In autonomous driving, vision-centric 3D object detection recognizes and localizes 3D objects from RGB images. However, due to high annotation costs and diverse outdoor scenes, training data often fails to cover all possible test scenarios, known as the out-of-distribution (OOD) issue. Training-free image editing offers a promising solution for improving model robustness by training data enhancement without any modifications to pre-trained diffusion models. Nevertheless, inversion-based methods often suffer from limited effectiveness and inherent inaccuracies, while recent rectified-flow-based approaches struggle to preserve objects with accurate 3D geometry. In this paper, we propose DriveFlow, a Rectified Flow Adaptation method for training data enhancement in autonomous driving based on pre-trained Text-to-Image flow models. Based on frequency decomposition, DriveFlow introduces two strategies to adapt noise-free editing paths derived from text-conditioned velocities. 1) High-Frequency Foreground Preservation: DriveFlow incorporates a high-frequency alignment loss for foreground to maintain precise 3D object geometry. 2) Dual-Frequency Background Optimization: DriveFlow also conducts dual-frequency optimization for background, balancing editing flexibility and semantic consistency. Comprehensive experiments validate the effectiveness and efficiency of DriveFlow, demonstrating comprehensive performance improvements on all categories across OOD scenarios.

3D Multimodal Large Language Models (MLLMs) have recently made substantial advancements. However, their potential remains untapped, primarily due to the limited quantity and suboptimal quality of 3D datasets. Current approaches attempt to transfer knowledge from 2D MLLMs to expand 3D instruction data, but still face modality and domain gaps. To this end, we introduce PiSA-Engine (Point-Self-Augmented-Engine), a new framework for generating instruction point-language datasets enriched with 3D spatial semantics. We observe that existing 3D MLLMs offer a comprehensive understanding of point clouds for annotation, while 2D MLLMs excel at cross-validation by providing complementary information. By integrating holistic 2D and 3D insights from off-the-shelf MLLMs, PiSA-Engine enables a continuous cycle of high-quality data generation. We select PointLLM as the baseline and adopt this co-evolution training framework to develop an enhanced 3D MLLM, termed PointLLM-PiSA. Additionally, we identify limitations in previous 3D benchmarks, which often feature coarse language captions and insufficient category diversity, resulting in inaccurate evaluations. To address this gap, we further introduce PiSA-Bench, a comprehensive 3D benchmark covering six key aspects with detailed and diverse labels. Experimental results demonstrate PointLLM-PiSA’s state-of-the-art performance in zero-shot 3D object captioning and generative classification on our PiSA-Bench, achieving significant improvements of 46.45% (+8.33%) and 63.75% (+16.25%), respectively. Project page: CG-ops/PiSA.

Although existing speech-driven talking face generation methods achieve significant progress, they are far from real-world application due to the avatar-specific training demand and unstable lip movements. To address the above issues, we propose the GSmoothFace, a novel two-stage generalized talking face generation model guided by a fine-grained 3D face model, which can synthesize smooth lip dynamics while preserving the speaker's identity. Our proposed GSmoothFace model mainly consists of the Audio to Expression Prediction (A2EP) module and the Target Adaptive Face Translation (TAFT) module. Specifically, we first develop the A2EP module to predict expression parameters synchronized with the driven speech. It uses a transformer to capture the long-term audio context and learns the parameters from the fine-grained 3D facial vertices, resulting in accurate and smooth lip-synchronization performance. Afterward, the well-designed TAFT module, empowered by Morphology Augmented Face Blending (MAFB), takes the predicted expression parameters and target video as inputs to modify the facial region of the target video without distorting the background content. The TAFT effectively exploits the identity appearance and background context in the target video, which makes it possible to generalize to different speakers without retraining. Both quantitative and qualitative experiments confirm the superiority of our method in terms of realism, lip-synchronization, and visual quality.

The field of 4D point cloud understanding is rapidly developing with the goal of analyzing dynamic 3D point cloud sequences. However, it remains a challenging task due to the sparsity and lack of texture in point clouds. Moreover, the irregularity of point cloud poses a difficulty in aligning temporal information within video sequences. To address these issues, we propose a novel cross-modal knowledge transfer framework, called X4D-SceneFormer. This framework enhances 4D-Scene understanding by transferring texture priors from RGB sequences using a Transformer architecture with temporal relationship mining. Specifically, the framework is designed with a dual-branch architecture, consisting of an 4D point cloud transformer and a Gradient-aware Image Transformer (GIT). The GIT combines visual texture and temporal correlation features to offer rich semantics and dynamics for better point cloud representation. During training, we employ multiple knowledge transfer techniques, including temporal consistency losses and masked self-attention, to strengthen the knowledge transfer between modalities. This leads to enhanced performance during inference using single-modal 4D point cloud inputs. Extensive experiments demonstrate the superior performance of our framework on various 4D point cloud video understanding tasks, including action recognition, action segmentation and semantic segmentation. The results achieve 1st places, i.e., 85.3% (+7.9%) accuracy and 47.3% (+5.0%) mIoU for 4D action segmentation and semantic segmentation, on the HOI4D challenge, outperforming previous state-of-the-art by a large margin. We release the code at https://github.com/jinglinglingling/X4D.

While 3D object bounding box (bbox) representation has been widely used in autonomous driving perception, it lacks the ability to capture the precise details of an object's intrinsic geometry. Recently, occupancy has emerged as a promising alternative for 3D scene perception. However, constructing a high-resolution occupancy map remains infeasible for large scenes due to computational constraints. Recognizing that foreground objects only occupy a small portion of the scene, we introduce object-centric occupancy as a supplement to object bboxes. This representation not only provides intricate details for detected objects but also enables higher voxel resolution in practical applications. We advance the development of object-centric occupancy perception from both data and algorithm perspectives. On the data side, we construct the first object-centric occupancy dataset from scratch using an automated pipeline. From the algorithmic standpoint, we introduce a novel object-centric occupancy completion network equipped with an implicit shape decoder that manages dynamic-size occupancy generation. This network accurately predicts the complete object-centric occupancy volume for inaccurate object proposals by leveraging temporal information from long sequences. Our method demonstrates robust performance in completing object shapes under noisy detection and tracking conditions. Additionally, we show that our occupancy features significantly enhance the detection results of state-of-the-art 3D object detectors, especially for incomplete or distant objects in the Waymo Open Dataset.

3D single object tracking in LiDAR point clouds (LiDAR SOT) plays a crucial role in autonomous driving. Current approaches all follow the Siamese paradigm based on appearance matching. However, LiDAR point clouds are usually textureless and incomplete, which hinders effective appearance matching. Besides, previous methods greatly overlook the critical motion clues among targets. In this work, beyond 3D Siamese tracking, we introduce amotion-centric paradigmto handle LiDAR SOT from a new perspective. Following this paradigm, we propose a matching-free two-stage trackerM$^{2}$2-Track. At the 1st-stage,$M^{2}$-Track localizes the target within successive frames viamotion transformation. Then it refines the target box throughmotion-assisted shape completion at the 2nd-stage. Due to the motion-centric nature, our method shows its impressive generalizability with limited training labels and provides good differentiability for end-to-end cycle training. This inspires us to explore semi-supervised LiDAR SOT by incorporating a pseudo-label-based motion augmentation and a self-supervised loss term. Under the fully-supervised setting, extensive experiments confirm that$M^{2}$-Track significantly outperforms previous state-of-the-arts on three large-scale datasets while running at57FPS($\sim$∼3%,$\sim$∼11%and$\sim$∼22%precision gains on KITTI, NuScenes, and Waymo Open Dataset respectively). While under the semi-supervised setting, our method performs on par with or even surpasses its fully-supervised counterpart using fewer than half labels from KITTI. Further analysis verifies each component's effectiveness and shows the motion-centric paradigm's promising potential for auto-labeling and unsupervised domain adaptation.

Single object tracking (SOT) is one of the most active research directions in the field of computer vision. Compared with the 2-D image-based SOT which has already been well-studied, SOT on 3-D point clouds is a relatively emerging research field. In this article, a novel approach, namely, the contextual-aware tracker (CAT), is investigated to achieve a superior 3-D SOT through spatially and temporally contextual learning from the LiDAR sequence. More precisely, in contrast to the previous 3-D SOT methods merely exploiting point clouds in the target bounding box as the template, CAT generates templates by adaptively including the surroundings outside the target box to use available ambient cues. This template generation strategy is more effective and rational than the previous area-fixed one, especially when the object has only a small number of points. Moreover, it is deduced that LiDAR point clouds in 3-D scenes are often incomplete and significantly vary from frame to another, which makes the learning process more difficult. To this end, a novel cross-frame aggregation (CFA) module is proposed to enhance the feature representation of the template by aggregating the features from a historical reference frame. Leveraging such schemes enables CAT to achieve a robust performance, even in the case of extremely sparse point clouds. The experiments confirm that the proposed CAT outperforms the state-of-the-art methods on both the KITTI and NuScenes benchmarks, achieving 3.9% and 5.6% improvements in terms of precision.

3D lane detection from monocular images is a fundamental yet challenging task in autonomous driving. Recent advances primarily rely on structural 3D surrogates (e.g., bird’s eye view) built from front-view image features and camera parameters. However, the depth ambiguity in monocular images inevitably causes misalignment between the constructed surrogate feature map and the original image, posing a great challenge for accurate lane detection. To address the above issue, we present a novel LATR model, an end-to-end 3D lane detector that uses 3D-aware front-view features without transformed view representation. Specifically, LATR detects 3D lanes via cross-attention based on query and key-value pairs, constructed using our lane-aware query generator and dynamic 3D ground positional embedding. On the one hand, each query is generated based on 2D lane-aware features and adopts a hybrid embedding to enhance the lane information. On the other hand, 3D space information is injected as positional embedding from an iteratively-updated 3D ground plane. LATR outperforms previous state-of-the-art methods on both synthetic Apollo and realistic OpenLane and ONCE-3DLanes by large margins (e.g., 11.4 gain in terms of F1 score on OpenLane). Code will be released at https://github.com/JMoonr/LATR.

Point clouds produced by 3D scanning are typically sparse, non-uniform, and noisy. Existing upsampling techniques directly learn the mapping from a sparse point set to a dense point set, which is often under-determined and ill-posed. To reduce the uncertainty and ambiguity of the upsampling mapping, this paper proposes a generic three-stage vector-quantization framework, which incorporates a Codebook lookup Transformer and knowledge distillation for Point Cloud Upsampling, named CPU. The proposed CPU reformulates the upsampling task into a relatively determinate code prediction task within a small, discrete proxy space. Since the traditional vector-quantization methods cannot be directly applied to point cloud upsampling scenarios, we introduce a knowledge distillation training scheme that facilitates efficient codebook learning and ensures full utilization of codebook entries. Specifically, we adopt a teacher-student training paradigm to avoid model collapse during codebook learning. In the first stage, we pre-train a vanilla auto-encoder of the dense point set as the teacher model, which provides rich guidance features to ensure sufficient codebook learning. In the second stage, we train a vector-quantized auto-encoder as a student model to capture high-fidelity geometric priors into a learned codebook with the aid of distillation. In the third stage, we propose a Codebook Lookup Transformer to model the global context of the sparse point set and predict the code indices. Then the coarse features of the sparse point set can be quantized and substituted by looking up the indices in the learned codebook. Benefiting from the expressive codebook priors and the distillation training scheme, the proposed CPU outperforms state-of-the-art methods quantitatively and qualitatively.

3D single object tracking (3D SOT) in LiDAR point clouds plays a crucial role in autonomous driving. Current approaches all follow the Siamese paradigm based on appearance matching. However, LiDAR point clouds are usually textureless and incomplete, which hinders effective appearance matching. Besides, previous methods greatly overlook the critical motion clues among targets. In this work, beyond 3D Siamese tracking, we introduce a motion-centric paradigm to handle 3D SOT from a new perspective. Following this paradigm, we propose a matching-free two-stage tracker M2-Track. At the 1st-stage, M2-Track localizes the target within successive frames via motion transformation. Then it refines the target box through motion-assisted shape completion at the 2nd-stage. Extensive experiments confirm that M2-Track significantly outperforms previous state-of-the-arts on three large-scale datasets while running at 57FPS (~ 8%, ~ 17% and ~ 22% precision gains on KITTI, NuScenes, and Waymo Open Dataset respectively). Further analysis verifies each component's effectiveness and shows the motioncentric paradigm's promising potential when combined with appearance matching. Code will be made available at https://github.com/Ghostish/Open3DSOT.

Although recent point cloud analysis achieves impressive progress, the paradigm of representation learning from single modality gradually meets its bottleneck. In this work, we take a step towards more discriminative 3D point cloud representation using 2D images, which inherently contain richer appearance information, e.g., texture, color, and shade. Specifically, this paper introduces a simple but effective point cloud cross-modality training (PointCMT) strategy, which utilizes view-images, i.e., rendered or projected 2D images of the 3D object, to boost point cloud classification. In practice, to effectively acquire auxiliary knowledge from view-images, we develop a teacher-student framework and formulate the cross-modal learning as a knowledge distillation problem. Through novel feature and classifier enhancement criteria, PointCMT eliminates the distribution discrepancy between different modalities and avoid potential negative transfer effectively. Note that PointCMT efficiently improves the point-only representation without any architecture modification. Sufficient experiments verify significant gains on various datasets based on several backbones, i.e., equipped with PointCMT, PointNet++ and PointMLP achieve state-of-the-art performance on two benchmarks, i.e., 94.4% and 86.7% accuracy on ModelNet40 and ScanObjectNN, respectively.

Current 3D single object tracking approaches track the target based on a feature comparison between the target template and the search area. However, due to the common occlusion in LiDAR scans, it is non-trivial to conduct accurate feature comparisons on severe sparse and incomplete shapes. In this work, we exploit the ground truth bounding box given in the first frame as a strong cue to enhance the feature description of the target object, enabling a more accurate feature comparison in a simple yet effective way. In particular, we first propose the BoxCloud, an informative and robust representation, to depict an object using the point-to-box relation. We further design an efficient box-aware feature fusion module, which leverages the aforementioned BoxCloud for reliable feature matching and embedding. Integrating the proposed general components into an existing model P2B [27], we construct a superior box-aware tracker (BAT)1. Experiments confirm that our proposed BAT outperforms the previous state-of-the-art by a large margin on both KITTI and NuScenes benchmarks, achieving a 12.8% improvement in terms of precision while running ∼20% faster.

In recent years, multi-view learning has emerged as a promising approach for 3D shape recognition, which identifies a 3D shape based on its 2D views taken from different viewpoints. Usually, the correspondences inside a view or across different views encode the spatial arrangement of object parts and the symmetry of the object, which provide useful geometric cues for recognition. However, such view correspondences have not been explicitly and fully exploited in existing work. In this paper, we propose a correspondence-aware representation (CAR) module, which explicitly finds potential intra-view correspondences and cross-view correspondences via k NN search in semantic space and then aggregates the shape features from the correspondences via learned transforms. Particularly, the spatial relations of correspondences in terms of their viewpoint positions and intra-view locations are taken into account for learning correspondence-aware features. Incorporating the CAR module into a ResNet-18 backbone, we propose an effective deep model called CAR-Net for 3D shape classification and retrieval. Extensive experiments have demonstrated the effectiveness of the CAR module as well as the excellent performance of the CAR-Net.

Raw point clouds data inevitably contains outliers or noise through acquisition from 3D sensors or reconstruction algorithms. In this paper, we present a novel end-to-end network for robust point clouds processing, named PointASNL, which can deal with point clouds with noise effectively. The key component in our approach is the adaptive sampling (AS) module. It first re-weights the neighbors around the initial sampled points from farthest point sampling (FPS), and then adaptively adjusts the sampled points beyond the entire point cloud. Our AS module can not only benefit the feature learning of point clouds, but also ease the biased effect of outliers. To further capture the neighbor and long-range dependencies of the sampled point, we proposed a local-nonlocal (L-NL) module inspired by the nonlocal operation. Such L-NL module enables the learning process insensitive to noise. Extensive experiments verify the robustness and superiority of our approach in point clouds processing tasks regardless of synthesis data, indoor data, and outdoor data with or without noise. Specifically, PointASNL achieves state-of-the-art robust performance for classification and segmentation tasks on all datasets, and significantly outperforms previous methods on real-world outdoor SemanticKITTI dataset with considerate noise.

3D shape recognition via deep learning is drawing more and more attention due to huge industry interests. As 3D deep learning methods emerged, the view-based approaches have gained considerable success in object classification. Most of these methods focus on designing a pooling scheme to aggregate CNN features of multi-view images into a single compact one. However, these view-wise pooling techniques suffer from loss of visual information. To deal with this issue, an adaptive rank pooling layer is introduced in this paper. Unlike max-pooling which only considers the maximum or mean-pooling that treats each element indiscriminately, the proposed pooling layer takes all the elements into account and dynamically adjusts their importances during the training. Experiments conducted on ModelNet40 and ModelNet10 shows both efficiency and accuracy gain when inserting such a layer into a baseline CNN architecture.