the neighbor grids. To update\nfeatures for a grid, GMPNet iteratively collects information from its\nneighbors, thus mining the motion cues in grids from nearby frames. To further\naggregate the long-term frames, we propose an Attentive Spatiotemporal\nTransformer GRU (AST-GRU), which contains a Spatial Transformer Attention (STA)\nmodule and a Temporal Transformer Attention (TTA) module. STA and TTA enhance\nthe vanilla GRU to focus on small objects and better align the moving objects.\nOur overall framework supports both online and offline video object detection\nin point clouds. We implement our algorithm based on prevalent anchor-based and\nanchor-free detectors. The evaluation results on the challenging nuScenes\nbenchmark show the superior performance of our method, achieving the 1st on the\nleaderboard without any bells and whistles, by the time the paper is submitted.\n", "date": "2022-07-26", "categories": []} {"id": "2209.06535", "title": "CRAFT: Camera-Radar 3D Object Detection with Spatio-Contextual Fusion\n Transformer", "abstract": " Camera and radar sensors have significant advantages in cost, reliability,\nand maintenance compared to LiDAR. Existing fusion methods often fuse the\noutputs of single modalities at the result-level, called the late fusion\nstrategy. This can benefit from using off-the-shelf single sensor detection\nalgorithms, but late fusion cannot fully exploit the complementary properties\nof sensors, thus having limited performance despite the huge potential of\ncamera-radar fusion. Here we propose a novel proposal-level early fusion\napproach that effectively exploits both spatial and contextual properties of\ncamera and radar for 3D object detection. Our fusion framework first associates\nimage proposal with radar points in the polar coordinate system to efficiently\nhandle the discrepancy between the coordinate system and spatial properties.\nUsing this as a first stage, following consecutive cross-attention based\nfeature fusion layers adaptively exchange spatio-contextual information between\ncamera and radar, leading to a robust and attentive fusion. Our camera-radar\nfusion approach achieves the state-of-the-art 41.1% mAP and 52.3% NDS on the\nnuScenes test set, which is 8.7 and 10.8 points higher than the camera-only\nbaseline, as well as yielding competitive performance on the LiDAR method.\n", "date": "2022-09-14", "categories": []} {"id": "2211.13858", "title": "Far3Det: Towards Far-Field 3D Detection", "abstract": " We focus on the task of far-field 3D detection (Far3Det) of objects beyond a\ncertain distance from an observer, e.g., $>$50m. Far3Det is particularly\nimportant for autonomous vehicles (AVs) operating at highway speeds, which\nrequire detections of far-field obstacles to ensure sufficient braking\ndistances. However, contemporary AV benchmarks such as nuScenes underemphasize\nthis problem because they evaluate performance only up to a certain distance\n(50m). One reason is that obtaining far-field 3D annotations is difficult,\nparticularly for lidar sensors that produce very few point returns for far-away\nobjects. Indeed, we find that almost 50% of far-field objects (beyond 50m)\ncontain zero lidar points. Secondly, current metrics for 3D detection employ a\n\"one-size-fits-all\" philosophy, using the same tolerance thresholds for near\nand far objects, inconsistent with tolerances for both human vision and stereo\ndisparities. Both factors lead to an incomplete analysis of the Far3Det task.\nFor example, while conventional wisdom tells us that high-resolution RGB\nsensors should be vital for 3D detection of far-away objects, lidar-based\nmethods still rank higher compared to RGB counterparts on the current benchmark\nleaderboards. As a first step towards a Far3Det benchmark, we develop a method\nto find well-annotated scenes from the nuScenes dataset and derive a\nwell-annotated far-field validation set. We also propose a Far3Det evaluation\nprotocol and explore various 3D detection methods for Far3Det. Our result\nconvincingly justifies the long-held conventional wisdom that high-resolution\nRGB improves 3D detection in the far-field. We further propose a simple yet\neffective method that fuses detections from RGB and lidar detectors based on\nnon-maximum suppression, which remarkably outperforms state-of-the-art 3D\ndetectors in the far-field.\n", "date": "2022-11-25", "categories": []} {"id": "2211.13067", "title": "Sparse2Dense: Learning to Densify 3D Features for 3D Object Detection", "abstract": " LiDAR-produced point clouds are the major source for most state-of-the-art 3D\nobject detectors. Yet, small, distant, and incomplete objects with sparse or\nfew points are often hard to detect. We present Sparse2Dense, a new framework\nto efficiently boost 3D detection performance by learning to densify point\nclouds in latent space. Specifically, we first train a dense point 3D detector\n(DDet) with a dense point cloud as input and design a sparse point 3D detector\n(SDet) with a regular point cloud as input. Importantly, we formulate the\nlightweight plug-in S2D module and the point cloud reconstruction module in\nSDet to densify 3D features and train SDet to produce 3D features, following\nthe dense 3D features in DDet. So, in inference, SDet can simulate dense 3D\nfeatures from regular (sparse) point cloud inputs without requiring dense\ninputs. We evaluate our method on the large-scale Waymo Open Dataset and the\nWaymo Domain Adaptation Dataset, showing its high performance and efficiency\nover the state of the arts.\n", "date": "2022-11-23", "categories": []} {"id": "2211.08398", "title": "Structured Knowledge Distillation Towards Efficient and Compact\n Multi-View 3D Detection", "abstract": " Detecting 3D objects from multi-view images is a fundamental problem in 3D\ncomputer vision. Recently, significant breakthrough has been made in multi-view\n3D detection tasks. However, the unprecedented detection performance of these\nvision BEV (bird's-eye-view) detection models is accompanied with enormous\nparameters and computation, which make them unaffordable on edge devices. To\naddress this problem, in this paper, we propose a structured knowledge\ndistillation framework, aiming to improve the efficiency of modern vision-only\nBEV detection models. The proposed framework mainly includes: (a)\nspatial-temporal distillation which distills teacher knowledge of information\nfusion from different timestamps and views, (b) BEV response distillation which\ndistills teacher response to different pillars, and (c) weight-inheriting which\nsolves the problem of inconsistent inputs between students and teacher in\nmodern transformer architectures. Experimental results show that our method\nleads to an average improvement of 2.16 mAP and 2.27 NDS on the nuScenes\nbenchmark, outperforming multiple baselines by a large margin.\n", "date": "2022-11-14", "categories": []} {"id": "2211.10593", "title": "MatrixVT: Efficient Multi-Camera to BEV Transformation for 3D Perception", "abstract": " This paper proposes an efficient multi-camera to Bird's-Eye-View (BEV) view\ntransformation method for 3D perception, dubbed MatrixVT. Existing view\ntransformers either suffer from poor transformation efficiency or rely on\ndevice-specific operators, hindering the broad application of BEV models. In\ncontrast, our method generates BEV features efficiently with only convolutions\nand matrix multiplications (MatMul). Specifically, we propose describing the\nBEV feature as the MatMul of image feature and a sparse Feature Transporting\nMatrix (FTM). A Prime Extraction module is then introduced to compress the\ndimension of image features and reduce FTM's sparsity. Moreover, we propose the\nRing \\& Ray Decomposition to replace the FTM with two matrices and reformulate\nour pipeline to reduce calculation further. Compared to existing methods,\nMatrixVT enjoys a faster speed and less memory footprint while remaining\ndeploy-friendly. Extensive experiments on the nuScenes benchmark demonstrate\nthat our method is highly efficient but obtains results on par with the SOTA\nmethod in object detection and map segmentation tasks\n", "date": "2022-11-19", "categories": []} {"id": "2211.10850", "title": "Context-Aware Data Augmentation for LIDAR 3D Object Detection", "abstract": " For 3D object detection, labeling lidar point cloud is difficult, so data\naugmentation is an important module to make full use of precious annotated\ndata. As a widely used data augmentation method, GT-sample effectively improves\ndetection performance by inserting groundtruths into the lidar frame during\ntraining. However, these samples are often placed in unreasonable areas, which\nmisleads model to learn the wrong context information between targets and\nbackgrounds. To address this problem, in this paper, we propose a context-aware\ndata augmentation method (CA-aug) , which ensures the reasonable placement of\ninserted objects by calculating the \"Validspace\" of the lidar point cloud.\nCA-aug is lightweight and compatible with other augmentation methods. Compared\nwith the GT-sample and the similar method in Lidar-aug(SOTA), it brings higher\naccuracy to the existing detectors. We also present an in-depth study of\naugmentation methods for the range-view-based(RV-based) models and find that\nCA-aug can fully exploit the potential of RV-based networks. The experiment on\nKITTI val split shows that CA-aug can improve the mAP of the test model by 8%.\n", "date": "2022-11-20", "categories": []} {"id": "2211.10439", "title": "BEVFormer v2: Adapting Modern Image Backbones to Bird's-Eye-View\n Recognition via Perspective Supervision", "abstract": " We present a novel bird's-eye-view (BEV) detector with perspective\nsupervision, which converges faster and better suits modern image backbones.\nExisting state-of-the-art BEV detectors are often tied to certain depth\npre-trained backbones like VoVNet, hindering the synergy between booming image\nbackbones and BEV detectors. To address this limitation, we prioritize easing\nthe optimization of BEV detectors by introducing perspective space supervision.\nTo this end, we propose a two-stage BEV detector, where proposals from the\nperspective head are fed into the bird's-eye-view head for final predictions.\nTo evaluate the effectiveness of our model, we conduct extensive ablation\nstudies focusing on the form of supervision and the generality of the proposed\ndetector. The proposed method is verified with a wide spectrum of traditional\nand modern image backbones and achieves new SoTA results on the large-scale\nnuScenes dataset. The code shall be released soon.\n", "date": "2022-11-18", "categories": []} {"id": "2211.09386", "title": "BEVDistill: Cross-Modal BEV Distillation for Multi-View 3D Object\n Detection", "abstract": " 3D object detection from multiple image views is a fundamental and\nchallenging task for visual scene understanding. Owing to its low cost and high\nefficiency, multi-view 3D object detection has demonstrated promising\napplication prospects. However, accurately detecting objects through\nperspective views is extremely difficult due to the lack of depth information.\nCurrent approaches tend to adopt heavy backbones for image encoders, making\nthem inapplicable for real-world deployment. Different from the images, LiDAR\npoints are superior in providing spatial cues, resulting in highly precise\nlocalization. In this paper, we explore the incorporation of LiDAR-based\ndetectors for multi-view 3D object detection. Instead of directly training a\ndepth prediction network, we unify the image and LiDAR features in the\nBird-Eye-View (BEV) space and adaptively transfer knowledge across\nnon-homogenous representations in a teacher-student paradigm. To this end, we\npropose \\textbf{BEVDistill}, a cross-modal BEV knowledge distillation (KD)\nframework for multi-view 3D object detection. Extensive experiments demonstrate\nthat the proposed method outperforms current KD approaches on a\nhighly-competitive baseline, BEVFormer, without introducing any extra cost in\nthe inference phase. Notably, our best model achieves 59.4 NDS on the nuScenes\ntest leaderboard, achieving new state-of-the-art in comparison with various\nimage-based detectors. Code will be available at\nhttps://github.com/zehuichen123/BEVDistill.\n", "date": "2022-11-17", "categories": []} {"id": "2211.09518", "title": "ImLiDAR: Cross-Sensor Dynamic Message Propagation Network for 3D Object\n Detection", "abstract": " LiDAR and camera, as two different sensors, supply geometric (point clouds)\nand semantic (RGB images) information of 3D scenes. However, it is still\nchallenging for existing methods to fuse data from the two cross sensors,\nmaking them complementary for quality 3D object detection (3OD). We propose\nImLiDAR, a new 3OD paradigm to narrow the cross-sensor discrepancies by\nprogressively fusing the multi-scale features of camera Images and LiDAR point\nclouds. ImLiDAR enables to provide the detection head with cross-sensor yet\nrobustly fused features. To achieve this, two core designs exist in ImLiDAR.\nFirst, we propose a cross-sensor dynamic message propagation module to combine\nthe best of the multi-scale image and point features. Second, we raise a direct\nset prediction problem that allows designing an effective set-based detector to\ntackle the inconsistency of the classification and localization confidences,\nand the sensitivity of hand-tuned hyperparameters. Besides, the novel set-based\ndetector can be detachable and easily integrated into various detection\nnetworks. Comparisons on both the KITTI and SUN-RGBD datasets show clear visual\nand numerical improvements of our ImLiDAR over twenty-three state-of-the-art\n3OD methods.\n", "date": "2022-11-17", "categories": []} {"id": "2211.07084", "title": "Boosting Semi-Supervised 3D Object Detection with Semi-Sampling", "abstract": " Current 3D object detection methods heavily rely on an enormous amount of\nannotations. Semi-supervised learning can be used to alleviate this issue.\nPrevious semi-supervised 3D object detection methods directly follow the\npractice of fully-supervised methods to augment labeled and unlabeled data,\nwhich is sub-optimal. In this paper, we design a data augmentation method for\nsemi-supervised learning, which we call Semi-Sampling. Specifically, we use\nground truth labels and pseudo labels to crop gt samples and pseudo samples on\nlabeled frames and unlabeled frames, respectively. Then we can generate a gt\nsample database and a pseudo sample database. When training a teacher-student\nsemi-supervised framework, we randomly select gt samples and pseudo samples to\nboth labeled frames and unlabeled frames, making a strong data augmentation for\nthem. Our semi-sampling can be regarded as an extension of gt-sampling to\nsemi-supervised learning. Our method is simple but effective. We consistently\nimprove state-of-the-art methods on ScanNet, SUN-RGBD, and KITTI benchmarks by\nlarge margins. For example, when training using only 10% labeled data on\nScanNet, we achieve 3.1 mAP and 6.4 mAP improvement upon 3DIoUMatch in terms of\nmAP@0.25 and mAP@0.5. When training using only 1% labeled data on KITTI, we\nboost 3DIoUMatch by 3.5 mAP, 6.7 mAP and 14.1 mAP on car, pedestrian and\ncyclist classes. Codes will be made publicly available at\nhttps://github.com/LittlePey/Semi-Sampling.\n", "date": "2022-11-14", "categories": []} {"id": "2211.08055", "title": "PAI3D: Painting Adaptive Instance-Prior for 3D Object Detection", "abstract": " 3D object detection is a critical task in autonomous driving. Recently\nmulti-modal fusion-based 3D object detection methods, which combine the\ncomplementary advantages of LiDAR and camera, have shown great performance\nimprovements over mono-modal methods. However, so far, no methods have\nattempted to utilize the instance-level contextual image semantics to guide the\n3D object detection. In this paper, we propose a simple and effective Painting\nAdaptive Instance-prior for 3D object detection (PAI3D) to fuse instance-level\nimage semantics flexibly with point cloud features. PAI3D is a multi-modal\nsequential instance-level fusion framework. It first extracts instance-level\nsemantic information from images, the extracted information, including objects\ncategorical label, point-to-object membership and object position, are then\nused to augment each LiDAR point in the subsequent 3D detection network to\nguide and improve detection performance. PAI3D outperforms the state-of-the-art\nwith a large margin on the nuScenes dataset, achieving 71.4 in mAP and 74.2 in\nNDS on the test split. Our comprehensive experiments show that instance-level\nimage semantics contribute the most to the performance gain, and PAI3D works\nwell with any good-quality instance segmentation models and any modern point\ncloud 3D encoders, making it a strong candidate for deployment on autonomous\nvehicles.\n", "date": "2022-11-15", "categories": []} {"id": "2211.08248", "title": "3D Cascade RCNN: High Quality Object Detection in Point Clouds", "abstract": " Recent progress on 2D object detection has featured Cascade RCNN, which\ncapitalizes on a sequence of cascade detectors to progressively improve\nproposal quality, towards high-quality object detection. However, there has not\nbeen evidence in support of building such cascade structures for 3D object\ndetection, a challenging detection scenario with highly sparse LiDAR point\nclouds. In this work, we present a simple yet effective cascade architecture,\nnamed 3D Cascade RCNN, that allocates multiple detectors based on the voxelized\npoint clouds in a cascade paradigm, pursuing higher quality 3D object detector\nprogressively. Furthermore, we quantitatively define the sparsity level of the\npoints within 3D bounding box of each object as the point completeness score,\nwhich is exploited as the task weight for each proposal to guide the learning\nof each stage detector. The spirit behind is to assign higher weights for\nhigh-quality proposals with relatively complete point distribution, while\ndown-weight the proposals with extremely sparse points that often incur noise\nduring training. This design of completeness-aware re-weighting elegantly\nupgrades the cascade paradigm to be better applicable for the sparse input\ndata, without increasing any FLOP budgets. Through extensive experiments on\nboth the KITTI dataset and Waymo Open Dataset, we validate the superiority of\nour proposed 3D Cascade RCNN, when comparing to state-of-the-art 3D object\ndetection techniques. The source code is publicly available at\n\\url{https://github.com/caiqi/Cascasde-3D}.\n", "date": "2022-11-15", "categories": []} {"id": "2211.08287", "title": "Towards 3D Object Detection with 2D Supervision", "abstract": " The great progress of 3D object detectors relies on large-scale data and 3D\nannotations. The annotation cost for 3D bounding boxes is extremely expensive\nwhile the 2D ones are easier and cheaper to collect. In this paper, we\nintroduce a hybrid training framework, enabling us to learn a visual 3D object\ndetector with massive 2D (pseudo) labels, even without 3D annotations. To break\nthrough the information bottleneck of 2D clues, we explore a new perspective:\nTemporal 2D Supervision. We propose a temporal 2D transformation to bridge the\n3D predictions with temporal 2D labels. Two steps, including homography wraping\nand 2D box deduction, are taken to transform the 3D predictions into 2D ones\nfor supervision. Experiments conducted on the nuScenes dataset show strong\nresults (nearly 90% of its fully-supervised performance) with only 25% 3D\nannotations. We hope our findings can provide new insights for using a large\nnumber of 2D annotations for 3D perception.\n", "date": "2022-11-15", "categories": []} {"id": "2206.01256", "title": "PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images", "abstract": " In this paper, we propose PETRv2, a unified framework for 3D perception from\nmulti-view images. Based on PETR, PETRv2 explores the effectiveness of temporal\nmodeling, which utilizes the temporal information of previous frames to boost\n3D object detection. More specifically, we extend the 3D position embedding (3D\nPE) in PETR for temporal modeling. The 3D PE achieves the temporal alignment on\nobject position of different frames. A feature-guided position encoder is\nfurther introduced to improve the data adaptability of 3D PE. To support for\nmulti-task learning (e.g., BEV segmentation and 3D lane detection), PETRv2\nprovides a simple yet effective solution by introducing task-specific queries,\nwhich are initialized under different spaces. PETRv2 achieves state-of-the-art\nperformance on 3D object detection, BEV segmentation and 3D lane detection.\nDetailed robustness analysis is also conducted on PETR framework. We hope\nPETRv2 can serve as a strong baseline for 3D perception. Code is available at\n\\url{https://github.com/megvii-research/PETR}.\n", "date": "2022-06-02", "categories": []} {"id": "2211.03407", "title": "3D Harmonic Loss: Towards Task-consistent and Time-friendly 3D Object\n Detection on Edge for V2X Orchestration", "abstract": " Edge computing-based 3D perception has received attention in intelligent\ntransportation systems (ITS) because real-time monitoring of traffic candidates\npotentially strengthens Vehicle-to-Everything (V2X) orchestration. Thanks to\nthe capability of precisely measuring the depth information on surroundings\nfrom LiDAR, the increasing studies focus on lidar-based 3D detection, which\nsignificantly promotes the development of 3D perception. Few methods met the\nreal-time requirement of edge deployment because of high computation-intensive\noperations. Moreover, an inconsistency problem of object detection remains\nuncovered in the pointcloud domain due to large sparsity. This paper thoroughly\nanalyses this problem, comprehensively roused by recent works on determining\ninconsistency problems in the image specialisation. Therefore, we proposed a 3D\nharmonic loss function to relieve the pointcloud based inconsistent\npredictions. Moreover, the feasibility of 3D harmonic loss is demonstrated from\na mathematical optimization perspective. The KITTI dataset and DAIR-V2X-I\ndataset are used for simulations, and our proposed method considerably improves\nthe performance than benchmark models. Further, the simulative deployment on an\nedge device (Jetson Xavier TX) validates our proposed model's efficiency.\n", "date": "2022-11-07", "categories": []} {"id": "2211.07171", "title": "Cross-Modality Knowledge Distillation Network for Monocular 3D Object\n Detection", "abstract": " Leveraging LiDAR-based detectors or real LiDAR point data to guide monocular\n3D detection has brought significant improvement, e.g., Pseudo-LiDAR methods.\nHowever, the existing methods usually apply non-end-to-end training strategies\nand insufficiently leverage the LiDAR information, where the rich potential of\nthe LiDAR data has not been well exploited. In this paper, we propose the\nCross-Modality Knowledge Distillation (CMKD) network for monocular 3D detection\nto efficiently and directly transfer the knowledge from LiDAR modality to image\nmodality on both features and responses. Moreover, we further extend CMKD as a\nsemi-supervised training framework by distilling knowledge from large-scale\nunlabeled data and significantly boost the performance. Until submission, CMKD\nranks $1^{st}$ among the monocular 3D detectors with publications on both KITTI\n$test$ set and Waymo $val$ set with significant performance gains compared to\nprevious state-of-the-art methods.\n", "date": "2022-11-14", "categories": []} {"id": "2207.03720", "title": "Bounding Box Disparity: 3D Metrics for Object Detection With Full Degree\n of Freedom", "abstract": " The most popular evaluation metric for object detection in 2D images is\nIntersection over Union (IoU). Existing implementations of the IoU metric for\n3D object detection usually neglect one or more degrees of freedom. In this\npaper, we first derive the analytic solution for three dimensional bounding\nboxes. As a second contribution, a closed-form solution of the volume-to-volume\ndistance is derived. Finally, the Bounding Box Disparity is proposed as a\ncombined positive continuous metric. We provide open source implementations of\nthe three metrics as standalone python functions, as well as extensions to the\nOpen3D library and as ROS nodes.\n", "date": "2022-07-08", "categories": []} {"id": "2211.05233", "title": "Plausibility Verification For 3D Object Detectors Using Energy-Based\n Optimization", "abstract": " Environmental perception obtained via object detectors have no predictable\nsafety layer encoded into their model schema, which creates the question of\ntrustworthiness about the system's prediction. As can be seen from recent\nadversarial attacks, most of the current object detection networks are\nvulnerable to input tampering, which in the real world could compromise the\nsafety of autonomous vehicles. The problem would be amplified even more when\nuncertainty errors could not propagate into the submodules, if these are not a\npart of the end-to-end system design. To address these concerns, a parallel\nmodule which verifies the predictions of the object proposals coming out of\nDeep Neural Networks are required. This work aims to verify 3D object proposals\nfrom MonoRUn model by proposing a plausibility framework that leverages cross\nsensor streams to reduce false positives. The verification metric being\nproposed uses prior knowledge in the form of four different energy functions,\neach utilizing a certain prior to output an energy value leading to a\nplausibility justification for the hypothesis under consideration. We also\nemploy a novel two-step schema to improve the optimization of the composite\nenergy function representing the energy model.\n", "date": "2022-11-02", "categories": []} {"id": "2211.05580", "title": "Hyperbolic Cosine Transformer for LiDAR 3D Object Detection", "abstract": " Recently, Transformer has achieved great success in computer vision. However,\nit is constrained because the spatial and temporal complexity grows\nquadratically with the number of large points in 3D object detection\napplications. Previous point-wise methods are suffering from time consumption\nand limited receptive fields to capture information among points. In this\npaper, we propose a two-stage hyperbolic cosine transformer (ChTR3D) for 3D\nobject detection from LiDAR point clouds. The proposed ChTR3D refines proposals\nby applying cosh-attention in linear computation complexity to encode rich\ncontextual relationships among points. The cosh-attention module reduces the\nspace and time complexity of the attention operation. The traditional softmax\noperation is replaced by non-negative ReLU activation and\nhyperbolic-cosine-based operator with re-weighting mechanism. Extensive\nexperiments on the widely used KITTI dataset demonstrate that, compared with\nvanilla attention, the cosh-attention significantly improves the inference\nspeed with competitive performance. Experiment results show that, among\ntwo-stage state-of-the-art methods using point-level features, the proposed\nChTR3D is the fastest one.\n", "date": "2022-11-10", "categories": []} {"id": "2211.01556", "title": "Ground Plane Matters: Picking Up Ground Plane Prior in Monocular 3D\n Object Detection", "abstract": " The ground plane prior is a very informative geometry clue in monocular 3D\nobject detection (M3OD). However, it has been neglected by most mainstream\nmethods. In this paper, we identify two key factors that limit the\napplicability of ground plane prior: the projection point localization issue\nand the ground plane tilt issue. To pick up the ground plane prior for M3OD, we\npropose a Ground Plane Enhanced Network (GPENet) which resolves both issues at\none go. For the projection point localization issue, instead of using the\nbottom vertices or bottom center of the 3D bounding box (BBox), we leverage the\nobject's ground contact points, which are explicit pixels in the image and easy\nfor the neural network to detect. For the ground plane tilt problem, our GPENet\nestimates the horizon line in the image and derives a novel mathematical\nexpression to accurately estimate the ground plane equation. An unsupervised\nvertical edge mining algorithm is also proposed to address the occlusion of the\nhorizon line. Furthermore, we design a novel 3D bounding box deduction method\nbased on a dynamic back projection algorithm, which could take advantage of the\naccurate contact points and the ground plane equation. Additionally, using only\nM3OD labels, contact point and horizon line pseudo labels can be easily\ngenerated with NO extra data collection and label annotation cost. Extensive\nexperiments on the popular KITTI benchmark show that our GPENet can outperform\nother methods and achieve state-of-the-art performance, well demonstrating the\neffectiveness and the superiority of the proposed approach. Moreover, our\nGPENet works better than other methods in cross-dataset evaluation on the\nnuScenes dataset. Our code and models will be published.\n", "date": "2022-11-03", "categories": []} {"id": "2211.01664", "title": "PointSee: Image Enhances Point Cloud", "abstract": " There is a trend to fuse multi-modal information for 3D object detection\n(3OD). However, the challenging problems of low lightweightness, poor\nflexibility of plug-and-play, and inaccurate alignment of features are still\nnot well-solved, when designing multi-modal fusion newtorks. We propose\nPointSee, a lightweight, flexible and effective multi-modal fusion solution to\nfacilitate various 3OD networks by semantic feature enhancement of LiDAR point\nclouds assembled with scene images. Beyond the existing wisdom of 3OD, PointSee\nconsists of a hidden module (HM) and a seen module (SM): HM decorates LiDAR\npoint clouds using 2D image information in an offline fusion manner, leading to\nminimal or even no adaptations of existing 3OD networks; SM further enriches\nthe LiDAR point clouds by acquiring point-wise representative semantic\nfeatures, leading to enhanced performance of existing 3OD networks. Besides the\nnew architecture of PointSee, we propose a simple yet efficient training\nstrategy, to ease the potential inaccurate regressions of 2D object detection\nnetworks. Extensive experiments on the popular outdoor/indoor benchmarks show\nnumerical improvements of our PointSee over twenty-two state-of-the-arts.\n", "date": "2022-11-03", "categories": []} {"id": "2211.00746", "title": "3DMODT: Attention-Guided Affinities for Joint Detection & Tracking in 3D\n Point Clouds", "abstract": " We propose a method for joint detection and tracking of multiple objects in\n3D point clouds, a task conventionally treated as a two-step process comprising\nobject detection followed by data association. Our method embeds both steps\ninto a single end-to-end trainable network eliminating the dependency on\nexternal object detectors. Our model exploits temporal information employing\nmultiple frames to detect objects and track them in a single network, thereby\nmaking it a utilitarian formulation for real-world scenarios. Computing\naffinity matrix by employing features similarity across consecutive point cloud\nscans forms an integral part of visual tracking. We propose an attention-based\nrefinement module to refine the affinity matrix by suppressing erroneous\ncorrespondences. The module is designed to capture the global context in\naffinity matrix by employing self-attention within each affinity matrix and\ncross-attention across a pair of affinity matrices. Unlike competing\napproaches, our network does not require complex post-processing algorithms,\nand processes raw LiDAR frames to directly output tracking results. We\ndemonstrate the effectiveness of our method on the three tracking benchmarks:\nJRDB, Waymo, and KITTI. Experimental evaluations indicate the ability of our\nmodel to generalize well across datasets.\n", "date": "2022-11-01", "categories": []} {"id": "2211.01142", "title": "OPA-3D: Occlusion-Aware Pixel-Wise Aggregation for Monocular 3D Object\n Detection", "abstract": " Despite monocular 3D object detection having recently made a significant leap\nforward thanks to the use of pre-trained depth estimators for pseudo-LiDAR\nrecovery, such two-stage methods typically suffer from overfitting and are\nincapable of explicitly encapsulating the geometric relation between depth and\nobject bounding box. To overcome this limitation, we instead propose OPA-3D, a\nsingle-stage, end-to-end, Occlusion-Aware Pixel-Wise Aggregation network that\nto jointly estimate dense scene depth with depth-bounding box residuals and\nobject bounding boxes, allowing a two-stream detection of 3D objects, leading\nto significantly more robust detections. Thereby, the geometry stream denoted\nas the Geometry Stream, combines visible depth and depth-bounding box residuals\nto recover the object bounding box via explicit occlusion-aware optimization.\nIn addition, a bounding box based geometry projection scheme is employed in an\neffort to enhance distance perception. The second stream, named as the Context\nStream, directly regresses 3D object location and size. This novel two-stream\nrepresentation further enables us to enforce cross-stream consistency terms\nwhich aligns the outputs of both streams, improving the overall performance.\nExtensive experiments on the public benchmark demonstrate that OPA-3D\noutperforms state-of-the-art methods on the main Car category, whilst keeping a\nreal-time inference speed. We plan to release all codes and trained models\nsoon.\n", "date": "2022-11-02", "categories": []} {"id": "2208.12079", "title": "Bridging the View Disparity Between Radar and Camera Features for\n Multi-modal Fusion 3D Object Detection", "abstract": " Environmental perception with the multi-modal fusion of radar and camera is\ncrucial in autonomous driving to increase accuracy, completeness, and\nrobustness. This paper focuses on utilizing millimeter-wave (MMW) radar and\ncamera sensor fusion for 3D object detection. A novel method that realizes the\nfeature-level fusion under the bird's-eye view (BEV) for a better feature\nrepresentation is proposed. Firstly, radar points are augmented with temporal\naccumulation and sent to a spatial-temporal encoder for radar feature\nextraction. Meanwhile, multi-scale image 2D features which adapt to various\nspatial scales are obtained by image backbone and neck model. Then, image\nfeatures are transformed to BEV with the designed view transformer. In\naddition, this work fuses the multi-modal features with a two-stage fusion\nmodel called point-fusion and ROI-fusion, respectively. Finally, a detection\nhead regresses objects category and 3D locations. Experimental results\ndemonstrate that the proposed method realizes the state-of-the-art (SOTA)\nperformance under the most crucial detection metrics-mean average precision\n(mAP) and nuScenes detection score (NDS) on the challenging nuScenes dataset.\n", "date": "2022-08-25", "categories": []} {"id": "2210.16574", "title": "Boosting Monocular 3D Object Detection with Object-Centric Auxiliary\n Depth Supervision", "abstract": " Recent advances in monocular 3D detection leverage a depth estimation network\nexplicitly as an intermediate stage of the 3D detection network. Depth map\napproaches yield more accurate depth to objects than other methods thanks to\nthe depth estimation network trained on a large-scale dataset. However, depth\nmap approaches can be limited by the accuracy of the depth map, and\nsequentially using two separated networks for depth estimation and 3D detection\nsignificantly increases computation cost and inference time. In this work, we\npropose a method to boost the RGB image-based 3D detector by jointly training\nthe detection network with a depth prediction loss analogous to the depth\nestimation task. In this way, our 3D detection network can be supervised by\nmore depth supervision from raw LiDAR points, which does not require any human\nannotation cost, to estimate accurate depth without explicitly predicting the\ndepth map. Our novel object-centric depth prediction loss focuses on depth\naround foreground objects, which is important for 3D object detection, to\nleverage pixel-wise depth supervision in an object-centric manner. Our depth\nregression model is further trained to predict the uncertainty of depth to\nrepresent the 3D confidence of objects. To effectively train the 3D detector\nwith raw LiDAR points and to enable end-to-end training, we revisit the\nregression target of 3D objects and design a network architecture. Extensive\nexperiments on KITTI and nuScenes benchmarks show that our method can\nsignificantly boost the monocular image-based 3D detector to outperform depth\nmap approaches while maintaining the real-time inference speed.\n", "date": "2022-10-29", "categories": []} {"id": "2210.17252", "title": "Multi-Camera Calibration Free BEV Representation for 3D Object Detection", "abstract": " In advanced paradigms of autonomous driving, learning Bird's Eye View (BEV)\nrepresentation from surrounding views is crucial for multi-task framework.\nHowever, existing methods based on depth estimation or camera-driven attention\nare not stable to obtain transformation under noisy camera parameters, mainly\nwith two challenges, accurate depth prediction and calibration. In this work,\nwe present a completely Multi-Camera Calibration Free Transformer (CFT) for\nrobust BEV representation, which focuses on exploring implicit mapping, not\nrelied on camera intrinsics and extrinsics. To guide better feature learning\nfrom image views to BEV, CFT mines potential 3D information in BEV via our\ndesigned position-aware enhancement (PA). Instead of camera-driven point-wise\nor global transformation, for interaction within more effective region and\nlower computation cost, we propose a view-aware attention which also reduces\nredundant computation and promotes converge. CFT achieves 49.7% NDS on the\nnuScenes detection task leaderboard, which is the first work removing camera\nparameters, comparable to other geometry-guided methods. Without temporal input\nand other modal information, CFT achieves second highest performance with a\nsmaller image input 1600 * 640. Thanks to view-attention variant, CFT reduces\nmemory and transformer FLOPs for vanilla attention by about 12% and 60%,\nrespectively, with improved NDS by 1.0%. Moreover, its natural robustness to\nnoisy camera parameters makes CFT more competitive.\n", "date": "2022-10-31", "categories": []} {"id": "2210.15933", "title": "PSFormer: Point Transformer for 3D Salient Object Detection", "abstract": " We propose PSFormer, an effective point transformer model for 3D salient\nobject detection. PSFormer is an encoder-decoder network that takes full\nadvantage of transformers to model the contextual information in both\nmulti-scale point- and scene-wise manners. In the encoder, we develop a Point\nContext Transformer (PCT) module to capture region contextual features at the\npoint level; PCT contains two different transformers to excavate the\nrelationship among points. In the decoder, we develop a Scene Context\nTransformer (SCT) module to learn context representations at the scene level;\nSCT contains both Upsampling-and-Transformer blocks and Multi-context\nAggregation units to integrate the global semantic and multi-level features\nfrom the encoder into the global scene context. Experiments show clear\nimprovements of PSFormer over its competitors and validate that PSFormer is\nmore robust to challenging cases such as small objects, multiple objects, and\nobjects with complex structures.\n", "date": "2022-10-28", "categories": []} {"id": "2210.15316", "title": "MSF3DDETR: Multi-Sensor Fusion 3D Detection Transformer for Autonomous\n Driving", "abstract": " 3D object detection is a significant task for autonomous driving. Recently\nwith the progress of vision transformers, the 2D object detection problem is\nbeing treated with the set-to-set loss. Inspired by these approaches on 2D\nobject detection and an approach for multi-view 3D object detection DETR3D, we\npropose MSF3DDETR: Multi-Sensor Fusion 3D Detection Transformer architecture to\nfuse image and LiDAR features to improve the detection accuracy. Our end-to-end\nsingle-stage, anchor-free and NMS-free network takes in multi-view images and\nLiDAR point clouds and predicts 3D bounding boxes. Firstly, we link the object\nqueries learnt from data to the image and LiDAR features using a novel\nMSF3DDETR cross-attention block. Secondly, the object queries interacts with\neach other in multi-head self-attention block. Finally, MSF3DDETR block is\nrepeated for $L$ number of times to refine the object queries. The MSF3DDETR\nnetwork is trained end-to-end on the nuScenes dataset using Hungarian algorithm\nbased bipartite matching and set-to-set loss inspired by DETR. We present both\nquantitative and qualitative results which are competitive to the\nstate-of-the-art approaches.\n", "date": "2022-10-27", "categories": []} {"id": "2210.15365", "title": "Li3DeTr: A LiDAR based 3D Detection Transformer", "abstract": " Inspired by recent advances in vision transformers for object detection, we\npropose Li3DeTr, an end-to-end LiDAR based 3D Detection Transformer for\nautonomous driving, that inputs LiDAR point clouds and regresses 3D bounding\nboxes. The LiDAR local and global features are encoded using sparse convolution\nand multi-scale deformable attention respectively. In the decoder head,\nfirstly, in the novel Li3DeTr cross-attention block, we link the LiDAR global\nfeatures to 3D predictions leveraging the sparse set of object queries learnt\nfrom the data. Secondly, the object query interactions are formulated using\nmulti-head self-attention. Finally, the decoder layer is repeated $L_{dec}$\nnumber of times to refine the object queries. Inspired by DETR, we employ\nset-to-set loss to train the Li3DeTr network. Without bells and whistles, the\nLi3DeTr network achieves 61.3% mAP and 67.6% NDS surpassing the\nstate-of-the-art methods with non-maximum suppression (NMS) on the nuScenes\ndataset and it also achieves competitive performance on the KITTI dataset. We\nalso employ knowledge distillation (KD) using a teacher and student model that\nslightly improves the performance of our network.\n", "date": "2022-10-27", "categories": []} {"id": "2210.10983", "title": "PSA-Det3D: Pillar Set Abstraction for 3D object Detection", "abstract": " Small object detection for 3D point cloud is a challenging problem because of\ntwo limitations: (1) Perceiving small objects is much more diffcult than normal\nobjects due to the lack of valid points. (2) Small objects are easily blocked\nwhich breaks the shape of their meshes in 3D point cloud. In this paper, we\npropose a pillar set abstraction (PSA) and foreground point compensation (FPC)\nand design a point-based detection network, PSA-Det3D, to improve the detection\nperformance for small object. The PSA embeds a pillar query operation on the\nbasis of set abstraction (SA) to expand its receptive field of the network,\nwhich can aggregate point-wise features effectively. To locate more occluded\nobjects, we persent a proposal generation layer consisting of a foreground\npoint segmentation and a FPC module. Both the foreground points and the\nestimated centers are finally fused together to generate the detection result.\nThe experiments on the KITTI 3D detection benchmark show that our proposed\nPSA-Det3D outperforms other algorithms with high accuracy for small object\ndetection.\n", "date": "2022-10-20", "categories": []} {"id": "2210.13488", "title": "LidarAugment: Searching for Scalable 3D LiDAR Data Augmentations", "abstract": " Data augmentations are important in training high-performance 3D object\ndetectors for point clouds. Despite recent efforts on designing new data\naugmentations, perhaps surprisingly, most state-of-the-art 3D detectors only\nuse a few simple data augmentations. In particular, different from 2D image\ndata augmentations, 3D data augmentations need to account for different\nrepresentations of input data and require being customized for different\nmodels, which introduces significant overhead. In this paper, we resort to a\nsearch-based approach, and propose LidarAugment, a practical and effective data\naugmentation strategy for 3D object detection. Unlike previous approaches where\nall augmentation policies are tuned in an exponentially large search space, we\npropose to factorize and align the search space of each data augmentation,\nwhich cuts down the 20+ hyperparameters to 2, and significantly reduces the\nsearch complexity. We show LidarAugment can be customized for different model\narchitectures with different input representations by a simple 2D grid search,\nand consistently improve both convolution-based UPillars/StarNet/RSN and\ntransformer-based SWFormer. Furthermore, LidarAugment mitigates overfitting and\nallows us to scale up 3D detectors to much larger capacity. In particular, by\ncombining with latest 3D detectors, our LidarAugment achieves a new\nstate-of-the-art 74.8 mAPH L2 on Waymo Open Dataset.\n", "date": "2022-10-24", "categories": []} {"id": "2210.13428", "title": "PseudoAugment: Learning to Use Unlabeled Data for Data Augmentation in\n Point Clouds", "abstract": " Data augmentation is an important technique to improve data efficiency and\nsave labeling cost for 3D detection in point clouds. Yet, existing augmentation\npolicies have so far been designed to only utilize labeled data, which limits\nthe data diversity. In this paper, we recognize that pseudo labeling and data\naugmentation are complementary, thus propose to leverage unlabeled data for\ndata augmentation to enrich the training data. In particular, we design three\nnovel pseudo-label based data augmentation policies (PseudoAugments) to fuse\nboth labeled and pseudo-labeled scenes, including frames (PseudoFrame), objecta\n(PseudoBBox), and background (PseudoBackground). PseudoAugments outperforms\npseudo labeling by mitigating pseudo labeling errors and generating diverse\nfused training scenes. We demonstrate PseudoAugments generalize across\npoint-based and voxel-based architectures, different model capacity and both\nKITTI and Waymo Open Dataset. To alleviate the cost of hyperparameter tuning\nand iterative pseudo labeling, we develop a population-based data augmentation\nframework for 3D detection, named AutoPseudoAugment. Unlike previous works that\nperform pseudo-labeling offline, our framework performs PseudoAugments and\nhyperparameter tuning in one shot to reduce computational cost. Experimental\nresults on the large-scale Waymo Open Dataset show our method outperforms\nstate-of-the-art auto data augmentation method (PPBA) and self-training method\n(pseudo labeling). In particular, AutoPseudoAugment is about 3X and 2X data\nefficient on vehicle and pedestrian tasks compared to prior arts. Notably,\nAutoPseudoAugment nearly matches the full dataset training results, with just\n10% of the labeled run segments on the vehicle detection task.\n", "date": "2022-10-24", "categories": []} {"id": "2210.11590", "title": "XC: Exploring Quantitative Use Cases for Explanations in 3D Object\n Detection", "abstract": " Explainable AI (XAI) methods are frequently applied to obtain qualitative\ninsights about deep models' predictions. However, such insights need to be\ninterpreted by a human observer to be useful. In this paper, we aim to use\nexplanations directly to make decisions without human observers. We adopt two\ngradient-based explanation methods, Integrated Gradients (IG) and backprop, for\nthe task of 3D object detection. Then, we propose a set of quantitative\nmeasures, named Explanation Concentration (XC) scores, that can be used for\ndownstream tasks. These scores quantify the concentration of attributions\nwithin the boundaries of detected objects. We evaluate the effectiveness of XC\nscores via the task of distinguishing true positive (TP) and false positive\n(FP) detected objects in the KITTI and Waymo datasets. The results demonstrate\nan improvement of more than 100\\% on both datasets compared to other heuristics\nsuch as random guesses and the number of LiDAR points in the bounding box,\nraising confidence in XC's potential for application in more use cases. Our\nresults also indicate that computationally expensive XAI methods like IG may\nnot be more valuable when used quantitatively compare to simpler methods.\n", "date": "2022-10-20", "categories": []} {"id": "2210.09267", "title": "CramNet: Camera-Radar Fusion with Ray-Constrained Cross-Attention for\n Robust 3D Object Detection", "abstract": " Robust 3D object detection is critical for safe autonomous driving. Camera\nand radar sensors are synergistic as they capture complementary information and\nwork well under different environmental conditions. Fusing camera and radar\ndata is challenging, however, as each of the sensors lacks information along a\nperpendicular axis, that is, depth is unknown to camera and elevation is\nunknown to radar. We propose the camera-radar matching network CramNet, an\nefficient approach to fuse the sensor readings from camera and radar in a joint\n3D space. To leverage radar range measurements for better camera depth\npredictions, we propose a novel ray-constrained cross-attention mechanism that\nresolves the ambiguity in the geometric correspondences between camera features\nand radar features. Our method supports training with sensor modality dropout,\nwhich leads to robust 3D object detection, even when a camera or radar sensor\nsuddenly malfunctions on a vehicle. We demonstrate the effectiveness of our\nfusion approach through extensive experiments on the RADIATE dataset, one of\nthe few large-scale datasets that provide radar radio frequency imagery. A\ncamera-only variant of our method achieves competitive performance in monocular\n3D object detection on the Waymo Open Dataset.\n", "date": "2022-10-17", "categories": []} {"id": "2210.09615", "title": "Homogeneous Multi-modal Feature Fusion and Interaction for 3D Object\n Detection", "abstract": " Multi-modal 3D object detection has been an active research topic in\nautonomous driving. Nevertheless, it is non-trivial to explore the cross-modal\nfeature fusion between sparse 3D points and dense 2D pixels. Recent approaches\neither fuse the image features with the point cloud features that are projected\nonto the 2D image plane or combine the sparse point cloud with dense image\npixels. These fusion approaches often suffer from severe information loss, thus\ncausing sub-optimal performance. To address these problems, we construct the\nhomogeneous structure between the point cloud and images to avoid projective\ninformation loss by transforming the camera features into the LiDAR 3D space.\nIn this paper, we propose a homogeneous multi-modal feature fusion and\ninteraction method (HMFI) for 3D object detection. Specifically, we first\ndesign an image voxel lifter module (IVLM) to lift 2D image features into the\n3D space and generate homogeneous image voxel features. Then, we fuse the\nvoxelized point cloud features with the image features from different regions\nby introducing the self-attention based query fusion mechanism (QFM). Next, we\npropose a voxel feature interaction module (VFIM) to enforce the consistency of\nsemantic information from identical objects in the homogeneous point cloud and\nimage voxel representations, which can provide object-level alignment guidance\nfor cross-modal feature fusion and strengthen the discriminative ability in\ncomplex backgrounds. We conduct extensive experiments on the KITTI and Waymo\nOpen Dataset, and the proposed HMFI achieves better performance compared with\nthe state-of-the-art multi-modal methods. Particularly, for the 3D detection of\ncyclist on the KITTI benchmark, HMFI surpasses all the published algorithms by\na large margin.\n", "date": "2022-10-18", "categories": []} {"id": "2210.07811", "title": "SAILOR: Scaling Anchors via Insights into Latent Object Representation", "abstract": " LiDAR 3D object detection models are inevitably biased towards their training\ndataset. The detector clearly exhibits this bias when employed on a target\ndataset, particularly towards object sizes. However, object sizes vary heavily\nbetween domains due to, for instance, different labeling policies or\ngeographical locations. State-of-the-art unsupervised domain adaptation\napproaches outsource methods to overcome the object size bias. Mainstream size\nadaptation approaches exploit target domain statistics, contradicting the\noriginal unsupervised assumption. Our novel unsupervised anchor calibration\nmethod addresses this limitation. Given a model trained on the source data, we\nestimate the optimal target anchors in a completely unsupervised manner. The\nmain idea stems from an intuitive observation: by varying the anchor sizes for\nthe target domain, we inevitably introduce noise or even remove valuable object\ncues. The latent object representation, perturbed by the anchor size, is\nclosest to the learned source features only under the optimal target anchors.\nWe leverage this observation for anchor size optimization. Our experimental\nresults show that, without any retraining, we achieve competitive results even\ncompared to state-of-the-art weakly-supervised size adaptation approaches. In\naddition, our anchor calibration can be combined with such existing methods,\nmaking them completely unsupervised.\n", "date": "2022-10-14", "categories": []} {"id": "2210.08375", "title": "Improving the Intra-class Long-tail in 3D Detection via Rare Example\n Mining", "abstract": " Continued improvements in deep learning architectures have steadily advanced\nthe overall performance of 3D object detectors to levels on par with humans for\ncertain tasks and datasets, where the overall performance is mostly driven by\ncommon examples. However, even the best performing models suffer from the most\nnaive mistakes when it comes to rare examples that do not appear frequently in\nthe training data, such as vehicles with irregular geometries. Most studies in\nthe long-tail literature focus on class-imbalanced classification problems with\nknown imbalanced label counts per class, but they are not directly applicable\nto the intra-class long-tail examples in problems with large intra-class\nvariations such as 3D object detection, where instances with the same class\nlabel can have drastically varied properties such as shapes and sizes. Other\nworks propose to mitigate this problem using active learning based on the\ncriteria of uncertainty, difficulty, or diversity. In this study, we identify a\nnew conceptual dimension - rareness - to mine new data for improving the\nlong-tail performance of models. We show that rareness, as opposed to\ndifficulty, is the key to data-centric improvements for 3D detectors, since\nrareness is the result of a lack in data support while difficulty is related to\nthe fundamental ambiguity in the problem. We propose a general and effective\nmethod to identify the rareness of objects based on density estimation in the\nfeature space using flow models, and propose a principled cost-aware\nformulation for mining rare object tracks, which improves overall model\nperformance, but more importantly - significantly improves the performance for\nrare objects (by 30.97\\%\n", "date": "2022-10-15", "categories": []} {"id": "2210.07372", "title": "SWFormer: Sparse Window Transformer for 3D Object Detection in Point\n Clouds", "abstract": " 3D object detection in point clouds is a core component for modern robotics\nand autonomous driving systems. A key challenge in 3D object detection comes\nfrom the inherent sparse nature of point occupancy within the 3D scene. In this\npaper, we propose Sparse Window Transformer (SWFormer ), a scalable and\naccurate model for 3D object detection, which can take full advantage of the\nsparsity of point clouds. Built upon the idea of window-based Transformers,\nSWFormer converts 3D points into sparse voxels and windows, and then processes\nthese variable-length sparse windows efficiently using a bucketing scheme. In\naddition to self-attention within each spatial window, our SWFormer also\ncaptures cross-window correlation with multi-scale feature fusion and window\nshifting operations. To further address the unique challenge of detecting 3D\nobjects accurately from sparse features, we propose a new voxel diffusion\ntechnique. Experimental results on the Waymo Open Dataset show our SWFormer\nachieves state-of-the-art 73.36 L2 mAPH on vehicle and pedestrian for 3D object\ndetection on the official test set, outperforming all previous single-stage and\ntwo-stage models, while being much more efficient.\n", "date": "2022-10-13", "categories": []} {"id": "2210.07424", "title": "Autoregressive Uncertainty Modeling for 3D Bounding Box Prediction", "abstract": " 3D bounding boxes are a widespread intermediate representation in many\ncomputer vision applications. However, predicting them is a challenging task,\nlargely due to partial observability, which motivates the need for a strong\nsense of uncertainty. While many recent methods have explored better\narchitectures for consuming sparse and unstructured point cloud data, we\nhypothesize that there is room for improvement in the modeling of the output\ndistribution and explore how this can be achieved using an autoregressive\nprediction head. Additionally, we release a simulated dataset, COB-3D, which\nhighlights new types of ambiguity that arise in real-world robotics\napplications, where 3D bounding box prediction has largely been underexplored.\nWe propose methods for leveraging our autoregressive model to make high\nconfidence predictions and meaningful uncertainty measures, achieving strong\nresults on SUN-RGBD, Scannet, KITTI, and our new dataset.\n", "date": "2022-10-13", "categories": []} {"id": "2206.00630", "title": "Unifying Voxel-based Representation with Transformer for 3D Object\n Detection", "abstract": " In this work, we present a unified framework for multi-modality 3D object\ndetection, named UVTR. The proposed method aims to unify multi-modality\nrepresentations in the voxel space for accurate and robust single- or\ncross-modality 3D detection. To this end, the modality-specific space is first\ndesigned to represent different inputs in the voxel feature space. Different\nfrom previous work, our approach preserves the voxel space without height\ncompression to alleviate semantic ambiguity and enable spatial connections. To\nmake full use of the inputs from different sensors, the cross-modality\ninteraction is then proposed, including knowledge transfer and modality fusion.\nIn this way, geometry-aware expressions in point clouds and context-rich\nfeatures in images are well utilized for better performance and robustness. The\ntransformer decoder is applied to efficiently sample features from the unified\nspace with learnable positions, which facilitates object-level interactions. In\ngeneral, UVTR presents an early attempt to represent different modalities in a\nunified framework. It surpasses previous work in single- or multi-modality\nentries. The proposed method achieves leading performance in the nuScenes test\nset for both object detection and the following object tracking task. Code is\nmade publicly available at https://github.com/dvlab-research/UVTR.\n", "date": "2022-06-01", "categories": []} {"id": "2210.05896", "title": "Common Corruption Robustness of Point Cloud Detectors: Benchmark and\n Enhancement", "abstract": " Object detection through LiDAR-based point cloud has recently been important\nin autonomous driving. Although achieving high accuracy on public benchmarks,\nthe state-of-the-art detectors may still go wrong and cause a heavy loss due to\nthe widespread corruptions in the real world like rain, snow, sensor noise,\netc. Nevertheless, there is a lack of a large-scale dataset covering diverse\nscenes and realistic corruption types with different severities to develop\npractical and robust point cloud detectors, which is challenging due to the\nheavy collection costs. To alleviate the challenge and start the first step for\nrobust point cloud detection, we propose the physical-aware simulation methods\nto generate degraded point clouds under different real-world common\ncorruptions. Then, for the first attempt, we construct a benchmark based on the\nphysical-aware common corruptions for point cloud detectors, which contains a\ntotal of 1,122,150 examples covering 7,481 scenes, 25 common corruption types,\nand 6 severities. With such a novel benchmark, we conduct extensive empirical\nstudies on 8 state-of-the-art detectors that contain 6 different detection\nframeworks. Thus we get several insight observations revealing the\nvulnerabilities of the detectors and indicating the enhancement directions.\nMoreover, we further study the effectiveness of existing robustness enhancement\nmethods based on data augmentation and data denoising. The benchmark can\npotentially be a new platform for evaluating point cloud detectors, opening a\ndoor for developing novel robustness enhancement methods.\n", "date": "2022-10-12", "categories": []} {"id": "2210.04264", "title": "CAGroup3D: Class-Aware Grouping for 3D Object Detection on Point Clouds", "abstract": " We present a novel two-stage fully sparse convolutional 3D object detection\nframework, named CAGroup3D. Our proposed method first generates some\nhigh-quality 3D proposals by leveraging the class-aware local group strategy on\nthe object surface voxels with the same semantic predictions, which considers\nsemantic consistency and diverse locality abandoned in previous bottom-up\napproaches. Then, to recover the features of missed voxels due to incorrect\nvoxel-wise segmentation, we build a fully sparse convolutional RoI pooling\nmodule to directly aggregate fine-grained spatial information from backbone for\nfurther proposal refinement. It is memory-and-computation efficient and can\nbetter encode the geometry-specific features of each 3D proposal. Our model\nachieves state-of-the-art 3D detection performance with remarkable gains of\n+\\textit{3.6\\%} on ScanNet V2 and +\\textit{2.6}\\% on SUN RGB-D in term of\nmAP@0.25. Code will be available at https://github.com/Haiyang-W/CAGroup3D.\n", "date": "2022-10-09", "categories": []} {"id": "2206.01772", "title": "Radar Guided Dynamic Visual Attention for Resource-Efficient RGB Object\n Detection", "abstract": " An autonomous system's perception engine must provide an accurate\nunderstanding of the environment for it to make decisions. Deep learning based\nobject detection networks experience degradation in the performance and\nrobustness for small and far away objects due to a reduction in object's\nfeature map as we move to higher layers of the network. In this work, we\npropose a novel radar-guided spatial attention for RGB images to improve the\nperception quality of autonomous vehicles operating in a dynamic environment.\nIn particular, our method improves the perception of small and long range\nobjects, which are often not detected by the object detectors in RGB mode. The\nproposed method consists of two RGB object detectors, namely the Primary\ndetector and a lightweight Secondary detector. The primary detector takes a\nfull RGB image and generates primary detections. Next, the radar proposal\nframework creates regions of interest (ROIs) for object proposals by projecting\nthe radar point cloud onto the 2D RGB image. These ROIs are cropped and fed to\nthe secondary detector to generate secondary detections which are then fused\nwith the primary detections via non-maximum suppression. This method helps in\nrecovering the small objects by preserving the object's spatial features\nthrough an increase in their receptive field. We evaluate our fusion method on\nthe challenging nuScenes dataset and show that our fusion method with SSD-lite\nas primary and secondary detector improves the baseline primary yolov3\ndetector's recall by 14% while requiring three times fewer computational\nresources.\n", "date": "2022-06-03", "categories": []} {"id": "2210.03331", "title": "Resolving Class Imbalance for LiDAR-based Object Detector by Dynamic\n Weight Average and Contextual Ground Truth Sampling", "abstract": " An autonomous driving system requires a 3D object detector, which must\nperceive all present road agents reliably to navigate an environment safely.\nHowever, real-world driving datasets often suffer from the problem of data\nimbalance, which causes difficulties in training a model that works well across\nall classes, resulting in an undesired imbalanced sub-optimal performance. In\nthis work, we propose a method to address this data imbalance problem. Our\nmethod consists of two main components: (i) a LiDAR-based 3D object detector\nwith per-class multiple detection heads where losses from each head are\nmodified by dynamic weight average to be balanced. (ii) Contextual ground truth\n(GT) sampling, where we improve conventional GT sampling techniques by\nleveraging semantic information to augment point cloud with sampled ground\ntruth GT objects. Our experiment with KITTI and nuScenes datasets confirms our\nproposed method's effectiveness in dealing with the data imbalance problem,\nproducing better detection accuracy compared to existing approaches.\n", "date": "2022-10-07", "categories": []} {"id": "2210.02493", "title": "Depth Is All You Need for Monocular 3D Detection", "abstract": " A key contributor to recent progress in 3D detection from single images is\nmonocular depth estimation. Existing methods focus on how to leverage depth\nexplicitly, by generating pseudo-pointclouds or providing attention cues for\nimage features. More recent works leverage depth prediction as a pretraining\ntask and fine-tune the depth representation while training it for 3D detection.\nHowever, the adaptation is insufficient and is limited in scale by manual\nlabels. In this work, we propose to further align depth representation with the\ntarget domain in unsupervised fashions. Our methods leverage commonly available\nLiDAR or RGB videos during training time to fine-tune the depth representation,\nwhich leads to improved 3D detectors. Especially when using RGB videos, we show\nthat our two-stage training by first generating pseudo-depth labels is critical\nbecause of the inconsistency in loss distribution between the two tasks. With\neither type of reference data, our multi-task learning approach improves over\nthe state of the art on both KITTI and NuScenes, while matching the test-time\ncomplexity of its single task sub-network.\n", "date": "2022-10-05", "categories": []} {"id": "2210.02443", "title": "Time Will Tell: New Outlooks and A Baseline for Temporal Multi-View 3D\n Object Detection", "abstract": " While recent camera-only 3D detection methods leverage multiple timesteps,\nthe limited history they use significantly hampers the extent to which temporal\nfusion can improve object perception. Observing that existing works' fusion of\nmulti-frame images are instances of temporal stereo matching, we find that\nperformance is hindered by the interplay between 1) the low granularity of\nmatching resolution and 2) the sub-optimal multi-view setup produced by limited\nhistory usage. Our theoretical and empirical analysis demonstrates that the\noptimal temporal difference between views varies significantly for different\npixels and depths, making it necessary to fuse many timesteps over long-term\nhistory. Building on our investigation, we propose to generate a cost volume\nfrom a long history of image observations, compensating for the coarse but\nefficient matching resolution with a more optimal multi-view matching setup.\nFurther, we augment the per-frame monocular depth predictions used for\nlong-term, coarse matching with short-term, fine-grained matching and find that\nlong and short term temporal fusion are highly complementary. While maintaining\nhigh efficiency, our framework sets new state-of-the-art on nuScenes, achieving\nfirst place on the test set and outperforming previous best art by 5.2% mAP and\n3.7% NDS on the validation set. Code will be released\n$\\href{https://github.com/Divadi/SOLOFusion}{here.}$\n", "date": "2022-10-05", "categories": []} {"id": "2210.01391", "title": "Bridged Transformer for Vision and Point Cloud 3D Object Detection", "abstract": " 3D object detection is a crucial research topic in computer vision, which\nusually uses 3D point clouds as input in conventional setups. Recently, there\nis a trend of leveraging multiple sources of input data, such as complementing\nthe 3D point cloud with 2D images that often have richer color and fewer\nnoises. However, due to the heterogeneous geometrics of the 2D and 3D\nrepresentations, it prevents us from applying off-the-shelf neural networks to\nachieve multimodal fusion. To that end, we propose Bridged Transformer (BrT),\nan end-to-end architecture for 3D object detection. BrT is simple and\neffective, which learns to identify 3D and 2D object bounding boxes from both\npoints and image patches. A key element of BrT lies in the utilization of\nobject queries for bridging 3D and 2D spaces, which unifies different sources\nof data representations in Transformer. We adopt a form of feature aggregation\nrealized by point-to-patch projections which further strengthen the\ncorrelations between images and points. Moreover, BrT works seamlessly for\nfusing the point cloud with multi-view images. We experimentally show that BrT\nsurpasses state-of-the-art methods on SUN RGB-D and ScanNetV2 datasets.\n", "date": "2022-10-04", "categories": []} {"id": "2207.10035", "title": "Fully Sparse 3D Object Detection", "abstract": " As the perception range of LiDAR increases, LiDAR-based 3D object detection\nbecomes a dominant task in the long-range perception task of autonomous\ndriving. The mainstream 3D object detectors usually build dense feature maps in\nthe network backbone and prediction head. However, the computational and\nspatial costs on the dense feature map are quadratic to the perception range,\nwhich makes them hardly scale up to the long-range setting. To enable efficient\nlong-range LiDAR-based object detection, we build a fully sparse 3D object\ndetector (FSD). The computational and spatial cost of FSD is roughly linear to\nthe number of points and independent of the perception range. FSD is built upon\nthe general sparse voxel encoder and a novel sparse instance recognition (SIR)\nmodule. SIR first groups the points into instances and then applies\ninstance-wise feature extraction and prediction. In this way, SIR resolves the\nissue of center feature missing, which hinders the design of the fully sparse\narchitecture for all center-based or anchor-based detectors. Moreover, SIR\navoids the time-consuming neighbor queries in previous point-based methods by\ngrouping points into instances. We conduct extensive experiments on the\nlarge-scale Waymo Open Dataset to reveal the working mechanism of FSD, and\nstate-of-the-art performance is reported. To demonstrate the superiority of FSD\nin long-range detection, we also conduct experiments on Argoverse 2 Dataset,\nwhich has a much larger perception range ($200m$) than Waymo Open Dataset\n($75m$). On such a large perception range, FSD achieves state-of-the-art\nperformance and is 2.4$\\times$ faster than the dense counterpart. Codes will be\nreleased at https://github.com/TuSimple/SST.\n", "date": "2022-07-20", "categories": []} {"id": "2210.00087", "title": "D-Align: Dual Query Co-attention Network for 3D Object Detection Based\n on Multi-frame Point Cloud Sequence", "abstract": " LiDAR sensors are widely used for 3D object detection in various mobile\nrobotics applications. LiDAR sensors continuously generate point cloud data in\nreal-time. Conventional 3D object detectors detect objects using a set of\npoints acquired over a fixed duration. However, recent studies have shown that\nthe performance of object detection can be further enhanced by utilizing\nspatio-temporal information obtained from point cloud sequences. In this paper,\nwe propose a new 3D object detector, named D-Align, which can effectively\nproduce strong bird's-eye-view (BEV) features by aligning and aggregating the\nfeatures obtained from a sequence of point sets. The proposed method includes a\nnovel dual-query co-attention network that uses two types of queries, including\ntarget query set (T-QS) and support query set (S-QS), to update the features of\ntarget and support frames, respectively. D-Align aligns S-QS to T-QS based on\nthe temporal context features extracted from the adjacent feature maps and then\naggregates S-QS with T-QS using a gated attention mechanism. The dual queries\nare updated through multiple attention layers to progressively enhance the\ntarget frame features used to produce the detection results. Our experiments on\nthe nuScenes dataset show that the proposed D-Align method greatly improved the\nperformance of a single frame-based baseline method and significantly\noutperformed the latest 3D object detectors.\n", "date": "2022-09-30", "categories": []} {"id": "2210.00519", "title": "Exploiting More Information in Sparse Point Cloud for 3D Single Object\n Tracking", "abstract": " 3D single object tracking is a key task in 3D computer vision. However, the\nsparsity of point clouds makes it difficult to compute the similarity and\nlocate the object, posing big challenges to the 3D tracker. Previous works\ntried to solve the problem and improved the tracking performance in some common\nscenarios, but they usually failed in some extreme sparse scenarios, such as\nfor tracking objects at long distances or partially occluded. To address the\nabove problems, in this letter, we propose a sparse-to-dense and\ntransformer-based framework for 3D single object tracking. First, we transform\nthe 3D sparse points into 3D pillars and then compress them into 2D BEV\nfeatures to have a dense representation. Then, we propose an attention-based\nencoder to achieve global similarity computation between template and search\nbranches, which could alleviate the influence of sparsity. Meanwhile, the\nencoder applies the attention on multi-scale features to compensate for the\nlack of information caused by the sparsity of point cloud and the single scale\nof features. Finally, we use set-prediction to track the object through a\ntwo-stage decoder which also utilizes attention. Extensive experiments show\nthat our method achieves very promising results on the KITTI and NuScenes\ndatasets.\n", "date": "2022-10-02", "categories": []} {"id": "2206.07959", "title": "Simple-BEV: What Really Matters for Multi-Sensor BEV Perception?", "abstract": " Building 3D perception systems for autonomous vehicles that do not rely on\nhigh-density LiDAR is a critical research problem because of the expense of\nLiDAR systems compared to cameras and other sensors. Recent research has\ndeveloped a variety of camera-only methods, where features are differentiably\n\"lifted\" from the multi-camera images onto the 2D ground plane, yielding a\n\"bird's eye view\" (BEV) feature representation of the 3D space around the\nvehicle. This line of work has produced a variety of novel \"lifting\" methods,\nbut we observe that other details in the training setups have shifted at the\nsame time, making it unclear what really matters in top-performing methods. We\nalso observe that using cameras alone is not a real-world constraint,\nconsidering that additional sensors like radar have been integrated into real\nvehicles for years already. In this paper, we first of all attempt to elucidate\nthe high-impact factors in the design and training protocol of BEV perception\nmodels. We find that batch size and input resolution greatly affect\nperformance, while lifting strategies have a more modest effect -- even a\nsimple parameter-free lifter works well. Second, we demonstrate that radar data\ncan provide a substantial boost to performance, helping to close the gap\nbetween camera-only and LiDAR-enabled systems. We analyze the radar usage\ndetails that lead to good performance, and invite the community to re-consider\nthis commonly-neglected part of the sensor platform.\n", "date": "2022-06-16", "categories": []} {"id": "2209.15252", "title": "PointPillars Backbone Type Selection For Fast and Accurate LiDAR Object\n Detection", "abstract": " 3D object detection from LiDAR sensor data is an important topic in the\ncontext of autonomous cars and drones. In this paper, we present the results of\nexperiments on the impact of backbone selection of a deep convolutional neural\nnetwork on detection accuracy and computation speed. We chose the PointPillars\nnetwork, which is characterised by a simple architecture, high speed, and\nmodularity that allows for easy expansion. During the experiments, we paid\nparticular attention to the change in detection efficiency (measured by the mAP\nmetric) and the total number of multiply-addition operations needed to process\none point cloud. We tested 10 different convolutional neural network\narchitectures that are widely used in image-based detection problems. For a\nbackbone like MobilenetV1, we obtained an almost 4x speedup at the cost of a\n1.13% decrease in mAP. On the other hand, for CSPDarknet we got an acceleration\nof more than 1.5x at an increase in mAP of 0.33%. We have thus demonstrated\nthat it is possible to significantly speed up a 3D object detector in LiDAR\npoint clouds with a small decrease in detection efficiency. This result can be\nused when PointPillars or similar algorithms are implemented in embedded\nsystems, including SoC FPGAs. The code is available at\nhttps://github.com/vision-agh/pointpillars\\_backbone.\n", "date": "2022-09-30", "categories": []} {"id": "2209.15258", "title": "Transformers for Object Detection in Large Point Clouds", "abstract": " We present TransLPC, a novel detection model for large point clouds that is\nbased on a transformer architecture. While object detection with transformers\nhas been an active field of research, it has proved difficult to apply such\nmodels to point clouds that span a large area, e.g. those that are common in\nautonomous driving, with lidar or radar data. TransLPC is able to remedy these\nissues: The structure of the transformer model is modified to allow for larger\ninput sequence lengths, which are sufficient for large point clouds. Besides\nthis, we propose a novel query refinement technique to improve detection\naccuracy, while retaining a memory-friendly number of transformer decoder\nqueries. The queries are repositioned between layers, moving them closer to the\nbounding box they are estimating, in an efficient manner. This simple technique\nhas a significant effect on detection accuracy, which is evaluated on the\nchallenging nuScenes dataset on real-world lidar data. Besides this, the\nproposed method is compatible with existing transformer-based solutions that\nrequire object detection, e.g. for joint multi-object tracking and detection,\nand enables them to be used in conjunction with large point clouds.\n", "date": "2022-09-30", "categories": []} {"id": "2209.14435", "title": "Out-of-Distribution Detection for LiDAR-based 3D Object Detection", "abstract": " 3D object detection is an essential part of automated driving, and deep\nneural networks (DNNs) have achieved state-of-the-art performance for this\ntask. However, deep models are notorious for assigning high confidence scores\nto out-of-distribution (OOD) inputs, that is, inputs that are not drawn from\nthe training distribution. Detecting OOD inputs is challenging and essential\nfor the safe deployment of models. OOD detection has been studied extensively\nfor the classification task, but it has not received enough attention for the\nobject detection task, specifically LiDAR-based 3D object detection. In this\npaper, we focus on the detection of OOD inputs for LiDAR-based 3D object\ndetection. We formulate what OOD inputs mean for object detection and propose\nto adapt several OOD detection methods for object detection. We accomplish this\nby our proposed feature extraction method. To evaluate OOD detection methods,\nwe develop a simple but effective technique of generating OOD objects for a\ngiven object detection model. Our evaluation based on the KITTI dataset shows\nthat different OOD detection methods have biases toward detecting specific OOD\nobjects. It emphasizes the importance of combined OOD detection methods and\nmore research in this direction.\n", "date": "2022-09-28", "categories": []} {"id": "2209.14201", "title": "Spatial Pruned Sparse Convolution for Efficient 3D Object Detection", "abstract": " 3D scenes are dominated by a large number of background points, which is\nredundant for the detection task that mainly needs to focus on foreground\nobjects. In this paper, we analyze major components of existing sparse 3D CNNs\nand find that 3D CNNs ignore the redundancy of data and further amplify it in\nthe down-sampling process, which brings a huge amount of extra and unnecessary\ncomputational overhead. Inspired by this, we propose a new convolution operator\nnamed spatial pruned sparse convolution (SPS-Conv), which includes two\nvariants, spatial pruned submanifold sparse convolution (SPSS-Conv) and spatial\npruned regular sparse convolution (SPRS-Conv), both of which are based on the\nidea of dynamically determining crucial areas for redundancy reduction. We\nvalidate that the magnitude can serve as important cues to determine crucial\nareas which get rid of the extra computations of learning-based methods. The\nproposed modules can easily be incorporated into existing sparse 3D CNNs\nwithout extra architectural modifications. Extensive experiments on the KITTI,\nWaymo and nuScenes datasets demonstrate that our method can achieve more than\n50% reduction in GFLOPs without compromising the performance.\n", "date": "2022-09-28", "categories": []} {"id": "2207.10047", "title": "Densely Constrained Depth Estimator for Monocular 3D Object Detection", "abstract": " Estimating accurate 3D locations of objects from monocular images is a\nchallenging problem because of lacking depth. Previous work shows that\nutilizing the object's keypoint projection constraints to estimate multiple\ndepth candidates boosts the detection performance. However, the existing\nmethods can only utilize vertical edges as projection constraints for depth\nestimation. So these methods only use a small number of projection constraints\nand produce insufficient depth candidates, leading to inaccurate depth\nestimation. In this paper, we propose a method that utilizes dense projection\nconstraints from edges of any direction. In this way, we employ much more\nprojection constraints and produce considerable depth candidates. Besides, we\npresent a graph matching weighting module to merge the depth candidates. The\nproposed method DCD (Densely Constrained Detector) achieves state-of-the-art\nperformance on the KITTI and WOD benchmarks. Code is released at\nhttps://github.com/BraveGroup/DCD.\n", "date": "2022-07-20", "categories": []} {"id": "2209.12729", "title": "DeepFusion: A Robust and Modular 3D Object Detector for Lidars, Cameras\n and Radars", "abstract": " We propose DeepFusion, a modular multi-modal architecture to fuse lidars,\ncameras and radars in different combinations for 3D object detection.\nSpecialized feature extractors take advantage of each modality and can be\nexchanged easily, making the approach simple and flexible. Extracted features\nare transformed into bird's-eye-view as a common representation for fusion.\nSpatial and semantic alignment is performed prior to fusing modalities in the\nfeature space. Finally, a detection head exploits rich multi-modal features for\nimproved 3D detection performance. Experimental results for lidar-camera,\nlidar-camera-radar and camera-radar fusion show the flexibility and\neffectiveness of our fusion approach. In the process, we study the largely\nunexplored task of faraway car detection up to 225 meters, showing the benefits\nof our lidar-camera fusion. Furthermore, we investigate the required density of\nlidar points for 3D object detection and illustrate implications at the example\nof robustness against adverse weather conditions. Moreover, ablation studies on\nour camera-radar fusion highlight the importance of accurate depth estimation.\n", "date": "2022-09-26", "categories": []} {"id": "2209.12254", "title": "From One to Many: Dynamic Cross Attention Networks for LiDAR and Camera\n Fusion", "abstract": " LiDAR and cameras are two complementary sensors for 3D perception in\nautonomous driving. LiDAR point clouds have accurate spatial and geometry\ninformation, while RGB images provide textural and color data for context\nreasoning. To exploit LiDAR and cameras jointly, existing fusion methods tend\nto align each 3D point to only one projected image pixel based on calibration,\nnamely one-to-one mapping. However, the performance of these approaches highly\nrelies on the calibration quality, which is sensitive to the temporal and\nspatial synchronization of sensors. Therefore, we propose a Dynamic Cross\nAttention (DCA) module with a novel one-to-many cross-modality mapping that\nlearns multiple offsets from the initial projection towards the neighborhood\nand thus develops tolerance to calibration error. Moreover, a \\textit{dynamic\nquery enhancement} is proposed to perceive the model-independent calibration,\nwhich further strengthens DCA's tolerance to the initial misalignment. The\nwhole fusion architecture named Dynamic Cross Attention Network (DCAN) exploits\nmulti-level image features and adapts to multiple representations of point\nclouds, which allows DCA to serve as a plug-in fusion module. Extensive\nexperiments on nuScenes and KITTI prove DCA's effectiveness. The proposed DCAN\noutperforms state-of-the-art methods on the nuScenes detection challenge.\n", "date": "2022-09-25", "categories": []} {"id": "2209.12419", "title": "Feature-based model selection for object detection from point cloud data", "abstract": " Smart monitoring using three-dimensional (3D) image sensors has been\nattracting attention in the context of smart cities. In smart monitoring,\nobject detection from point cloud data acquired by 3D image sensors is\nimplemented for detecting moving objects such as vehicles and pedestrians to\nensure safety on the road. However, the features of point cloud data are\ndiversified due to the characteristics of light detection and ranging (LIDAR)\nunits used as 3D image sensors or the install position of the 3D image sensors.\nAlthough a variety of deep learning (DL) models for object detection from point\ncloud data have been studied to date, no research has considered how to use\nmultiple DL models in accordance with the features of the point cloud data. In\nthis work, we propose a feature-based model selection framework that creates\nvarious DL models by using multiple DL methods and by utilizing training data\nwith pseudo incompleteness generated by two artificial techniques: sampling and\nnoise adding. It selects the most suitable DL model for the object detection\ntask in accordance with the features of the point cloud data acquired in the\nreal environment. To demonstrate the effectiveness of the proposed framework,\nwe compare the performance of multiple DL models using benchmark datasets\ncreated from the KITTI dataset and present example results of object detection\nobtained through a real outdoor experiment. Depending on the situation, the\ndetection accuracy varies up to 32% between DL models, which confirms the\nimportance of selecting an appropriate DL model according to the situation.\n", "date": "2022-09-26", "categories": []} {"id": "2209.10733", "title": "FusionRCNN: LiDAR-Camera Fusion for Two-stage 3D Object Detection", "abstract": " 3D object detection with multi-sensors is essential for an accurate and\nreliable perception system of autonomous driving and robotics. Existing 3D\ndetectors significantly improve the accuracy by adopting a two-stage paradigm\nwhich merely relies on LiDAR point clouds for 3D proposal refinement. Though\nimpressive, the sparsity of point clouds, especially for the points far away,\nmaking it difficult for the LiDAR-only refinement module to accurately\nrecognize and locate objects.To address this problem, we propose a novel\nmulti-modality two-stage approach named FusionRCNN, which effectively and\nefficiently fuses point clouds and camera images in the Regions of\nInterest(RoI). FusionRCNN adaptively integrates both sparse geometry\ninformation from LiDAR and dense texture information from camera in a unified\nattention mechanism. Specifically, it first utilizes RoIPooling to obtain an\nimage set with a unified size and gets the point set by sampling raw points\nwithin proposals in the RoI extraction step; then leverages an intra-modality\nself-attention to enhance the domain-specific features, following by a\nwell-designed cross-attention to fuse the information from two\nmodalities.FusionRCNN is fundamentally plug-and-play and supports different\none-stage methods with almost no architectural changes. Extensive experiments\non KITTI and Waymo benchmarks demonstrate that our method significantly boosts\nthe performances of popular detectors.Remarkably, FusionRCNN significantly\nimproves the strong SECOND baseline by 6.14% mAP on Waymo, and outperforms\ncompeting two-stage approaches. Code will be released soon at\nhttps://github.com/xxlbigbrother/Fusion-RCNN.\n", "date": "2022-09-22", "categories": []} {"id": "2209.10248", "title": "BEVStereo: Enhancing Depth Estimation in Multi-view 3D Object Detection\n with Dynamic Temporal Stereo", "abstract": " Bounded by the inherent ambiguity of depth perception, contemporary\ncamera-based 3D object detection methods fall into the performance bottleneck.\nIntuitively, leveraging temporal multi-view stereo (MVS) technology is the\nnatural knowledge for tackling this ambiguity. However, traditional attempts of\nMVS are flawed in two aspects when applying to 3D object detection scenes: 1)\nThe affinity measurement among all views suffers expensive computation cost; 2)\nIt is difficult to deal with outdoor scenarios where objects are often mobile.\nTo this end, we introduce an effective temporal stereo method to dynamically\nselect the scale of matching candidates, enable to significantly reduce\ncomputation overhead. Going one step further, we design an iterative algorithm\nto update more valuable candidates, making it adaptive to moving candidates. We\ninstantiate our proposed method to multi-view 3D detector, namely BEVStereo.\nBEVStereo achieves the new state-of-the-art performance (i.e., 52.5% mAP and\n61.0% NDS) on the camera-only track of nuScenes dataset. Meanwhile, extensive\nexperiments reflect our method can deal with complex outdoor scenarios better\nthan contemporary MVS approaches. Codes have been released at\nhttps://github.com/Megvii-BaseDetection/BEVStereo.\n", "date": "2022-09-21", "categories": []} {"id": "2209.10471", "title": "Sample, Crop, Track: Self-Supervised Mobile 3D Object Detection for\n Urban Driving LiDAR", "abstract": " Deep learning has led to great progress in the detection of mobile (i.e.\nmovement-capable) objects in urban driving scenes in recent years. Supervised\napproaches typically require the annotation of large training sets; there has\nthus been great interest in leveraging weakly, semi- or self-supervised methods\nto avoid this, with much success. Whilst weakly and semi-supervised methods\nrequire some annotation, self-supervised methods have used cues such as motion\nto relieve the need for annotation altogether. However, a complete absence of\nannotation typically degrades their performance, and ambiguities that arise\nduring motion grouping can inhibit their ability to find accurate object\nboundaries. In this paper, we propose a new self-supervised mobile object\ndetection approach called SCT. This uses both motion cues and expected object\nsizes to improve detection performance, and predicts a dense grid of 3D\noriented bounding boxes to improve object discovery. We significantly\noutperform the state-of-the-art self-supervised mobile object detection method\nTCR on the KITTI tracking benchmark, and achieve performance that is within 30%\nof the fully supervised PV-RCNN++ method for IoUs <= 0.5.\n", "date": "2022-09-21", "categories": []} {"id": "2209.09486", "title": "Self-supervised 3D Object Detection from Monocular Pseudo-LiDAR", "abstract": " There have been attempts to detect 3D objects by fusion of stereo camera\nimages and LiDAR sensor data or using LiDAR for pre-training and only monocular\nimages for testing, but there have been less attempts to use only monocular\nimage sequences due to low accuracy. In addition, when depth prediction using\nonly monocular images, only scale-inconsistent depth can be predicted, which is\nthe reason why researchers are reluctant to use monocular images alone.\nTherefore, we propose a method for predicting absolute depth and detecting 3D\nobjects using only monocular image sequences by enabling end-to-end learning of\ndetection networks and depth prediction networks. As a result, the proposed\nmethod surpasses other existing methods in performance on the KITTI 3D dataset.\nEven when monocular image and 3D LiDAR are used together during training in an\nattempt to improve performance, ours exhibit is the best performance compared\nto other methods using the same input. In addition, end-to-end learning not\nonly improves depth prediction performance, but also enables absolute depth\nprediction, because our network utilizes the fact that the size of a 3D object\nsuch as a car is determined by the approximate size.\n", "date": "2022-09-20", "categories": []} {"id": "2209.08844", "title": "A Dual-Cycled Cross-View Transformer Network for Unified Road Layout\n Estimation and 3D Object Detection in the Bird's-Eye-View", "abstract": " The bird's-eye-view (BEV) representation allows robust learning of multiple\ntasks for autonomous driving including road layout estimation and 3D object\ndetection. However, contemporary methods for unified road layout estimation and\n3D object detection rarely handle the class imbalance of the training dataset\nand multi-class learning to reduce the total number of networks required. To\novercome these limitations, we propose a unified model for road layout\nestimation and 3D object detection inspired by the transformer architecture and\nthe CycleGAN learning framework. The proposed model deals with the performance\ndegradation due to the class imbalance of the dataset utilizing the focal loss\nand the proposed dual cycle loss. Moreover, we set up extensive learning\nscenarios to study the effect of multi-class learning for road layout\nestimation in various situations. To verify the effectiveness of the proposed\nmodel and the learning scheme, we conduct a thorough ablation study and a\ncomparative study. The experiment results attest the effectiveness of our\nmodel; we achieve state-of-the-art performance in both the road layout\nestimation and 3D object detection tasks.\n", "date": "2022-09-19", "categories": []} {"id": "2209.07419", "title": "FFPA-Net: Efficient Feature Fusion with Projection Awareness for 3D\n Object Detection", "abstract": " Promising complementarity exists between the texture features of color images\nand the geometric information of LiDAR point clouds. However, there still\npresent many challenges for efficient and robust feature fusion in the field of\n3D object detection. In this paper, first, unstructured 3D point clouds are\nfilled in the 2D plane and 3D point cloud features are extracted faster using\nprojection-aware convolution layers. Further, the corresponding indexes between\ndifferent sensor signals are established in advance in the data preprocessing,\nwhich enables faster cross-modal feature fusion. To address LiDAR points and\nimage pixels misalignment problems, two new plug-and-play fusion modules,\nLiCamFuse and BiLiCamFuse, are proposed. In LiCamFuse, soft query weights with\nperceiving the Euclidean distance of bimodal features are proposed. In\nBiLiCamFuse, the fusion module with dual attention is proposed to deeply\ncorrelate the geometric and textural features of the scene. The quantitative\nresults on the KITTI dataset demonstrate that the proposed method achieves\nbetter feature-level fusion. In addition, the proposed network shows a shorter\nrunning time compared to existing methods.\n", "date": "2022-09-15", "categories": []} {"id": "2209.06407", "title": "Viewer-Centred Surface Completion for Unsupervised Domain Adaptation in\n 3D Object Detection", "abstract": " Every autonomous driving dataset has a different configuration of sensors,\noriginating from distinct geographic regions and covering various scenarios. As\na result, 3D detectors tend to overfit the datasets they are trained on. This\ncauses a drastic decrease in accuracy when the detectors are trained on one\ndataset and tested on another. We observe that lidar scan pattern differences\nform a large component of this reduction in performance. We address this in our\napproach, SEE-VCN, by designing a novel viewer-centred surface completion\nnetwork (VCN) to complete the surfaces of objects of interest within an\nunsupervised domain adaptation framework, SEE. With SEE-VCN, we obtain a\nunified representation of objects across datasets, allowing the network to\nfocus on learning geometry, rather than overfitting on scan patterns. By\nadopting a domain-invariant representation, SEE-VCN can be classed as a\nmulti-target domain adaptation approach where no annotations or re-training is\nrequired to obtain 3D detections for new scan patterns. Through extensive\nexperiments, we show that our approach outperforms previous domain adaptation\nmethods in multiple domain adaptation settings. Our code and data are available\nat https://github.com/darrenjkt/SEE-VCN.\n", "date": "2022-09-14", "categories": []} {"id": "2209.06641", "title": "CMR3D: Contextualized Multi-Stage Refinement for 3D Object Detection", "abstract": " Existing deep learning-based 3D object detectors typically rely on the\nappearance of individual objects and do not explicitly pay attention to the\nrich contextual information of the scene. In this work, we propose\nContextualized Multi-Stage Refinement for 3D Object Detection (CMR3D)\nframework, which takes a 3D scene as input and strives to explicitly integrate\nuseful contextual information of the scene at multiple levels to predict a set\nof object bounding-boxes along with their corresponding semantic labels. To\nthis end, we propose to utilize a context enhancement network that captures the\ncontextual information at different levels of granularity followed by a\nmulti-stage refinement module to progressively refine the box positions and\nclass predictions. Extensive experiments on the large-scale ScanNetV2 benchmark\nreveal the benefits of our proposed method, leading to an absolute improvement\nof 2.0% over the baseline. In addition to 3D object detection, we investigate\nthe effectiveness of our CMR3D framework for the problem of 3D object counting.\nOur source code will be publicly released.\n", "date": "2022-09-13", "categories": []} {"id": "2209.05588", "title": "CenterFormer: Center-based Transformer for 3D Object Detection", "abstract": " Query-based transformer has shown great potential in constructing long-range\nattention in many image-domain tasks, but has rarely been considered in\nLiDAR-based 3D object detection due to the overwhelming size of the point cloud\ndata. In this paper, we propose CenterFormer, a center-based transformer\nnetwork for 3D object detection. CenterFormer first uses a center heatmap to\nselect center candidates on top of a standard voxel-based point cloud encoder.\nIt then uses the feature of the center candidate as the query embedding in the\ntransformer. To further aggregate features from multiple frames, we design an\napproach to fuse features through cross-attention. Lastly, regression heads are\nadded to predict the bounding box on the output center feature representation.\nOur design reduces the convergence difficulty and computational complexity of\nthe transformer structure. The results show significant improvements over the\nstrong baseline of anchor-free object detection networks. CenterFormer achieves\nstate-of-the-art performance for a single model on the Waymo Open Dataset, with\n73.7% mAPH on the validation set and 75.6% mAPH on the test set, significantly\noutperforming all previously published CNN and transformer-based methods. Our\ncode is publicly available at https://github.com/TuSimple/centerformer\n", "date": "2022-09-12", "categories": []} {"id": "2209.04966", "title": "Multi-modal Streaming 3D Object Detection", "abstract": " Modern autonomous vehicles rely heavily on mechanical LiDARs for perception.\nCurrent perception methods generally require 360{\\deg} point clouds, collected\nsequentially as the LiDAR scans the azimuth and acquires consecutive\nwedge-shaped slices. The acquisition latency of a full scan (~ 100ms) may lead\nto outdated perception which is detrimental to safe operation. Recent streaming\nperception works proposed directly processing LiDAR slices and compensating for\nthe narrow field of view (FOV) of a slice by reusing features from preceding\nslices. These works, however, are all based on a single modality and require\npast information which may be outdated. Meanwhile, images from high-frequency\ncameras can support streaming models as they provide a larger FoV compared to a\nLiDAR slice. However, this difference in FoV complicates sensor fusion. To\naddress this research gap, we propose an innovative camera-LiDAR streaming 3D\nobject detection framework that uses camera images instead of past LiDAR slices\nto provide an up-to-date, dense, and wide context for streaming perception. The\nproposed method outperforms prior streaming models on the challenging NuScenes\nbenchmark. It also outperforms powerful full-scan detectors while being much\nfaster. Our method is shown to be robust to missing camera images, narrow LiDAR\nslices, and small camera-LiDAR miscalibration.\n", "date": "2022-09-12", "categories": []} {"id": "2209.01745", "title": "SEFormer: Structure Embedding Transformer for 3D Object Detection", "abstract": " Effectively preserving and encoding structure features from objects in\nirregular and sparse LiDAR points is a key challenge to 3D object detection on\npoint cloud. Recently, Transformer has demonstrated promising performance on\nmany 2D and even 3D vision tasks. Compared with the fixed and rigid convolution\nkernels, the self-attention mechanism in Transformer can adaptively exclude the\nunrelated or noisy points and thus suitable for preserving the local spatial\nstructure in irregular LiDAR point cloud. However, Transformer only performs a\nsimple sum on the point features, based on the self-attention mechanism, and\nall the points share the same transformation for value. Such isotropic\noperation lacks the ability to capture the direction-distance-oriented local\nstructure which is important for 3D object detection. In this work, we propose\na Structure-Embedding transFormer (SEFormer), which can not only preserve local\nstructure as traditional Transformer but also have the ability to encode the\nlocal structure. Compared to the self-attention mechanism in traditional\nTransformer, SEFormer learns different feature transformations for value points\nbased on the relative directions and distances to the query point. Then we\npropose a SEFormer based network for high-performance 3D object detection.\nExtensive experiments show that the proposed architecture can achieve SOTA\nresults on Waymo Open Dataset, the largest 3D detection benchmark for\nautonomous driving. Specifically, SEFormer achieves 79.02% mAP, which is 1.2%\nhigher than existing works. We will release the codes.\n", "date": "2022-09-05", "categories": []} {"id": "2207.12654", "title": "ProposalContrast: Unsupervised Pre-training for LiDAR-based 3D Object\n Detection", "abstract": " Existing approaches for unsupervised point cloud pre-training are constrained\nto either scene-level or point/voxel-level instance discrimination. Scene-level\nmethods tend to lose local details that are crucial for recognizing the road\nobjects, while point/voxel-level methods inherently suffer from limited\nreceptive field that is incapable of perceiving large objects or context\nenvironments. Considering region-level representations are more suitable for 3D\nobject detection, we devise a new unsupervised point cloud pre-training\nframework, called ProposalContrast, that learns robust 3D representations by\ncontrasting region proposals. Specifically, with an exhaustive set of region\nproposals sampled from each point cloud, geometric point relations within each\nproposal are modeled for creating expressive proposal representations. To\nbetter accommodate 3D detection properties, ProposalContrast optimizes with\nboth inter-cluster and inter-proposal separation, i.e., sharpening the\ndiscriminativeness of proposal representations across semantic classes and\nobject instances. The generalizability and transferability of ProposalContrast\nare verified on various 3D detectors (i.e., PV-RCNN, CenterPoint, PointPillars\nand PointRCNN) and datasets (i.e., KITTI, Waymo and ONCE).\n", "date": "2022-07-26", "categories": []} {"id": "2208.14738", "title": "Scatter Points in Space: 3D Detection from Multi-view Monocular Images", "abstract": " 3D object detection from monocular image(s) is a challenging and\nlong-standing problem of computer vision. To combine information from different\nperspectives without troublesome 2D instance tracking, recent methods tend to\naggregate multiview feature by sampling regular 3D grid densely in space, which\nis inefficient. In this paper, we attempt to improve multi-view feature\naggregation by proposing a learnable keypoints sampling method, which scatters\npseudo surface points in 3D space, in order to keep data sparsity. The\nscattered points augmented by multi-view geometric constraints and visual\nfeatures are then employed to infer objects location and shape in the scene. To\nmake up the limitations of single frame and model multi-view geometry\nexplicitly, we further propose a surface filter module for noise suppression.\nExperimental results show that our method achieves significantly better\nperformance than previous works in terms of 3D detection (more than 0.1 AP\nimprovement on some categories of ScanNet). The code will be publicly\navailable.\n", "date": "2022-08-31", "categories": []} {"id": "2208.14796", "title": "3DLG-Detector: 3D Object Detection via Simultaneous Local-Global Feature\n Learning", "abstract": " Capturing both local and global features of irregular point clouds is\nessential to 3D object detection (3OD). However, mainstream 3D detectors, e.g.,\nVoteNet and its variants, either abandon considerable local features during\npooling operations or ignore many global features in the whole scene context.\nThis paper explores new modules to simultaneously learn local-global features\nof scene point clouds that serve 3OD positively. To this end, we propose an\neffective 3OD network via simultaneous local-global feature learning (dubbed\n3DLG-Detector). 3DLG-Detector has two key contributions. First, it develops a\nDynamic Points Interaction (DPI) module that preserves effective local features\nduring pooling. Besides, DPI is detachable and can be incorporated into\nexisting 3OD networks to boost their performance. Second, it develops a Global\nContext Aggregation module to aggregate multi-scale features from different\nlayers of the encoder to achieve scene context-awareness. Our method shows\nimprovements over thirteen competitors in terms of detection accuracy and\nrobustness on both the SUN RGB-D and ScanNet datasets. Source code will be\navailable upon publication.\n", "date": "2022-08-31", "categories": []} {"id": "2208.13414", "title": "PV-RCNN++: Semantical Point-Voxel Feature Interaction for 3D Object\n Detection", "abstract": " Large imbalance often exists between the foreground points (i.e., objects)\nand the background points in outdoor LiDAR point clouds. It hinders\ncutting-edge detectors from focusing on informative areas to produce accurate\n3D object detection results. This paper proposes a novel object detection\nnetwork by semantical point-voxel feature interaction, dubbed PV-RCNN++. Unlike\nmost of existing methods, PV-RCNN++ explores the semantic information to\nenhance the quality of object detection. First, a semantic segmentation module\nis proposed to retain more discriminative foreground keypoints. Such a module\nwill guide our PV-RCNN++ to integrate more object-related point-wise and\nvoxel-wise features in the pivotal areas. Then, to make points and voxels\ninteract efficiently, we utilize voxel query based on Manhattan distance to\nquickly sample voxel-wise features around keypoints. Such the voxel query will\nreduce the time complexity from O(N) to O(K), compared to the ball query.\nFurther, to avoid being stuck in learning only local features, an\nattention-based residual PointNet module is designed to expand the receptive\nfield to adaptively aggregate the neighboring voxel-wise features into\nkeypoints. Extensive experiments on the KITTI dataset show that PV-RCNN++\nachieves 81.60$\\%$, 40.18$\\%$, 68.21$\\%$ 3D mAP on Car, Pedestrian, and\nCyclist, achieving comparable or even better performance to the\nstate-of-the-arts.\n", "date": "2022-08-29", "categories": []} {"id": "2208.10056", "title": "Minkowski Tracker: A Sparse Spatio-Temporal R-CNN for Joint Object\n Detection and Tracking", "abstract": " Recent research in multi-task learning reveals the benefit of solving related\nproblems in a single neural network. 3D object detection and multi-object\ntracking (MOT) are two heavily intertwined problems predicting and associating\nan object instance location across time. However, most previous works in 3D MOT\ntreat the detector as a preceding separated pipeline, disjointly taking the\noutput of the detector as an input to the tracker. In this work, we present\nMinkowski Tracker, a sparse spatio-temporal R-CNN that jointly solves object\ndetection and tracking. Inspired by region-based CNN (R-CNN), we propose to\nsolve tracking as a second stage of the object detector R-CNN that predicts\nassignment probability to tracks. First, Minkowski Tracker takes 4D point\nclouds as input to generate a spatio-temporal Bird's-eye-view (BEV) feature map\nthrough a 4D sparse convolutional encoder network. Then, our proposed\nTrackAlign aggregates the track region-of-interest (ROI) features from the BEV\nfeatures. Finally, Minkowski Tracker updates the track and its confidence score\nbased on the detection-to-track match probability predicted from the ROI\nfeatures. We show in large-scale experiments that the overall performance gain\nof our method is due to four factors: 1. The temporal reasoning of the 4D\nencoder improves the detection performance 2. The multi-task learning of object\ndetection and MOT jointly enhances each other 3. The detection-to-track match\nscore learns implicit motion model to enhance track assignment 4. The\ndetection-to-track match score improves the quality of the track confidence\nscore. As a result, Minkowski Tracker achieved the state-of-the-art performance\non Nuscenes dataset tracking task without hand-designed motion models.\n", "date": "2022-08-22", "categories": []} {"id": "2208.11658", "title": "AGO-Net: Association-Guided 3D Point Cloud Object Detection Network", "abstract": " The human brain can effortlessly recognize and localize objects, whereas\ncurrent 3D object detection methods based on LiDAR point clouds still report\ninferior performance for detecting occluded and distant objects: the point\ncloud appearance varies greatly due to occlusion, and has inherent variance in\npoint densities along the distance to sensors. Therefore, designing feature\nrepresentations robust to such point clouds is critical. Inspired by human\nassociative recognition, we propose a novel 3D detection framework that\nassociates intact features for objects via domain adaptation. We bridge the gap\nbetween the perceptual domain, where features are derived from real scenes with\nsub-optimal representations, and the conceptual domain, where features are\nextracted from augmented scenes that consist of non-occlusion objects with rich\ndetailed information. A feasible method is investigated to construct conceptual\nscenes without external datasets. We further introduce an attention-based\nre-weighting module that adaptively strengthens the feature adaptation of more\ninformative regions. The network's feature enhancement ability is exploited\nwithout introducing extra cost during inference, which is plug-and-play in\nvarious 3D detection frameworks. We achieve new state-of-the-art performance on\nthe KITTI 3D detection benchmark in both accuracy and speed. Experiments on\nnuScenes and Waymo datasets also validate the versatility of our method.\n", "date": "2022-08-24", "categories": []} {"id": "2208.09870", "title": "Objects Can Move: 3D Change Detection by Geometric Transformation\n Constistency", "abstract": " AR/VR applications and robots need to know when the scene has changed. An\nexample is when objects are moved, added, or removed from the scene. We propose\na 3D object discovery method that is based only on scene changes. Our method\ndoes not need to encode any assumptions about what is an object, but rather\ndiscovers objects by exploiting their coherent move. Changes are initially\ndetected as differences in the depth maps and segmented as objects if they\nundergo rigid motions. A graph cut optimization propagates the changing labels\nto geometrically consistent regions. Experiments show that our method achieves\nstate-of-the-art performance on the 3RScan dataset against competitive\nbaselines. The source code of our method can be found at\nhttps://github.com/katadam/ObjectsCanMove.\n", "date": "2022-08-21", "categories": []} {"id": "2208.10035", "title": "A Simple Baseline for Multi-Camera 3D Object Detection", "abstract": " 3D object detection with surrounding cameras has been a promising direction\nfor autonomous driving. In this paper, we present SimMOD, a Simple baseline for\nMulti-camera Object Detection, to solve the problem. To incorporate multi-view\ninformation as well as build upon previous efforts on monocular 3D object\ndetection, the framework is built on sample-wise object proposals and designed\nto work in a two-stage manner. First, we extract multi-scale features and\ngenerate the perspective object proposals on each monocular image. Second, the\nmulti-view proposals are aggregated and then iteratively refined with\nmulti-view and multi-scale visual features in the DETR3D-style. The refined\nproposals are end-to-end decoded into the detection results. To further boost\nthe performance, we incorporate the auxiliary branches alongside the proposal\ngeneration to enhance the feature learning. Also, we design the methods of\ntarget filtering and teacher forcing to promote the consistency of two-stage\ntraining. We conduct extensive experiments on the 3D object detection benchmark\nof nuScenes to demonstrate the effectiveness of SimMOD and achieve new\nstate-of-the-art performance. Code will be available at\nhttps://github.com/zhangyp15/SimMOD.\n", "date": "2022-08-22", "categories": []} {"id": "2208.10145", "title": "STS: Surround-view Temporal Stereo for Multi-view 3D Detection", "abstract": " Learning accurate depth is essential to multi-view 3D object detection.\nRecent approaches mainly learn depth from monocular images, which confront\ninherent difficulties due to the ill-posed nature of monocular depth learning.\nInstead of using a sole monocular depth method, in this work, we propose a\nnovel Surround-view Temporal Stereo (STS) technique that leverages the geometry\ncorrespondence between frames across time to facilitate accurate depth\nlearning. Specifically, we regard the field of views from all cameras around\nthe ego vehicle as a unified view, namely surroundview, and conduct temporal\nstereo matching on it. The resulting geometrical correspondence between\ndifferent frames from STS is utilized and combined with the monocular depth to\nyield final depth prediction. Comprehensive experiments on nuScenes show that\nSTS greatly boosts 3D detection ability, notably for medium and long distance\nobjects. On BEVDepth with ResNet-50 backbone, STS improves mAP and NDS by 2.6%\nand 1.4%, respectively. Consistent improvements are observed when using a\nlarger backbone and a larger image resolution, demonstrating its effectiveness\n", "date": "2022-08-22", "categories": []} {"id": "2208.09394", "title": "PersDet: Monocular 3D Detection in Perspective Bird's-Eye-View", "abstract": " Currently, detecting 3D objects in Bird's-Eye-View (BEV) is superior to other\n3D detectors for autonomous driving and robotics. However, transforming image\nfeatures into BEV necessitates special operators to conduct feature sampling.\nThese operators are not supported on many edge devices, bringing extra\nobstacles when deploying detectors. To address this problem, we revisit the\ngeneration of BEV representation and propose detecting objects in perspective\nBEV -- a new BEV representation that does not require feature sampling. We\ndemonstrate that perspective BEV features can likewise enjoy the benefits of\nthe BEV paradigm. Moreover, the perspective BEV improves detection performance\nby addressing issues caused by feature sampling. We propose PersDet for\nhigh-performance object detection in perspective BEV space based on this\ndiscovery. While implementing a simple and memory-efficient structure, PersDet\noutperforms existing state-of-the-art monocular methods on the nuScenes\nbenchmark, reaching 34.6% mAP and 40.8% NDS when using ResNet-50 as the\nbackbone.\n", "date": "2022-08-19", "categories": []} {"id": "2208.08621", "title": "Ret3D: Rethinking Object Relations for Efficient 3D Object Detection in\n Driving Scenes", "abstract": " Current efficient LiDAR-based detection frameworks are lacking in exploiting\nobject relations, which naturally present in both spatial and temporal manners.\nTo this end, we introduce a simple, efficient, and effective two-stage\ndetector, termed as Ret3D. At the core of Ret3D is the utilization of novel\nintra-frame and inter-frame relation modules to capture the spatial and\ntemporal relations accordingly. More Specifically, intra-frame relation module\n(IntraRM) encapsulates the intra-frame objects into a sparse graph and thus\nallows us to refine the object features through efficient message passing. On\nthe other hand, inter-frame relation module (InterRM) densely connects each\nobject in its corresponding tracked sequences dynamically, and leverages such\ntemporal information to further enhance its representations efficiently through\na lightweight transformer network. We instantiate our novel designs of IntraRM\nand InterRM with general center-based or anchor-based detectors and evaluate\nthem on Waymo Open Dataset (WOD). With negligible extra overhead, Ret3D\nachieves the state-of-the-art performance, being 5.5% and 3.2% higher than the\nrecent competitor in terms of the LEVEL 1 and LEVEL 2 mAPH metrics on vehicle\ndetection, respectively.\n", "date": "2022-08-18", "categories": []} {"id": "2208.07137", "title": "An Empirical Study of Pseudo-Labeling for Image-based 3D Object\n Detection", "abstract": " Image-based 3D detection is an indispensable component of the perception\nsystem for autonomous driving. However, it still suffers from the unsatisfying\nperformance, one of the main reasons for which is the limited training data.\nUnfortunately, annotating the objects in the 3D space is extremely\ntime/resource-consuming, which makes it hard to extend the training set\narbitrarily. In this work, we focus on the semi-supervised manner and explore\nthe feasibility of a cheaper alternative, i.e. pseudo-labeling, to leverage the\nunlabeled data. For this purpose, we conduct extensive experiments to\ninvestigate whether the pseudo-labels can provide effective supervision for the\nbaseline models under varying settings. The experimental results not only\ndemonstrate the effectiveness of the pseudo-labeling mechanism for image-based\n3D detection (e.g. under monocular setting, we achieve 20.23 AP for moderate\nlevel on the KITTI-3D testing set without bells and whistles, improving the\nbaseline model by 6.03 AP), but also show several interesting and noteworthy\nfindings (e.g. the models trained with pseudo-labels perform better than that\ntrained with ground-truth annotations based on the same training data). We hope\nthis work can provide insights for the image-based 3D detection community under\na semi-supervised setting. The codes, pseudo-labels, and pre-trained models\nwill be publicly available.\n", "date": "2022-08-15", "categories": []} {"id": "2208.03624", "title": "Graph R-CNN: Towards Accurate 3D Object Detection with\n Semantic-Decorated Local Graph", "abstract": " Two-stage detectors have gained much popularity in 3D object detection. Most\ntwo-stage 3D detectors utilize grid points, voxel grids, or sampled keypoints\nfor RoI feature extraction in the second stage. Such methods, however, are\ninefficient in handling unevenly distributed and sparse outdoor points. This\npaper solves this problem in three aspects. 1) Dynamic Point Aggregation. We\npropose the patch search to quickly search points in a local region for each 3D\nproposal. The dynamic farthest voxel sampling is then applied to evenly sample\nthe points. Especially, the voxel size varies along the distance to accommodate\nthe uneven distribution of points. 2) RoI-graph Pooling. We build local graphs\non the sampled points to better model contextual information and mine point\nrelations through iterative message passing. 3) Visual Features Augmentation.\nWe introduce a simple yet effective fusion strategy to compensate for sparse\nLiDAR points with limited semantic cues. Based on these modules, we construct\nour Graph R-CNN as the second stage, which can be applied to existing one-stage\ndetectors to consistently improve the detection performance. Extensive\nexperiments show that Graph R-CNN outperforms the state-of-the-art 3D detection\nmodels by a large margin on both the KITTI and Waymo Open Dataset. And we rank\nfirst place on the KITTI BEV car detection leaderboard. Code will be available\nat \\url{https://github.com/Nightmare-n/GraphRCNN}.\n", "date": "2022-08-07", "categories": []} {"id": "2208.03141", "title": "TransPillars: Coarse-to-Fine Aggregation for Multi-Frame 3D Object\n Detection", "abstract": " 3D object detection using point clouds has attracted increasing attention due\nto its wide applications in autonomous driving and robotics. However, most\nexisting studies focus on single point cloud frames without harnessing the\ntemporal information in point cloud sequences. In this paper, we design\nTransPillars, a novel transformer-based feature aggregation technique that\nexploits temporal features of consecutive point cloud frames for multi-frame 3D\nobject detection. TransPillars aggregates spatial-temporal point cloud features\nfrom two perspectives. First, it fuses voxel-level features directly from\nmulti-frame feature maps instead of pooled instance features to preserve\ninstance details with contextual information that are essential to accurate\nobject localization. Second, it introduces a hierarchical coarse-to-fine\nstrategy to fuse multi-scale features progressively to effectively capture the\nmotion of moving objects and guide the aggregation of fine features. Besides, a\nvariant of deformable transformer is introduced to improve the effectiveness of\ncross-frame feature matching. Extensive experiments show that our proposed\nTransPillars achieves state-of-art performance as compared to existing\nmulti-frame detection approaches. Code will be released.\n", "date": "2022-08-04", "categories": []} {"id": "2206.00214", "title": "LiDAR-MIMO: Efficient Uncertainty Estimation for LiDAR-based 3D Object\n Detection", "abstract": " The estimation of uncertainty in robotic vision, such as 3D object detection,\nis an essential component in developing safe autonomous systems aware of their\nown performance. However, the deployment of current uncertainty estimation\nmethods in 3D object detection remains challenging due to timing and\ncomputational constraints. To tackle this issue, we propose LiDAR-MIMO, an\nadaptation of the multi-input multi-output (MIMO) uncertainty estimation method\nto the LiDAR-based 3D object detection task. Our method modifies the original\nMIMO by performing multi-input at the feature level to ensure the detection,\nuncertainty estimation, and runtime performance benefits are retained despite\nthe limited capacity of the underlying detector and the large computational\ncosts of point cloud processing. We compare LiDAR-MIMO with MC dropout and\nensembles as baselines and show comparable uncertainty estimation results with\nonly a small number of output heads. Further, LiDAR-MIMO can be configured to\nbe twice as fast as MC dropout and ensembles, while achieving higher mAP than\nMC dropout and approaching that of ensembles.\n", "date": "2022-06-01", "categories": []} {"id": "2207.12655", "title": "Semi-supervised 3D Object Detection with Proficient Teachers", "abstract": " Dominated point cloud-based 3D object detectors in autonomous driving\nscenarios rely heavily on the huge amount of accurately labeled samples,\nhowever, 3D annotation in the point cloud is extremely tedious, expensive and\ntime-consuming. To reduce the dependence on large supervision, semi-supervised\nlearning (SSL) based approaches have been proposed. The Pseudo-Labeling\nmethodology is commonly used for SSL frameworks, however, the low-quality\npredictions from the teacher model have seriously limited its performance. In\nthis work, we propose a new Pseudo-Labeling framework for semi-supervised 3D\nobject detection, by enhancing the teacher model to a proficient one with\nseveral necessary designs. First, to improve the recall of pseudo labels, a\nSpatialtemporal Ensemble (STE) module is proposed to generate sufficient seed\nboxes. Second, to improve the precision of recalled boxes, a Clusteringbased\nBox Voting (CBV) module is designed to get aggregated votes from the clustered\nseed boxes. This also eliminates the necessity of sophisticated thresholds to\nselect pseudo labels. Furthermore, to reduce the negative influence of wrongly\npseudo-labeled samples during the training, a soft supervision signal is\nproposed by considering Box-wise Contrastive Learning (BCL). The effectiveness\nof our model is verified on both ONCE and Waymo datasets. For example, on ONCE,\nour approach significantly improves the baseline by 9.51 mAP. Moreover, with\nhalf annotations, our model outperforms the oracle model with full annotations\non Waymo.\n", "date": "2022-07-26", "categories": []} {"id": "2207.12716", "title": "MV-FCOS3D++: Multi-View Camera-Only 4D Object Detection with Pretrained\n Monocular Backbones", "abstract": " In this technical report, we present our solution, dubbed MV-FCOS3D++, for\nthe Camera-Only 3D Detection track in Waymo Open Dataset Challenge 2022. For\nmulti-view camera-only 3D detection, methods based on bird-eye-view or 3D\ngeometric representations can leverage the stereo cues from overlapped regions\nbetween adjacent views and directly perform 3D detection without hand-crafted\npost-processing. However, it lacks direct semantic supervision for 2D\nbackbones, which can be complemented by pretraining simple monocular-based\ndetectors. Our solution is a multi-view framework for 4D detection following\nthis paradigm. It is built upon a simple monocular detector FCOS3D++,\npretrained only with object annotations of Waymo, and converts multi-view\nfeatures to a 3D grid space to detect 3D objects thereon. A dual-path neck for\nsingle-frame understanding and temporal stereo matching is devised to\nincorporate multi-frame information. Our method finally achieves 49.75% mAPL\nwith a single model and wins 2nd place in the WOD challenge, without any\nLiDAR-based depth supervision during training. The code will be released at\nhttps://github.com/Tai-Wang/Depth-from-Motion.\n", "date": "2022-07-26", "categories": []} {"id": "2207.11479", "title": "3D Labeling Tool", "abstract": " Training and testing supervised object detection models require a large\ncollection of images with ground truth labels. Labels define object classes in\nthe image, as well as their locations, shape, and possibly other information\nsuch as pose. The labeling process has proven extremely time consuming, even\nwith the presence of manpower. We introduce a novel labeling tool for 2D images\nas well as 3D triangular meshes: 3D Labeling Tool (3DLT). This is a standalone,\nfeature-heavy and cross-platform software that does not require installation\nand can run on Windows, macOS and Linux-based distributions. Instead of\nlabeling the same object on every image separately like current tools, we use\ndepth information to reconstruct a triangular mesh from said images and label\nthe object only once on the aforementioned mesh. We use registration to\nsimplify 3D labeling, outlier detection to improve 2D bounding box calculation\nand surface reconstruction to expand labeling possibility to large point\nclouds. Our tool is tested against state of the art methods and it greatly\nsurpasses them in terms of speed while preserving accuracy and ease of use.\n", "date": "2022-07-23", "categories": []} {"id": "2207.11753", "title": "Label-Guided Auxiliary Training Improves 3D Object Detector", "abstract": " Detecting 3D objects from point clouds is a practical yet challenging task\nthat has attracted increasing attention recently. In this paper, we propose a\nLabel-Guided auxiliary training method for 3D object detection (LG3D), which\nserves as an auxiliary network to enhance the feature learning of existing 3D\nobject detectors. Specifically, we propose two novel modules: a\nLabel-Annotation-Inducer that maps annotations and point clouds in bounding\nboxes to task-specific representations and a Label-Knowledge-Mapper that\nassists the original features to obtain detection-critical representations. The\nproposed auxiliary network is discarded in inference and thus has no extra\ncomputational cost at test time. We conduct extensive experiments on both\nindoor and outdoor datasets to verify the effectiveness of our approach. For\nexample, our proposed LG3D improves VoteNet by 2.5% and 3.1% mAP on the SUN\nRGB-D and ScanNetV2 datasets, respectively.\n", "date": "2022-07-24", "categories": []} {"id": "2207.08531", "title": "DID-M3D: Decoupling Instance Depth for Monocular 3D Object Detection", "abstract": " Monocular 3D detection has drawn much attention from the community due to its\nlow cost and setup simplicity. It takes an RGB image as input and predicts 3D\nboxes in the 3D space. The most challenging sub-task lies in the instance depth\nestimation. Previous works usually use a direct estimation method. However, in\nthis paper we point out that the instance depth on the RGB image is\nnon-intuitive. It is coupled by visual depth clues and instance attribute\nclues, making it hard to be directly learned in the network. Therefore, we\npropose to reformulate the instance depth to the combination of the instance\nvisual surface depth (visual depth) and the instance attribute depth (attribute\ndepth). The visual depth is related to objects' appearances and positions on\nthe image. By contrast, the attribute depth relies on objects' inherent\nattributes, which are invariant to the object affine transformation on the\nimage. Correspondingly, we decouple the 3D location uncertainty into visual\ndepth uncertainty and attribute depth uncertainty. By combining different types\nof depths and associated uncertainties, we can obtain the final instance depth.\nFurthermore, data augmentation in monocular 3D detection is usually limited due\nto the physical nature, hindering the boost of performance. Based on the\nproposed instance depth disentanglement strategy, we can alleviate this\nproblem. Evaluated on KITTI, our method achieves new state-of-the-art results,\nand extensive ablation studies validate the effectiveness of each component in\nour method. The codes are released at https://github.com/SPengLiang/DID-M3D.\n", "date": "2022-07-18", "categories": []} {"id": "2207.10758", "title": "DEVIANT: Depth EquiVarIAnt NeTwork for Monocular 3D Object Detection", "abstract": " Modern neural networks use building blocks such as convolutions that are\nequivariant to arbitrary 2D translations. However, these vanilla blocks are not\nequivariant to arbitrary 3D translations in the projective manifold. Even then,\nall monocular 3D detectors use vanilla blocks to obtain the 3D coordinates, a\ntask for which the vanilla blocks are not designed for. This paper takes the\nfirst step towards convolutions equivariant to arbitrary 3D translations in the\nprojective manifold. Since the depth is the hardest to estimate for monocular\ndetection, this paper proposes Depth EquiVarIAnt NeTwork (DEVIANT) built with\nexisting scale equivariant steerable blocks. As a result, DEVIANT is\nequivariant to the depth translations in the projective manifold whereas\nvanilla networks are not. The additional depth equivariance forces the DEVIANT\nto learn consistent depth estimates, and therefore, DEVIANT achieves\nstate-of-the-art monocular 3D detection results on KITTI and Waymo datasets in\nthe image-only category and performs competitively to methods using extra\ninformation. Moreover, DEVIANT works better than vanilla networks in\ncross-dataset evaluation. Code and models at\nhttps://github.com/abhi1kumar/DEVIANT\n", "date": "2022-07-21", "categories": []} {"id": "2207.10205", "title": "On the Robustness of 3D Object Detectors", "abstract": " In recent years, significant progress has been achieved for 3D object\ndetection on point clouds thanks to the advances in 3D data collection and deep\nlearning techniques. Nevertheless, 3D scenes exhibit a lot of variations and\nare prone to sensor inaccuracies as well as information loss during\npre-processing. Thus, it is crucial to design techniques that are robust\nagainst these variations. This requires a detailed analysis and understanding\nof the effect of such variations. This work aims to analyze and benchmark\npopular point-based 3D object detectors against several data corruptions. To\nthe best of our knowledge, we are the first to investigate the robustness of\npoint-based 3D object detectors. To this end, we design and evaluate\ncorruptions that involve data addition, reduction, and alteration. We further\nstudy the robustness of different modules against local and global variations.\nOur experimental results reveal several intriguing findings. For instance, we\nshow that methods that integrate Transformers at a patch or object level lead\nto increased robustness, compared to using Transformers at the point level.\n", "date": "2022-07-20", "categories": []} {"id": "2207.10316", "title": "AutoAlignV2: Deformable Feature Aggregation for Dynamic Multi-Modal 3D\n Object Detection", "abstract": " Point clouds and RGB images are two general perceptional sources in\nautonomous driving. The former can provide accurate localization of objects,\nand the latter is denser and richer in semantic information. Recently,\nAutoAlign presents a learnable paradigm in combining these two modalities for\n3D object detection. However, it suffers from high computational cost\nintroduced by the global-wise attention. To solve the problem, we propose\nCross-Domain DeformCAFA module in this work. It attends to sparse learnable\nsampling points for cross-modal relational modeling, which enhances the\ntolerance to calibration error and greatly speeds up the feature aggregation\nacross different modalities. To overcome the complex GT-AUG under multi-modal\nsettings, we design a simple yet effective cross-modal augmentation strategy on\nconvex combination of image patches given their depth information. Moreover, by\ncarrying out a novel image-level dropout training scheme, our model is able to\ninfer in a dynamic manner. To this end, we propose AutoAlignV2, a faster and\nstronger multi-modal 3D detection framework, built on top of AutoAlign.\nExtensive experiments on nuScenes benchmark demonstrate the effectiveness and\nefficiency of AutoAlignV2. Notably, our best model reaches 72.4 NDS on nuScenes\ntest leaderboard, achieving new state-of-the-art results among all published\nmulti-modal 3D object detectors. Code will be available at\nhttps://github.com/zehuichen123/AutoAlignV2.\n", "date": "2022-07-21", "categories": []} {"id": "2207.10589", "title": "Boosting 3D Object Detection via Object-Focused Image Fusion", "abstract": " 3D object detection has achieved remarkable progress by taking point clouds\nas the only input. However, point clouds often suffer from incomplete geometric\nstructures and the lack of semantic information, which makes detectors hard to\naccurately classify detected objects. In this work, we focus on how to\neffectively utilize object-level information from images to boost the\nperformance of point-based 3D detector. We present DeMF, a simple yet effective\nmethod to fuse image information into point features. Given a set of point\nfeatures and image feature maps, DeMF adaptively aggregates image features by\ntaking the projected 2D location of the 3D point as reference. We evaluate our\nmethod on the challenging SUN RGB-D dataset, improving state-of-the-art results\nby a large margin (+2.1 mAP@0.25 and +2.3mAP@0.5). Code is available at\nhttps://github.com/haoy945/DeMF.\n", "date": "2022-07-21", "categories": []} {"id": "2207.09332", "title": "Rethinking IoU-based Optimization for Single-stage 3D Object Detection", "abstract": " Since Intersection-over-Union (IoU) based optimization maintains the\nconsistency of the final IoU prediction metric and losses, it has been widely\nused in both regression and classification branches of single-stage 2D object\ndetectors. Recently, several 3D object detection methods adopt IoU-based\noptimization and directly replace the 2D IoU with 3D IoU. However, such a\ndirect computation in 3D is very costly due to the complex implementation and\ninefficient backward operations. Moreover, 3D IoU-based optimization is\nsub-optimal as it is sensitive to rotation and thus can cause training\ninstability and detection performance deterioration. In this paper, we propose\na novel Rotation-Decoupled IoU (RDIoU) method that can mitigate the\nrotation-sensitivity issue, and produce more efficient optimization objectives\ncompared with 3D IoU during the training stage. Specifically, our RDIoU\nsimplifies the complex interactions of regression parameters by decoupling the\nrotation variable as an independent term, yet preserving the geometry of 3D\nIoU. By incorporating RDIoU into both the regression and classification\nbranches, the network is encouraged to learn more precise bounding boxes and\nconcurrently overcome the misalignment issue between classification and\nregression. Extensive experiments on the benchmark KITTI and Waymo Open Dataset\nvalidate that our RDIoU method can bring substantial improvement for the\nsingle-stage 3D object detection.\n", "date": "2022-07-19", "categories": []} {"id": "2207.09805", "title": "Multimodal Transformer for Automatic 3D Annotation and Object Detection", "abstract": " Despite a growing number of datasets being collected for training 3D object\ndetection models, significant human effort is still required to annotate 3D\nboxes on LiDAR scans. To automate the annotation and facilitate the production\nof various customized datasets, we propose an end-to-end multimodal transformer\n(MTrans) autolabeler, which leverages both LiDAR scans and images to generate\nprecise 3D box annotations from weak 2D bounding boxes. To alleviate the\npervasive sparsity problem that hinders existing autolabelers, MTrans densifies\nthe sparse point clouds by generating new 3D points based on 2D image\ninformation. With a multi-task design, MTrans segments the\nforeground/background, densifies LiDAR point clouds, and regresses 3D boxes\nsimultaneously. Experimental results verify the effectiveness of the MTrans for\nimproving the quality of the generated labels. By enriching the sparse point\nclouds, our method achieves 4.48\\% and 4.03\\% better 3D AP on KITTI moderate\nand hard samples, respectively, versus the state-of-the-art autolabeler. MTrans\ncan also be extended to improve the accuracy for 3D object detection, resulting\nin a remarkable 89.45\\% AP on KITTI hard samples. Codes are at\n\\url{https://github.com/Cliu2/MTrans}.\n", "date": "2022-07-20", "categories": []} {"id": "2207.09412", "title": "Det6D: A Ground-Aware Full-Pose 3D Object Detector for Improving Terrain\n Robustness", "abstract": " Accurate 3D object detection with LiDAR is critical for autonomous driving.\nExisting research is all based on the flat-world assumption. However, the\nactual road can be complex with steep sections, which breaks the premise.\nCurrent methods suffer from performance degradation in this case due to\ndifficulty correctly detecting objects on sloped terrain. In this work, we\npropose Det6D, the first full-degree-of-freedom 3D object detector without\nspatial and postural limitations, to improve terrain robustness. We choose the\npoint-based framework by founding their capability of detecting objects in the\nentire spatial range. To predict full-degree poses, including pitch and roll,\nwe design a ground-aware orientation branch that leverages the local ground\nconstraints. Given the difficulty of long-tail non-flat scene data collection\nand 6D pose annotation, we present Slope-Aug, a data augmentation method for\nsynthesizing non-flat terrain from existing datasets recorded in flat scenes.\nExperiments on various datasets demonstrate the effectiveness and robustness of\nour method in different terrains. We further conducted an extended experiment\nto explore how the network predicts the two extra poses. The proposed modules\nare plug-and-play for existing point-based frameworks. The code is available at\nhttps://github.com/HITSZ-NRSL/De6D.\n", "date": "2022-07-19", "categories": []} {"id": "2207.01030", "title": "Boosting Single-Frame 3D Object Detection by Simulating Multi-Frame\n Point Clouds", "abstract": " To boost a detector for single-frame 3D object detection, we present a new\napproach to train it to simulate features and responses following a detector\ntrained on multi-frame point clouds. Our approach needs multi-frame point\nclouds only when training the single-frame detector, and once trained, it can\ndetect objects with only single-frame point clouds as inputs during the\ninference. We design a novel Simulated Multi-Frame Single-Stage object Detector\n(SMF-SSD) framework to realize the approach: multi-view dense object fusion to\ndensify ground-truth objects to generate a multi-frame point cloud;\nself-attention voxel distillation to facilitate one-to-many knowledge transfer\nfrom multi- to single-frame voxels; multi-scale BEV feature distillation to\ntransfer knowledge in low-level spatial and high-level semantic BEV features;\nand adaptive response distillation to activate single-frame responses of high\nconfidence and accurate localization. Experimental results on the Waymo test\nset show that our SMF-SSD consistently outperforms all state-of-the-art\nsingle-frame 3D object detectors for all object classes of difficulty levels 1\nand 2 in terms of both mAP and mAPH.\n", "date": "2022-07-03", "categories": []} {"id": "2207.05200", "title": "Real-Time And Robust 3D Object Detection with Roadside LiDARs", "abstract": " This work aims to address the challenges in autonomous driving by focusing on\nthe 3D perception of the environment using roadside LiDARs. We design a 3D\nobject detection model that can detect traffic participants in roadside LiDARs\nin real-time. Our model uses an existing 3D detector as a baseline and improves\nits accuracy. To prove the effectiveness of our proposed modules, we train and\nevaluate the model on three different vehicle and infrastructure datasets. To\nshow the domain adaptation ability of our detector, we train it on an\ninfrastructure dataset from China and perform transfer learning on a different\ndataset recorded in Germany. We do several sets of experiments and ablation\nstudies for each module in the detector that show that our model outperforms\nthe baseline by a significant margin, while the inference speed is at 45 Hz (22\nms). We make a significant contribution with our LiDAR-based 3D detector that\ncan be used for smart city applications to provide connected and automated\nvehicles with a far-reaching view. Vehicles that are connected to the roadside\nsensors can get information about other vehicles around the corner to improve\ntheir path and maneuver planning and to increase road traffic safety.\n", "date": "2022-07-11", "categories": []} {"id": "2207.05497", "title": "Paint and Distill: Boosting 3D Object Detection with Semantic Passing\n Network", "abstract": " 3D object detection task from lidar or camera sensors is essential for\nautonomous driving. Pioneer attempts at multi-modality fusion complement the\nsparse lidar point clouds with rich semantic texture information from images at\nthe cost of extra network designs and overhead. In this work, we propose a\nnovel semantic passing framework, named SPNet, to boost the performance of\nexisting lidar-based 3D detection models with the guidance of rich context\npainting, with no extra computation cost during inference. Our key design is to\nfirst exploit the potential instructive semantic knowledge within the\nground-truth labels by training a semantic-painted teacher model and then guide\nthe pure-lidar network to learn the semantic-painted representation via\nknowledge passing modules at different granularities: class-wise passing,\npixel-wise passing and instance-wise passing. Experimental results show that\nthe proposed SPNet can seamlessly cooperate with most existing 3D detection\nframeworks with 1~5% AP gain and even achieve new state-of-the-art 3D detection\nperformance on the KITTI test benchmark. Code is available at:\nhttps://github.com/jb892/SPNet.\n", "date": "2022-07-12", "categories": []} {"id": "2207.04448", "title": "Mix-Teaching: A Simple, Unified and Effective Semi-Supervised Learning\n Framework for Monocular 3D Object Detection", "abstract": " Monocular 3D object detection is an essential perception task for autonomous\ndriving. However, the high reliance on large-scale labeled data make it costly\nand time-consuming during model optimization. To reduce such over-reliance on\nhuman annotations, we propose Mix-Teaching, an effective semi-supervised\nlearning framework applicable to employ both labeled and unlabeled images in\ntraining stage. Mix-Teaching first generates pseudo-labels for unlabeled images\nby self-training. The student model is then trained on the mixed images\npossessing much more intensive and precise labeling by merging instance-level\nimage patches into empty backgrounds or labeled images. This is the first to\nbreak the image-level limitation and put high-quality pseudo labels from multi\nframes into one image for semi-supervised training. Besides, as a result of the\nmisalignment between confidence score and localization quality, it's hard to\ndiscriminate high-quality pseudo-labels from noisy predictions using only\nconfidence-based criterion. To that end, we further introduce an\nuncertainty-based filter to help select reliable pseudo boxes for the above\nmixing operation. To the best of our knowledge, this is the first unified SSL\nframework for monocular 3D object detection. Mix-Teaching consistently improves\nMonoFlex and GUPNet by significant margins under various labeling ratios on\nKITTI dataset. For example, our method achieves around +6.34% AP@0.7\nimprovement against the GUPNet baseline on validation set when using only 10%\nlabeled data. Besides, by leveraging full training set and the additional 48K\nraw images of KITTI, it can further improve the MonoFlex by +4.65% improvement\non AP@0.7 for car detection, reaching 18.54% AP@0.7, which ranks the 1st place\namong all monocular based methods on KITTI test leaderboard. The code and\npretrained models will be released at\nhttps://github.com/yanglei18/Mix-Teaching.\n", "date": "2022-07-10", "categories": []} {"id": "2207.04781", "title": "MT-Net Submission to the Waymo 3D Detection Leaderboard", "abstract": " In this technical report, we introduce our submission to the Waymo 3D\nDetection leaderboard. Our network is based on the Centerpoint architecture,\nbut with significant improvements. We design a 2D backbone to utilize\nmulti-scale features for better detecting objects with various sizes, together\nwith an optimal transport-based target assignment strategy, which dynamically\nassigns richer supervision signals to the detection candidates. We also apply\ntest-time augmentation and model-ensemble for further improvements. Our\nsubmission currently ranks 4th place with 78.45 mAPH on the Waymo 3D Detection\nleaderboard.\n", "date": "2022-07-11", "categories": []} {"id": "2207.01987", "title": "Open-Vocabulary 3D Detection via Image-level Class and Debiased\n Cross-modal Contrastive Learning", "abstract": " Current point-cloud detection methods have difficulty detecting the\nopen-vocabulary objects in the real world, due to their limited generalization\ncapability. Moreover, it is extremely laborious and expensive to collect and\nfully annotate a point-cloud detection dataset with numerous classes of\nobjects, leading to the limited classes of existing point-cloud datasets and\nhindering the model to learn general representations to achieve open-vocabulary\npoint-cloud detection. As far as we know, we are the first to study the problem\nof open-vocabulary 3D point-cloud detection. Instead of seeking a point-cloud\ndataset with full labels, we resort to ImageNet1K to broaden the vocabulary of\nthe point-cloud detector. We propose OV-3DETIC, an Open-Vocabulary 3D DETector\nusing Image-level Class supervision. Specifically, we take advantage of two\nmodalities, the image modality for recognition and the point-cloud modality for\nlocalization, to generate pseudo labels for unseen classes. Then we propose a\nnovel debiased cross-modal contrastive learning method to transfer the\nknowledge from image modality to point-cloud modality during training. Without\nhurting the latency during inference, OV-3DETIC makes the point-cloud detector\ncapable of achieving open-vocabulary detection. Extensive experiments\ndemonstrate that the proposed OV-3DETIC achieves at least 10.77 % mAP\nimprovement (absolute value) and 9.56 % mAP improvement (absolute value) by a\nwide range of baselines on the SUN-RGBD dataset and ScanNet dataset,\nrespectively. Besides, we conduct sufficient experiments to shed light on why\nthe proposed OV-3DETIC works.\n", "date": "2022-07-05", "categories": []} {"id": "2206.14971", "title": "Boosting 3D Object Detection by Simulating Multimodality on Point Clouds", "abstract": " This paper presents a new approach to boost a single-modality (LiDAR) 3D\nobject detector by teaching it to simulate features and responses that follow a\nmulti-modality (LiDAR-image) detector. The approach needs LiDAR-image data only\nwhen training the single-modality detector, and once well-trained, it only\nneeds LiDAR data at inference. We design a novel framework to realize the\napproach: response distillation to focus on the crucial response samples and\navoid the background samples; sparse-voxel distillation to learn voxel\nsemantics and relations from the estimated crucial voxels; a fine-grained\nvoxel-to-point distillation to better attend to features of small and distant\nobjects; and instance distillation to further enhance the deep-feature\nconsistency. Experimental results on the nuScenes dataset show that our\napproach outperforms all SOTA LiDAR-only 3D detectors and even surpasses the\nbaseline LiDAR-image detector on the key NDS metric, filling 72% mAP gap\nbetween the single- and multi-modality detectors.\n", "date": "2022-06-30", "categories": []} {"id": "2206.13858", "title": "Accurate and Real-time Pseudo Lidar Detection: Is Stereo Neural Network\n Really Necessary?", "abstract": " The proposal of Pseudo-Lidar representation has significantly narrowed the\ngap between visual-based and active Lidar-based 3D object detection. However,\ncurrent researches exclusively focus on pushing the accuracy improvement of\nPseudo-Lidar by taking the advantage of complex and time-consuming neural\nnetworks. Seldom explore the profound characteristics of Pseudo-Lidar\nrepresentation to obtain the promoting opportunities. In this paper, we dive\ndeep into the pseudo Lidar representation and argue that the performance of 3D\nobject detection is not fully dependent on the high precision stereo depth\nestimation. We demonstrate that even for the unreliable depth estimation, with\nproper data processing and refining, it can achieve comparable 3D object\ndetection accuracy. With this finding, we further show the possibility that\nutilizing fast but inaccurate stereo matching algorithms in the Pseudo-Lidar\nsystem to achieve low latency responsiveness. In the experiments, we develop a\nsystem with a less powerful stereo matching predictor and adopt the proposed\nrefinement schemes to improve the accuracy. The evaluation on the KITTI\nbenchmark shows that the presented system achieves competitive accuracy to the\nstate-of-the-art approaches with only 23 ms computing, showing it is a suitable\ncandidate for deploying to real car-hold applications.\n", "date": "2022-06-28", "categories": []} {"id": "2206.12738", "title": "Self-Supervised 3D Monocular Object Detection by Recycling Bounding\n Boxes", "abstract": " Modern object detection architectures are moving towards employing\nself-supervised learning (SSL) to improve performance detection with related\npretext tasks. Pretext tasks for monocular 3D object detection have not yet\nbeen explored yet in literature. The paper studies the application of\nestablished self-supervised bounding box recycling by labeling random windows\nas the pretext task. The classifier head of the 3D detector is trained to\nclassify random windows containing different proportions of the ground truth\nobjects, thus handling the foreground-background imbalance. We evaluate the\npretext task using the RTM3D detection model as baseline, with and without the\napplication of data augmentation. We demonstrate improvements of between 2-3 %\nin mAP 3D and 0.9-1.5 % BEV scores using SSL over the baseline scores. We\npropose the inverse class frequency re-weighted (ICFW) mAP score that\nhighlights improvements in detection for low frequency classes in a class\nimbalanced dataset with long tails. We demonstrate improvements in ICFW both\nmAP 3D and BEV scores to take into account the class imbalance in the KITTI\nvalidation dataset. We see 4-5 % increase in ICFW metric with the pretext task.\n", "date": "2022-06-25", "categories": []} {"id": "2206.03657", "title": "Delving into the Pre-training Paradigm of Monocular 3D Object Detection", "abstract": " The labels of monocular 3D object detection (M3OD) are expensive to obtain.\nMeanwhile, there usually exists numerous unlabeled data in practical\napplications, and pre-training is an efficient way of exploiting the knowledge\nin unlabeled data. However, the pre-training paradigm for M3OD is hardly\nstudied. We aim to bridge this gap in this work. To this end, we first draw two\nobservations: (1) The guideline of devising pre-training tasks is imitating the\nrepresentation of the target task. (2) Combining depth estimation and 2D object\ndetection is a promising M3OD pre-training baseline. Afterwards, following the\nguideline, we propose several strategies to further improve this baseline,\nwhich mainly include target guided semi-dense depth estimation, keypoint-aware\n2D object detection, and class-level loss adjustment. Combining all the\ndeveloped techniques, the obtained pre-training framework produces pre-trained\nbackbones that improve M3OD performance significantly on both the KITTI-3D and\nnuScenes benchmarks. For example, by applying a DLA34 backbone to a naive\ncenter-based M3OD detector, the moderate ${\\rm AP}_{3D}70$ score of Car on the\nKITTI-3D testing set is boosted by 18.71\\% and the NDS score on the nuScenes\nvalidation set is improved by 40.41\\% relatively.\n", "date": "2022-06-08", "categories": []} {"id": "2206.07372", "title": "MonoGround: Detecting Monocular 3D Objects from the Ground", "abstract": " Monocular 3D object detection has attracted great attention for its\nadvantages in simplicity and cost. Due to the ill-posed 2D to 3D mapping\nessence from the monocular imaging process, monocular 3D object detection\nsuffers from inaccurate depth estimation and thus has poor 3D detection\nresults. To alleviate this problem, we propose to introduce the ground plane as\na prior in the monocular 3d object detection. The ground plane prior serves as\nan additional geometric condition to the ill-posed mapping and an extra source\nin depth estimation. In this way, we can get a more accurate depth estimation\nfrom the ground. Meanwhile, to take full advantage of the ground plane prior,\nwe propose a depth-align training strategy and a precise two-stage depth\ninference method tailored for the ground plane prior. It is worth noting that\nthe introduced ground plane prior requires no extra data sources like LiDAR,\nstereo images, and depth information. Extensive experiments on the KITTI\nbenchmark show that our method could achieve state-of-the-art results compared\nwith other methods while maintaining a very fast speed. Our code and models are\navailable at https://github.com/cfzd/MonoGround.\n", "date": "2022-06-15", "categories": []} {"id": "2206.03666", "title": "Depth Estimation Matters Most: Improving Per-Object Depth Estimation for\n Monocular 3D Detection and Tracking", "abstract": " Monocular image-based 3D perception has become an active research area in\nrecent years owing to its applications in autonomous driving. Approaches to\nmonocular 3D perception including detection and tracking, however, often yield\ninferior performance when compared to LiDAR-based techniques. Through\nsystematic analysis, we identified that per-object depth estimation accuracy is\na major factor bounding the performance. Motivated by this observation, we\npropose a multi-level fusion method that combines different representations\n(RGB and pseudo-LiDAR) and temporal information across multiple frames for\nobjects (tracklets) to enhance per-object depth estimation. Our proposed fusion\nmethod achieves the state-of-the-art performance of per-object depth estimation\non the Waymo Open Dataset, the KITTI detection dataset, and the KITTI MOT\ndataset. We further demonstrate that by simply replacing estimated depth with\nfusion-enhanced depth, we can achieve significant improvements in monocular 3D\nperception tasks, including detection and tracking.\n", "date": "2022-06-08", "categories": []} {"id": "2206.00960", "title": "SparseDet: Towards End-to-End 3D Object Detection", "abstract": " In this paper, we propose SparseDet for end-to-end 3D object detection from\npoint cloud. Existing works on 3D object detection rely on dense object\ncandidates over all locations in a 3D or 2D grid following the mainstream\nmethods for object detection in 2D images. However, this dense paradigm\nrequires expertise in data to fulfill the gap between label and detection. As a\nnew detection paradigm, SparseDet maintains a fixed set of learnable proposals\nto represent latent candidates and directly perform classification and\nlocalization for 3D objects through stacked transformers. It demonstrates that\neffective 3D object detection can be achieved with none of post-processing such\nas redundant removal and non-maximum suppression. With a properly designed\nnetwork, SparseDet achieves highly competitive detection accuracy while running\nwith a more efficient speed of 34.5 FPS. We believe this end-to-end paradigm of\nSparseDet will inspire new thinking on the sparsity of 3D object detection.\n", "date": "2022-06-02", "categories": []} {"id": "2606.21309", "title": "WildBox: A Dataset and Benchmark for Aerial Monocular 3D Detection of African Savanna Wildlife", "abstract": "We introduce WildBox, a dataset and benchmark for monocular 3D detection of wildlife from drone video, comprising 237,505 3D bounding box annotations across seven African savanna species grouped into six benchmark classes. Annotations follow a KITTI/Omni3D-compatible format in a per-segment scale-normalised camera frame, with instance identities maintained across each segment. We evaluate two open-vocabulary monocular 3D architectures, OVMono3D-LIFT and DetAny3D, under zero-shot, ground-truth 2D box prompt, and supervised fine-tuning protocols. Open-vocabulary 2D foundation models provide usable zero-shot wildlife localisation (50.55 AP@50), but zero-shot 3D detection collapses to 0.00 AP across both architectures and every 2D-input condition tested, including ground-truth 2D box prompts, thus isolating the failure to the 3D stage. Fine-tuning on WildBox recovers performance to 8.68 +/- 0.47 AP-BEV@0.50 and 13.17 +/- 0.69 AP3D macro. Depth contributes 84% of normalised Hausdorff distance after fine-tuning and over 99% in zero-shot, identifying monocular aerial depth as the dominant open problem in this regime. A coarse-to-fine curriculum, i.e. pretraining on a merged zebra class before fine-tuning on the Grevy's/plains split, improves macro 3D performance with less total compute, with the largest gains on the two zebra subclasses. WildBox is released with video-level splits, evaluation code, and baseline checkpoints to enable progress in 3D wildlife perception from drone video.", "date": "2026-06-19", "categories": []} {"id": "2606.24805", "title": "DDStereo: Efficient Dual Decoder Transformers for Stereo 3D Road Anomaly Detection", "abstract": "Stereo-based 3D object detection still faces two critical safety challenges: real-time performance and open-set generalization. Existing stereo 3D methods typically achieve twice the accuracy of monocular methods but suffer from significantly lower inference speeds, making them unsuitable for real-time applications. Meanwhile, recent advances in open-world detection have introduced open-set and open-vocabulary algorithms in monocular 2D and 3D settings, yet stereo-based open-set detection remains largely unexplored. To bridge this gap, we propose DDStereo, a novel Dual-Decoder Stereo Transformer for real-time open-set 3D object detection. DDStereo features two lightweight decoder branches: one for open-set foreground 2D detection and the other for 3D attribute regression. These decoders share object-level queries to achieve unified target-level alignment. To enhance inference efficiency, we designed a compact disparity feature extractor and a streamlined decoder architecture. Experiments on public stereo 3D benchmarks demonstrate that DDStereo achieves state-of-the-art accuracy under both closed-set and open-set protocols. Notably, our method surpasses existing stereo 3D detectors in inference speed and, for the first time, achieves real-time performance comparable to monocular approaches.", "date": "2026-06-23", "categories": []} {"id": "2606.25652", "title": "Auto-Labelling-Based Domain Transfer for 3D Object Detection on a Bicycle-Mounted LiDAR Platform", "abstract": "Reliable 3D perception of vulnerable road users (VRUs) such as cyclists and pedestrians is essential for their safety in urban traffic and a core requirement for autonomous driving (AD). Alongside advances in vehicle-based perception, research increasingly equips bicycles with sensors to study traffic from a perspective native to VRUs. Such platforms still rely on LiDAR detectors originally trained on vehicle data, yet annotated 3D data from a cyclist's perspective is scarce. How well these detectors generalise to this setting has not been evaluated. We present a 3D object detection benchmark of 1,027 annotated LiDAR keyframes (over 18,000 3D bounding boxes) from the FUSE-Bike platform in urban Munich. We evaluate four nuScenes-pre-trained detectors against 1,854 human-verified ground-truth (GT) boxes both in their original form and after finetuning on training labels produced by a VRU-dedicated auto-labelling pipeline that requires no manual annotation. The zero-shot domain gap is concentrated on the VRU classes. Finetuning recovers most of it, improving mean average precision (mAP) by up to 23.4 points with the largest gains on pedestrians and cyclists, and the adapted detectors even surpass the quality of the auto-labels they were trained on. The benchmark provides a reproducible baseline for VRU-centric 3D detection and shows that auto-labels are a viable substitute for manual annotation when adapting vehicle-trained detectors to a cyclist platform.", "date": "2026-06-24", "categories": []} {"id": "2606.29963", "title": "Explainability-Aware Frustum Attack: Exposing Structural Vulnerabilities in LiDAR-Based 3D Object Detectors", "abstract": "The structural vulnerabilities of point cloud-based 3D object detectors remain poorly understood. Prior work has studied adversarial robustness primarily on isolated 3D object models, while recent LiDAR spoofing attacks target richer and more realistic driving scenes but focus mainly on physical realizability rather than understanding detector behavior or attack efficiency. In this work, we investigate how LiDAR-based detectors rely on spatial evidence in complex scenes and whether these reliance patterns can be exploited to induce failures more efficiently. To this end, we propose an explainability-guided adversarial analysis methodology. We introduce the Saliency-LiDAR (SALL) method, which aggregates Integrated Gradient attributions across scenes to produce universal saliency maps for LiDAR-based 3D object detectors. Guided by these maps, we design the Explainability-aware Frustum Attack (EFA), which selectively perturbs only the most influential frustums rather than uniformly attacking entire object regions. Experiments on KITTI and nuScenes, across detectors such as PointPillars and SECOND, show that EFA reduces detection recall by more than 15 percentage points while requiring 25-50% fewer perturbed frustums than the state-of-the-art non-saliency-aware baseline. These findings reveal that modern 3D detectors concentrate discriminative evidence in a small subset of spatial regions, exposing a structural robustness vulnerability in current LiDAR perception systems. Our code is released at https://github.com/SecMindLab/Saliency_LiDAR.", "date": "2026-06-29", "categories": []} {"id": "2606.31096", "title": "Horizon3D: Sparse Radar-Camera Fusion for Long-Range 3D Perception in Autonomous Driving", "abstract": "Long-range 3D object detection is critical for safe autonomous driving at highway speeds, yet existing radar-camera fusion methods remain limited at extended ranges. BEV-based methods capture scene-level context but incur rapidly growing computation and often lose fine-grained object detail, while query-based methods are efficient but provide limited scene-level context. Temporal fusion further requires both multi-frame accumulation for sparse distant observations and object-level motion modeling for fast-moving objects. We propose Horizon3D, a sparse radar-camera fusion framework for long-range 3D object detection that combines Gaussian primitives with sparse BEV features. Horizon3D initializes Gaussian primitives at radar- and camera-estimated object keypoints using Keypoint-Guided Gaussian Initialization, refines them through Object-Centric Sparse Fusion, and splats them onto the BEV plane to fuse object-level detail with sparse radar BEV context. It further introduces Dual-Path Temporal Fusion, which aggregates temporal cues through a BEV path for scene-level accumulation and a Gaussian path for object-level motion propagation. Experiments on TruckScenes show that Horizon3D achieves state-of-the-art radar-camera 3D detection performance. On the validation set, it outperforms the previous best method by +3.0 NDS and +1.6 mAP while maintaining competitive inference speed.", "date": "2026-06-30", "categories": []} {"id": "2606.31895", "title": "RESOLVE: A Multi-Resolution and Multi-Modal Dataset for Roadside Cooperative Perception", "abstract": "LiDAR has increasingly been integrated into traffic cameras to expand coverage and mitigate occlusion in roadside cooperative perception. However, how unimodal and camera-LiDAR fusion architectures behave under variations in LiDAR point sparsity induced by sensor configurations and scene-dependent sensing conditions remains underexplored. We introduce RESOLVE, a large-scale real-world benchmark dataset featuring multi-resolution roadside LiDAR and synchronized camera-LiDAR sensing for systematic evaluation of unimodal and fusion-based architectures in roadside 3D detection and tracking. RESOLVE contains over 100k images and 26k point cloud frames with 220k manually annotated bounding boxes, captured at a real-world urban intersection across diverse lighting and weather conditions and spanning 10 classes of traffic participants. In particular, RESOLVE enables controlled evaluation across three LiDAR resolution levels while keeping all other sensing and environmental factors fixed. This allows fair cross-architecture comparisons under point cloud distribution shifts resulting from resolution variations, sensing distance, and training-inference resolution mismatches. Results from extensive benchmark experiments reveal insights into how multimodal fusion can compensate for LiDAR point sparsity, offering clues for designing cost-efficient roadside multimodal perception. The dataset and benchmark codes are available at https://github.com/ASU-Suo-Lab/RESOLVE.", "date": "2026-06-30", "categories": []} {"id": "2607.00273", "title": "MVDGC: Joint 3D and 2D Multi-view Pedestrian Detection via Dual Geometric Constraints", "abstract": "The core challenge in multi-view pedestrian detection (MVPD) lies in effective aggregation of visual features from different viewpoints for robust occlusion reasoning. Recent approaches have addressed this by first projecting image-view features onto a Bird's Eye View (BEV) map, where ground localization is then performed. Despite impressive performance, the perspective transformation induces severe distortion, causing spatial structure break and degrading the quality of object feature extraction. The blurred and ambiguous features hinder accurate BEV point localization, especially in densely populated regions. Moreover, the strong mutual relationship between the BEV ground point and image bounding boxes is not capitalized on. Although multi-view consistency of 2D detections can serve as a powerful constraint in BEV space, these detections are commonly treated as auxiliary signals rather than being jointly optimized with the primary task.In this work, we propose \\textbf{MVDGC}, a unified framework that \\emph{jointly estimates pedestrian locations on the BEV plane and 2D bounding boxes in image views}. MVDGC employs a \\emph{sparse set of 3D cylindrical queries} that embraces geometric context across both BEV and image views, enforcing dual spatial constraints for precise localization. Specifically, the geometric constraints is established by modeling each pedestrian as a vertical cylinder whose center lies on the BEV plane and whose projection casts a rectangular box in the image views. These queries function as shape anchors that directly extract 2D features from the intact image-view features using camera projection, eliminating projection-induced distortions. The 3D cylindrical query enables the unification of BEV and ImV localization into a single task: 3D cylinder position and shape refinement. Code is available at: https://github.com/UARK-AICV/MVDGC", "date": "2026-06-30", "categories": []}