3D semantic occupancy prediction offers a nuanced representation of the surrounding environment, which is crucial for ensuring the safety of autonomous driving. However, fine-grained scene representations inevitably result in cubic growth in data scale, which imposes substantial demands on model architecture and computational complexity, especially in high-resolution scenarios. Existing approaches for handling high-resolution scenes typically obtain fine-grained features by grid sampling on low-resolution feature map, resulting in limited sparsity and insufficient feature interaction. This paper presents a framework leveraging SParse representation and SCalable feature interaction to address the aforementioned challenges, called SPSC. Specifically, we maintain sparsity by progressively pruning unoccupied queries during the coarse-to-fine process, thereby reducing the scale of data that the model needs to handle. Subsequently, we introduce query serialization, which transforms queries into an ordered sequence while preserving their spatial structure, This enables fine-grained feature interaction while maintaining linear computational complexity and a larger receptive field. Without complex architectural designs, SPSC significantly outperforms SOTA approaches, relatively enhances the mIoU by 12.0%, 11.0% and 4.8% on nuScenes-Occupancy dataset under the muli-modal, LiDAR and camera settings, respectively.

The costly annotation process has driven the development of semi-supervised learning (SSL) approaches. Existing semi-supervised LiDAR segmentation methods typically process entire point clouds directly, aiming to assign labels to all points at the scene scale. However, the large number of points, combined with their sparse and irregular nature, makes it challenging to learn scene-level optimization objectives, especially in SSL settings where labeled data are insufficient. This paper presents a Cluster-based pseudo-LAbeling Semi-Supervised technique, called CLASS. CLASS is designed to divide point clouds into several small, pure clusters, thereby decomposing challenging scene-scale segmentation task into more manageable cluster-scale classification and segmentation tasks, enabling the generation of high-quality pseudo labels for unlabeled data. CLASS possesses three key properties. i) Task simplicity: our pseudo-labeling process is based on simpler cluster-scale classification and segmentation tasks, resulting in ease of learning. ii) Labeling effectiveness: CLASS can generate pseudo-labels comparable to ground truth using only approximately 10% labeled data. iii) Universal versatility: CLASS exhibits flexibility regarding LiDAR representations (e.g., BEV, voxel, and range view). Comprehensive experiments on popular LiDAR segmentation benchmarks demonstrate its superiority.

The Global Positioning System (GPS) provides precise geographic locations for our vehicles. Nevertheless, it is frequently subject to disruptions, potentially resulting in incomplete or absent trajectory data. To address this challenge, we present DictRoadNet, a framework designed for GPS trajectory completion, which uses a clustering-based dictionary module for initial trajectory generation and a road network module for refining results based on road network data. First, we introduce a dictionary that employs a clustering-based strategy for selecting key-value pairs, which can be used in GPS data processing. This dictionary can provide auxiliary general information acquired from trajectory clusters, enhancing the generation of rational trajectories with additional details. Second, we propose a Road Network Module that utilizes a directed graph to store road network information derived from historical GPS trajectories. This module refines the output by aligning it with an empirically constructed road network, ensuring that trajectory completions are plausible and closely adhere to actual road paths. We achieved enhancements across all tasks when assessed against Average and Final Displacement Error, with the highest enhancement reaching up to 9.50% compared to state-of-the-art methods.

In recent years, graph-partition-based clustering algorithms have attracted increasing attention. These algorithms first construct a graph over the data points and then partition this graph, regarding each connected subgraph in the partitioned graph as a cluster. However, traditional graph-partition-based clustering algorithms face challenges when clustering datasets with highly imbalanced density distributions. This is because, during the process of graph partitioning, they mainly rely on edge lengths and largely ignore local density variations. To address this issue, we propose the Adaptive Graph Partitioning (AGP) clustering algorithm. AGP integrates local density information into the partitioning process and adaptively normalizes the magnitudes of edge weights in both sparse and dense regions. This enhancement allows AGP to avoid the over-partitioning of dense clusters, effectively addressing a common problem in traditional graph-partition-based clustering algorithms. Additionally, the mechanism of connectivity domain differences is introduced into AGP, further enhancing the algorithm’s ability to discriminate between neighboring clusters that are difficult to separate. Extensive experiments on 13 benchmark datasets show that AGP achieves the best clustering performance on 7 datasets and performs competitively on the others, especially when density imbalance is severe. Moreover, scalability experiments on datasets with up to 100,000 data points, together with complexity analysis, demonstrate that AGP enjoys favorable time and memory efficiency compared with representative baselines.

Recently, more and more images are compressed and sent to the back-end devices for machine analysis tasks (e.g., object detection) instead of being purely watched by humans. However, most traditional or learned image codecs are designed to minimize the distortion of the human visual system without considering the increased demand from machine vision systems. In this work, we propose a preprocessing enhanced image compression method for machine vision tasks to address this challenge. Instead of relying on the learned image codecs for end-to-end optimization, our framework is built upon the traditional non-differential codecs, which means it is standard compatible and can be easily deployed in practical applications. Specifically, we propose a neural preprocessing module before the encoder to maintain the useful semantic information for the downstream tasks and suppress the irrelevant information for bitrate saving. Furthermore, our neural preprocessing module is quantization adaptive and can be used in different compression ratios. More importantly, to jointly optimize the preprocessing module with the downstream machine vision tasks, we introduce the proxy network for the traditional non-differential codecs in the back-propagation stage. We provide extensive experiments by evaluating our compression method for several representative downstream tasks with different backbone networks. Experimental results show our method achieves a better trade-off between the coding bitrate and the performance of the downstream machine vision tasks by saving about 20% bitrate.

Discovering causal relationships among observed variables is an important research focus in data mining. Existing causal discovery approaches are mainly based on constraint-based methods and functional causal models (FCMs). However, the constraint-based method cannot identify the Markov equivalence class and the functional causal models cannot identify the complex interrelationships when multiple variables affect one variable. To address the two aforementioned problems, we propose a new graph structure Causal Star Graph (CSG) and a corresponding framework Causal Discovery via Causal Star Graphs (CD-CSG) to divide a causal directed acyclic graph into multiple CSGs for causal discovery. In this framework, we also propose a generalized learning in CSGs based on a variational approach to learn the representative intermediate variable of CSG’s non-central variables. Through the generalized learning in CSGs, the asymmetry in the forward and backward model of CD-CSG can be found to identify the causal directions in the directed acyclic graphs. We further divide the CSGs into three categories and provide the causal identification principle under each category in our proposed framework. Experiments using synthetic data show that the causal relationships between variables can be effectively identified with CD-CSG and the accuracy of CD-CSG is higher than the best existing model. By applying CD-CSG to real-world data, our proposed method can greatly augment the applicability and effectiveness of causal discovery.

Discovering causal relationships among observed variables is a new research focus in the area of data mining. Methods based on the additive noise model have been proved to be efficient in the identification of cause-effect pairs. However, when trying to determine many-to-one causality, additive noise models often fail to identify the causal direction due to the complex interrelationships and interactions even though the generation of each causal relation follows the additive noise model, and become unreliable in practical applications. In this work, to identify the causal direction, we propose a Hierarchical Additive Noise Model (HANM) to convert many-to-one causality into an approximate one-to-one causality by generalizing multiple factors into an intermediate variable with a variational approach, and use asymmetry in the forward model and backward model of HANM to identify causal direction. Experiments using synthetic data show that many-to-one causality can be effectively identified through asymmetry with our proposed HANM and the accuracy of HANM is higher than the best existing model. By applying the model to real-world data, it can be seen that HANM can greatly augment the application scope of functional causal models for causal discovery.

Multi-modality (i.e., multi-sensor) data is widely used in various vision tasks for more accurate or robust perception. However, the increased data modalities bring new challenges for data storage and transmission. The existing data compression approaches usually adopt individual codecs for each modality without considering the correlation between different modalities. This work proposes a multi-modality compression framework for infrared and visible image pairs by exploiting the cross-modality redun-dancy. Specifically, given the image in the reference modality (e.g., the infrared image), we use the channel-wise alignment module to produce the aligned features based on the affine transform. Then the aligned feature is used as the context information for compressing the image in the current modality (e.g., the visible image), and the corresponding affine coefficients are losslessly compressed at negligible cost. Furthermore, we introduce the Transformer-based spatial alignment module to exploit the correlation between the intermediate features in the decoding procedures for different modalities. Our framework is very flexible and easily extended for multi-modality video compression. Experimental results show our proposed framework outperforms the traditional and learning-based single modality compression methods on the FLIR and KAIST datasets.