We demonstrate the feasibility of integrating physics-based animations of solids and fluids with 3D Gaussian Splatting (3DGS) to create novel effects in virtual scenes reconstructed using 3DGS. Leveraging the coherence of the Gaussian splatting and position-based dynamics (PBD) in the underlying representation, we manage rendering, view synthesis, and the dynamics of solids and fluids in a cohesive manner. Similar to Gaussian shader, we enhance each Gaussian kernel with an added normal, aligning the kernel's orientation with the surface normal to refine the PBD simulation. This approach effectively eliminates spiky noises that arise from rotational deformation in solids. It also allows us to integrate physically based rendering to augment the dynamic surface reflections on fluids. Consequently, our framework is capable of realistically reproducing surface highlights on dynamic fluids and facilitating interactions between scene objects and fluids from new views. For more information, please visit our project page at \url{https://amysteriouscat.github.io/GaussianSplashing/}.
Vision-language pretrained models have seen remarkable success, but their application to safety-critical settings is limited by their lack of interpretability. To improve the interpretability of vision-language models such as CLIP, we propose a multi-modal information bottleneck (M2IB) approach that learns latent representations that compress irrelevant information while preserving relevant visual and textual features. We demonstrate how M2IB can be applied to attribution analysis of vision-language pretrained models, increasing attribution accuracy and improving the interpretability of such models when applied to safety-critical domains such as healthcare. Crucially, unlike commonly used unimodal attribution methods, M2IB does not require ground truth labels, making it possible to audit representations of vision-language pretrained models when multiple modalities but no ground-truth data is available. Using CLIP as an example, we demonstrate the effectiveness of M2IB attribution and show that it outperforms gradient-based, perturbation-based, and attention-based attribution methods both qualitatively and quantitatively.
Rich textual and topological information of textual graphs need to be modeled in real-world applications such as webpages, e-commerce, and academic articles. Practitioners have been long following the path of adopting a shallow text encoder and a subsequent graph neural network (GNN) to solve this problem. In light of recent advancements in large language models (LLMs), it is apparent that integrating LLMs for enhanced textual encoding can substantially improve the performance of textual graphs. Nevertheless, the efficiency of these methods poses a significant challenge. In this paper, we propose ENGINE, a parameter- and memory-efficient fine-tuning method for textual graphs with an LLM encoder. The key insight is to combine the LLMs and GNNs through a tunable side structure, which significantly reduces the training complexity without impairing the joint model's capacity. Extensive experiments on textual graphs demonstrate our method's effectiveness by achieving the best model performance, meanwhile having the lowest training cost compared to previous methods. Moreover, we introduce two variants with caching and dynamic early exit to further enhance training and inference speed. Specifically, caching accelerates ENGINE's training by 12x, and dynamic early exit achieves up to 5x faster inference with a negligible performance drop (at maximum 1.17% relevant drop across 7 datasets).
Holistic understanding and reasoning in 3D scenes play a vital role in the success of autonomous driving systems. The evolution of 3D semantic occupancy prediction as a pretraining task for autonomous driving and robotic downstream tasks captures finer 3D details compared to methods like 3D detection. Existing approaches predominantly focus on spatial cues, often overlooking temporal cues. Query-based methods tend to converge on computationally intensive Voxel representation for encoding 3D scene information. This study introduces S2TPVFormer, an extension of TPVFormer, utilizing a spatiotemporal transformer architecture for coherent 3D semantic occupancy prediction. Emphasizing the importance of spatiotemporal cues in 3D scene perception, particularly in 3D semantic occupancy prediction, our work explores the less-explored realm of temporal cues. Leveraging Tri-Perspective View (TPV) representation, our spatiotemporal encoder generates temporally rich embeddings, improving prediction coherence while maintaining computational efficiency. To achieve this, we propose a novel Temporal Cross-View Hybrid Attention (TCVHA) mechanism, facilitating effective spatiotemporal information exchange across TPV views. Experimental evaluations on the nuScenes dataset demonstrate a substantial 3.1% improvement in mean Intersection over Union (mIoU) for 3D Semantic Occupancy compared to TPVFormer, confirming the effectiveness of the proposed S2TPVFormer in enhancing 3D scene perception.
We present a novel vision-based control method to make a group of ground mobile robots achieve a specified formation shape with unspecified size. Our approach uses multiple aerial control units equipped with downward-facing cameras, each observing a partial subset of the multirobot team. The units compute the control commands from the ground robots' image projections, using neither calibration nor scene scale information, and transmit them to the robots. The control strategy relies on the calculation of image similarity transformations, and we show it to be asymptotically stable if the overlaps between the subsets of controlled robots satisfy certain conditions. The presence of the supervisory units, which coordinate their motions to guarantee a correct control performance, gives rise to a hybrid system topology. All in all, the proposed system provides relevant practical advantages in simplicity and flexibility. Within the problem of controlling a team shape, our contribution lies in addressing several simultaneous challenges: the controller needs only partial information of the robotic group, does not use distance measurements or global reference frames, is designed for unicycle agents, and can accommodate topology changes. We present illustrative simulation results.
Smartwatch health sensor data is increasingly utilized in smart health applications and patient monitoring, including stress detection. However, such medical data often comprises sensitive personal information and is resource-intensive to acquire for research purposes. In response to this challenge, we introduce the privacy-aware synthetization of multi-sensor smartwatch health readings related to moments of stress. Our method involves the generation of synthetic sequence data through Generative Adversarial Networks (GANs), coupled with the implementation of Differential Privacy (DP) safeguards for protecting patient information during model training. To ensure the integrity of our synthetic data, we employ a range of quality assessments and monitor the plausibility between synthetic and original data. To test the usefulness, we create private machine learning models on a commonly used, albeit small, stress detection dataset, exploring strategies for enhancing the existing data foundation with our synthetic data. Through our GAN-based augmentation methods, we observe improvements in model performance, both in non-private (0.45% F1) and private (11.90-15.48% F1) training scenarios. We underline the potential of differentially private synthetic data in optimizing utility-privacy trade-offs, especially with limited availability of real training samples.
The enhancement of 3D object detection is pivotal for precise environmental perception and improved task execution capabilities in autonomous driving. LiDAR point clouds, offering accurate depth information, serve as a crucial information for this purpose. Our study focuses on key challenges in 3D target detection. To tackle the challenge of expanding the receptive field of a 3D convolutional kernel, we introduce the Dynamic Feature Fusion Module (DFFM). This module achieves adaptive expansion of the 3D convolutional kernel's receptive field, balancing the expansion with acceptable computational loads. This innovation reduces operations, expands the receptive field, and allows the model to dynamically adjust to different object requirements. Simultaneously, we identify redundant information in 3D features. Employing the Feature Selection Module (FSM) quantitatively evaluates and eliminates non-important features, achieving the separation of output box fitting and feature extraction. This innovation enables the detector to focus on critical features, resulting in model compression, reduced computational burden, and minimized candidate frame interference. Extensive experiments confirm that both DFFM and FSM not only enhance current benchmarks, particularly in small target detection, but also accelerate network performance. Importantly, these modules exhibit effective complementarity.
Property prediction is a fundamental task in crystal material research. To model atoms and structures, structures represented as graphs are widely used and graph learning-based methods have achieved significant progress. Bond angles and bond distances are two key structural information that greatly influence crystal properties. However, most of the existing works only consider bond distances and overlook bond angles. The main challenge lies in the time cost of handling bond angles, which leads to a significant increase in inference time. To solve this issue, we first propose a crystal structure modeling based on dual scale neighbor partitioning mechanism, which uses a larger scale cutoff for edge neighbors and a smaller scale cutoff for angle neighbors. Then, we propose a novel Atom-Distance-Angle Graph Neural Network (ADA-GNN) for property prediction tasks, which can process node information and structural information separately. The accuracy of predictions and inference time are improved with the dual scale modeling and the specially designed architecture of ADA-GNN. The experimental results validate that our approach achieves state-of-the-art results in two large-scale material benchmark datasets on property prediction tasks.
Semi-supervised domain adaptation (SSDA) aims to bridge source and target domain distributions, with a small number of target labels available, achieving better classification performance than unsupervised domain adaptation (UDA). However, existing SSDA work fails to make full use of label information from both source and target domains for feature alignment across domains, resulting in label mismatch in the label space during model testing. This paper presents a novel SSDA approach, Inter-domain Mixup with Neighborhood Expansion (IDMNE), to tackle this issue. Firstly, we introduce a cross-domain feature alignment strategy, Inter-domain Mixup, that incorporates label information into model adaptation. Specifically, we employ sample-level and manifold-level data mixing to generate compatible training samples. These newly established samples, combined with reliable and actual label information, display diversity and compatibility across domains, while such extra supervision thus facilitates cross-domain feature alignment and mitigates label mismatch. Additionally, we utilize Neighborhood Expansion to leverage high-confidence pseudo-labeled samples in the target domain, diversifying the label information of the target domain and thereby further increasing the performance of the adaptation model. Accordingly, the proposed approach outperforms existing state-of-the-art methods, achieving significant accuracy improvements on popular SSDA benchmarks, including DomainNet, Office-Home, and Office-31.
Information extraction (IE) aims to extract structural knowledge (such as entities, relations, and events) from plain natural language texts. Recently, generative Large Language Models (LLMs) have demonstrated remarkable capabilities in text understanding and generation, allowing for generalization across various domains and tasks. As a result, numerous works have been proposed to harness abilities of LLMs and offer viable solutions for IE tasks based on a generative paradigm. To conduct a comprehensive systematic review and exploration of LLM efforts for IE tasks, in this study, we survey the most recent advancements in this field. We first present an extensive overview by categorizing these works in terms of various IE subtasks and learning paradigms, then we empirically analyze the most advanced methods and discover the emerging trend of IE tasks with LLMs. Based on thorough review conducted, we identify several insights in technique and promising research directions that deserve further exploration in future studies. We maintain a public repository and consistently update related resources at: \url{https://github.com/quqxui/Awesome-LLM4IE-Papers}.