This paper investigates the problem of understanding dynamic 3D scenes from egocentric observations, a key challenge in robotics and embodied AI. Unlike prior studies that explored this as long-form video understanding and utilized egocentric video only, we instead propose an LLM-based agent, Embodied VideoAgent, which constructs scene memory from both egocentric video and embodied sensory inputs (e.g. depth and pose sensing). We further introduce a VLM-based approach to automatically update the memory when actions or activities over objects are perceived. Embodied VideoAgent attains significant advantages over counterparts in challenging reasoning and planning tasks in 3D scenes, achieving gains of 4.9% on Ego4D-VQ3D, 5.8% on OpenEQA, and 11.7% on EnvQA. We have also demonstrated its potential in various embodied AI tasks including generating embodied interactions and perception for robot manipulation. The code and demo will be made public.
We present STORM, a spatio-temporal reconstruction model designed for reconstructing dynamic outdoor scenes from sparse observations. Existing dynamic reconstruction methods often rely on per-scene optimization, dense observations across space and time, and strong motion supervision, resulting in lengthy optimization times, limited generalization to novel views or scenes, and degenerated quality caused by noisy pseudo-labels for dynamics. To address these challenges, STORM leverages a data-driven Transformer architecture that directly infers dynamic 3D scene representations--parameterized by 3D Gaussians and their velocities--in a single forward pass. Our key design is to aggregate 3D Gaussians from all frames using self-supervised scene flows, transforming them to the target timestep to enable complete (i.e., "amodal") reconstructions from arbitrary viewpoints at any moment in time. As an emergent property, STORM automatically captures dynamic instances and generates high-quality masks using only reconstruction losses. Extensive experiments on public datasets show that STORM achieves precise dynamic scene reconstruction, surpassing state-of-the-art per-scene optimization methods (+4.3 to 6.6 PSNR) and existing feed-forward approaches (+2.1 to 4.7 PSNR) in dynamic regions. STORM reconstructs large-scale outdoor scenes in 200ms, supports real-time rendering, and outperforms competitors in scene flow estimation, improving 3D EPE by 0.422m and Acc5 by 28.02%. Beyond reconstruction, we showcase four additional applications of our model, illustrating the potential of self-supervised learning for broader dynamic scene understanding.
Open-vocabulary scene understanding using 3D Gaussian (3DGS) representations has garnered considerable attention. However, existing methods mostly lift knowledge from large 2D vision models into 3DGS on a scene-by-scene basis, restricting the capabilities of open-vocabulary querying within their training scenes so that lacking the generalizability to novel scenes. In this work, we propose \textbf{OVGaussian}, a generalizable \textbf{O}pen-\textbf{V}ocabulary 3D semantic segmentation framework based on the 3D \textbf{Gaussian} representation. We first construct a large-scale 3D scene dataset based on 3DGS, dubbed \textbf{SegGaussian}, which provides detailed semantic and instance annotations for both Gaussian points and multi-view images. To promote semantic generalization across scenes, we introduce Generalizable Semantic Rasterization (GSR), which leverages a 3D neural network to learn and predict the semantic property for each 3D Gaussian point, where the semantic property can be rendered as multi-view consistent 2D semantic maps. In the next, we propose a Cross-modal Consistency Learning (CCL) framework that utilizes open-vocabulary annotations of 2D images and 3D Gaussians within SegGaussian to train the 3D neural network capable of open-vocabulary semantic segmentation across Gaussian-based 3D scenes. Experimental results demonstrate that OVGaussian significantly outperforms baseline methods, exhibiting robust cross-scene, cross-domain, and novel-view generalization capabilities. Code and the SegGaussian dataset will be released. (https://github.com/runnanchen/OVGaussian).
Generating desired images conditioned on given text descriptions has received lots of attention. Recently, diffusion models and autoregressive models have demonstrated their outstanding expressivity and gradually replaced GAN as the favored architectures for text-to-image synthesis. However, they still face some obstacles: slow inference speed and expensive training costs. To achieve more powerful and faster text-to-image synthesis under complex scenes, we propose TIGER, a text-to-image GAN with pretrained representations. To be specific, we propose a vision-empowered discriminator and a high-capacity generator. (i) The vision-empowered discriminator absorbs the complex scene understanding ability and the domain generalization ability from pretrained vision models to enhance model performance. Unlike previous works, we explore stacking multiple pretrained models in our discriminator to collect multiple different representations. (ii) The high-capacity generator aims to achieve effective text-image fusion while increasing the model capacity. The high-capacity generator consists of multiple novel high-capacity fusion blocks (HFBlock). And the HFBlock contains several deep fusion modules and a global fusion module, which play different roles to benefit our model. Extensive experiments demonstrate the outstanding performance of our proposed TIGER both on standard and zero-shot text-to-image synthesis tasks. On the standard text-to-image synthesis task, TIGER achieves state-of-the-art performance on two challenging datasets, which obtain a new FID 5.48 (COCO) and 9.38 (CUB). On the zero-shot text-to-image synthesis task, we achieve comparable performance with fewer model parameters, smaller training data size and faster inference speed. Additionally, more experiments and analyses are conducted in the Supplementary Material.
Dynamic 3D scene representation and novel view synthesis from captured videos are crucial for enabling immersive experiences required by AR/VR and metaverse applications. However, this task is challenging due to the complexity of unconstrained real-world scenes and their temporal dynamics. In this paper, we frame dynamic scenes as a spatio-temporal 4D volume learning problem, offering a native explicit reformulation with minimal assumptions about motion, which serves as a versatile dynamic scene learning framework. Specifically, we represent a target dynamic scene using a collection of 4D Gaussian primitives with explicit geometry and appearance features, dubbed as 4D Gaussian splatting (4DGS). This approach can capture relevant information in space and time by fitting the underlying spatio-temporal volume. Modeling the spacetime as a whole with 4D Gaussians parameterized by anisotropic ellipses that can rotate arbitrarily in space and time, our model can naturally learn view-dependent and time-evolved appearance with 4D spherindrical harmonics. Notably, our 4DGS model is the first solution that supports real-time rendering of high-resolution, photorealistic novel views for complex dynamic scenes. To enhance efficiency, we derive several compact variants that effectively reduce memory footprint and mitigate the risk of overfitting. Extensive experiments validate the superiority of 4DGS in terms of visual quality and efficiency across a range of dynamic scene-related tasks (e.g., novel view synthesis, 4D generation, scene understanding) and scenarios (e.g., single object, indoor scenes, driving environments, synthetic and real data).
Multimodal large language models (MLLMs) have shown satisfactory effects in many autonomous driving tasks. In this paper, MLLMs are utilized to solve joint semantic scene understanding and risk localization tasks, while only relying on front-view images. In the proposed MLLM-SUL framework, a dual-branch visual encoder is first designed to extract features from two resolutions, and rich visual information is conducive to the language model describing risk objects of different sizes accurately. Then for the language generation, LLaMA model is fine-tuned to predict scene descriptions, containing the type of driving scenario, actions of risk objects, and driving intentions and suggestions of ego-vehicle. Ultimately, a transformer-based network incorporating a regression token is trained to locate the risk objects. Extensive experiments on the existing DRAMA-ROLISP dataset and the extended DRAMA-SRIS dataset demonstrate that our method is efficient, surpassing many state-of-the-art image-based and video-based methods. Specifically, our method achieves 80.1% BLEU-1 score and 298.5% CIDEr score in the scene understanding task, and 59.6% accuracy in the localization task. Codes and datasets are available at https://github.com/fjq-tongji/MLLM-SUL.
A 3D scene graph represents a compact scene model, storing information about the objects and the semantic relationships between them, making its use promising for robotic tasks. When interacting with a user, an embodied intelligent agent should be capable of responding to various queries about the scene formulated in natural language. Large Language Models (LLMs) are beneficial solutions for user-robot interaction due to their natural language understanding and reasoning abilities. Recent methods for creating learnable representations of 3D scenes have demonstrated the potential to improve the quality of LLMs responses by adapting to the 3D world. However, the existing methods do not explicitly utilize information about the semantic relationships between objects, limiting themselves to information about their coordinates. In this work, we propose a method 3DGraphLLM for constructing a learnable representation of a 3D scene graph. The learnable representation is used as input for LLMs to perform 3D vision-language tasks. In our experiments on popular ScanRefer, RIORefer, Multi3DRefer, ScanQA, Sqa3D, and Scan2cap datasets, we demonstrate the advantage of this approach over baseline methods that do not use information about the semantic relationships between objects. The code is publicly available at https://github.com/CognitiveAISystems/3DGraphLLM.
Autonomous driving in high-speed racing, as opposed to urban environments, presents significant challenges in scene understanding due to rapid changes in the track environment. Traditional sequential network approaches may struggle to meet the real-time knowledge and decision-making demands of an autonomous agent covering large displacements in a short time. This paper proposes a novel baseline architecture for developing sophisticated models capable of true hardware-enabled parallelism, achieving neural processing speeds that mirror the agent's high velocity. The proposed model (Parallel Perception Network (PPN)) consists of two independent neural networks, segmentation and reconstruction networks, running parallelly on separate accelerated hardware. The model takes raw 3D point cloud data from the LiDAR sensor as input and converts it into a 2D Bird's Eye View Map on both devices. Each network independently extracts its input features along space and time dimensions and produces outputs parallelly. The proposed method's model is trained on a system with two NVIDIA T4 GPUs, using a combination of loss functions, including edge preservation, and demonstrates a 2x speedup in model inference time compared to a sequential configuration. Implementation is available at: https://github.com/suwesh/Parallel-Perception-Network. Learned parameters of the trained networks are provided at: https://huggingface.co/suwesh/ParallelPerceptionNetwork.
We present UniPLV, a powerful framework that unifies point clouds, images and text in a single learning paradigm for open-world 3D scene understanding. UniPLV employs the image modal as a bridge to co-embed 3D points with pre-aligned images and text in a shared feature space without requiring carefully crafted point cloud text pairs. To accomplish multi-modal alignment, we propose two key strategies:(i) logit and feature distillation modules between images and point clouds, and (ii) a vison-point matching module is given to explicitly correct the misalignment caused by points to pixels projection. To further improve the performance of our unified framework, we adopt four task-specific losses and a two-stage training strategy. Extensive experiments show that our method outperforms the state-of-the-art methods by an average of 15.6% and 14.8% for semantic segmentation over Base-Annotated and Annotation-Free tasks, respectively. The code will be released later.
In this era of technological advancements, several cutting-edge techniques are being implemented to enhance Autonomous Driving (AD) systems, focusing on improving safety, efficiency, and adaptability in complex driving environments. However, AD still faces some problems including performance limitations. To address this problem, we conducted an in-depth study on implementing the Multi-modal Large Language Model. We constructed a Virtual Question Answering (VQA) dataset to fine-tune the model and address problems with the poor performance of MLLM on AD. We then break down the AD decision-making process by scene understanding, prediction, and decision-making. Chain of Thought has been used to make the decision more perfectly. Our experiments and detailed analysis of Autonomous Driving give an idea of how important MLLM is for AD.
Applying Gaussian Splatting to perception tasks for 3D scene understanding is becoming increasingly popular. Most existing works primarily focus on rendering 2D feature maps from novel viewpoints, which leads to an imprecise 3D language field with outlier languages, ultimately failing to align objects in 3D space. By utilizing masked images for feature extraction, these approaches also lack essential contextual information, leading to inaccurate feature representation. To this end, we propose a Language-Embedded Surface Field (LangSurf), which accurately aligns the 3D language fields with the surface of objects, facilitating precise 2D and 3D segmentation with text query, widely expanding the downstream tasks such as removal and editing. The core of LangSurf is a joint training strategy that flattens the language Gaussian on the object surfaces using geometry supervision and contrastive losses to assign accurate language features to the Gaussians of objects. In addition, we also introduce the Hierarchical-Context Awareness Module to extract features at the image level for contextual information then perform hierarchical mask pooling using masks segmented by SAM to obtain fine-grained language features in different hierarchies. Extensive experiments on open-vocabulary 2D and 3D semantic segmentation demonstrate that LangSurf outperforms the previous state-of-the-art method LangSplat by a large margin. As shown in Fig. 1, our method is capable of segmenting objects in 3D space, thus boosting the effectiveness of our approach in instance recognition, removal, and editing, which is also supported by comprehensive experiments. \url{https://langsurf.github.io}.
Monitoring animal populations is crucial for assessing the health of ecosystems. Traditional methods, which require extensive fieldwork, are increasingly being supplemented by time-lapse camera-trap imagery combined with an automatic analysis of the image data. The latter usually involves some object detector aimed at detecting relevant targets (commonly animals) in each image, followed by some postprocessing to gather activity and population data. In this paper, we show that the performance of an object detector in a single frame of a time-lapse sequence can be improved by including spatio-temporal features from the prior frames. We propose a method that leverages temporal information by integrating two additional spatial feature channels which capture stationary and non-stationary elements of the scene and consequently improve scene understanding and reduce the number of stationary false positives. The proposed technique achieves a significant improvement of 24\% in mean average precision ([email omitted]:0.95) over the baseline (temporal feature-free, single frame) object detector on a large dataset of breeding tropical seabirds. We envisage our method will be widely applicable to other wildlife monitoring applications that use time-lapse imaging.
3D scene understanding is an important task, and there has been a recent surge of research interest in aligning 3D representations of point clouds with text to empower embodied AI. However, due to the lack of comprehensive 3D benchmarks, the capabilities of 3D models in real-world scenes, particularly those that are challenging with subtly distinguished objects, remain insufficiently investigated. To facilitate a more thorough evaluation of 3D models' capabilities, we propose a scheme, ObjVariantEnsemble, to systematically introduce more scenes with specified object classes, colors, shapes, quantities, and spatial relationships to meet model evaluation needs. More importantly, we intentionally construct scenes with similar objects to a certain degree and design an LLM-VLM-cooperated annotator to capture key distinctions as annotations. The resultant benchmark can better challenge 3D models, reveal their shortcomings in understanding, and potentially aid in the further development of 3D models.
Although multiview fusion has demonstrated potential in LiDAR segmentation, its dependence on computationally intensive point-based interactions, arising from the lack of fixed correspondences between views such as range view and Bird's-Eye View (BEV), hinders its practical deployment. This paper challenges the prevailing notion that multiview fusion is essential for achieving high performance. We demonstrate that significant gains can be realized by directly fusing Polar and Cartesian partitioning strategies within the BEV space. Our proposed BEV-only segmentation model leverages the inherent fixed grid correspondences between these partitioning schemes, enabling a fusion process that is orders of magnitude faster (170$\times$ speedup) than conventional point-based methods. Furthermore, our approach facilitates dense feature fusion, preserving richer contextual information compared to sparse point-based alternatives. To enhance scene understanding while maintaining inference efficiency, we also introduce a hybrid Transformer-CNN architecture. Extensive evaluation on the SemanticKITTI and nuScenes datasets provides compelling evidence that our method outperforms previous multiview fusion approaches in terms of both performance and inference speed, highlighting the potential of BEV-based fusion for LiDAR segmentation. Code is available at \url{https://github.com/skyshoumeng/PC-BEV.}
Recent advancements in large vision language models (VLMs) tailored for autonomous driving (AD) have shown strong scene understanding and reasoning capabilities, making them undeniable candidates for end-to-end driving systems. However, limited work exists on studying the trustworthiness of DriveVLMs -- a critical factor that directly impacts public transportation safety. In this paper, we introduce AutoTrust, a comprehensive trustworthiness benchmark for large vision-language models in autonomous driving (DriveVLMs), considering diverse perspectives -- including trustfulness, safety, robustness, privacy, and fairness. We constructed the largest visual question-answering dataset for investigating trustworthiness issues in driving scenarios, comprising over 10k unique scenes and 18k queries. We evaluated six publicly available VLMs, spanning from generalist to specialist, from open-source to commercial models. Our exhaustive evaluations have unveiled previously undiscovered vulnerabilities of DriveVLMs to trustworthiness threats. Specifically, we found that the general VLMs like LLaVA-v1.6 and GPT-4o-mini surprisingly outperform specialized models fine-tuned for driving in terms of overall trustworthiness. DriveVLMs like DriveLM-Agent are particularly vulnerable to disclosing sensitive information. Additionally, both generalist and specialist VLMs remain susceptible to adversarial attacks and struggle to ensure unbiased decision-making across diverse environments and populations. Our findings call for immediate and decisive action to address the trustworthiness of DriveVLMs -- an issue of critical importance to public safety and the welfare of all citizens relying on autonomous transportation systems. Our benchmark is publicly available at \url{https://github.com/taco-group/AutoTrust}, and the leaderboard is released at \url{https://taco-group.github.io/AutoTrust/}.
3D open-vocabulary scene understanding, which accurately perceives complex semantic properties of objects in space, has gained significant attention in recent years. In this paper, we propose GAGS, a framework that distills 2D CLIP features into 3D Gaussian splatting, enabling open-vocabulary queries for renderings on arbitrary viewpoints. The main challenge of distilling 2D features for 3D fields lies in the multiview inconsistency of extracted 2D features, which provides unstable supervision for the 3D feature field. GAGS addresses this challenge with two novel strategies. First, GAGS associates the prompt point density of SAM with the camera distances, which significantly improves the multiview consistency of segmentation results. Second, GAGS further decodes a granularity factor to guide the distillation process and this granularity factor can be learned in a unsupervised manner to only select the multiview consistent 2D features in the distillation process. Experimental results on two datasets demonstrate significant performance and stability improvements of GAGS in visual grounding and semantic segmentation, with an inference speed 2$\times$ faster than baseline methods. The code and additional results are available at https://pz0826.github.io/GAGS-Webpage/ .
Neural radiance fields are an emerging 3D scene representation and recently even been extended to learn features for scene understanding by distilling open-vocabulary features from vision-language models. However, current method primarily focus on object-centric representations, supporting object segmentation or detection, while understanding semantic relationships between objects remains largely unexplored. To address this gap, we propose RelationField, the first method to extract inter-object relationships directly from neural radiance fields. RelationField represents relationships between objects as pairs of rays within a neural radiance field, effectively extending its formulation to include implicit relationship queries. To teach RelationField complex, open-vocabulary relationships, relationship knowledge is distilled from multi-modal LLMs. To evaluate RelationField, we solve open-vocabulary 3D scene graph generation tasks and relationship-guided instance segmentation, achieving state-of-the-art performance in both tasks. See the project website at https://relationfield.github.io.
We address an advanced challenge of predicting pedestrian occupancy as an extension of multi-view pedestrian detection in urban traffic. To support this, we have created a new synthetic dataset called MVP-Occ, designed for dense pedestrian scenarios in large-scale scenes. Our dataset provides detailed representations of pedestrians using voxel structures, accompanied by rich semantic scene understanding labels, facilitating visual navigation and insights into pedestrian spatial information. Furthermore, we present a robust baseline model, termed OmniOcc, capable of predicting both the voxel occupancy state and panoptic labels for the entire scene from multi-view images. Through in-depth analysis, we identify and evaluate the key elements of our proposed model, highlighting their specific contributions and importance.
In the field of audio-visual learning, most research tasks focus exclusively on short videos. This paper focuses on the more practical Dense Audio-Visual Event Localization (DAVEL) task, advancing audio-visual scene understanding for longer, untrimmed videos. This task seeks to identify and temporally pinpoint all events simultaneously occurring in both audio and visual streams. Typically, each video encompasses dense events of multiple classes, which may overlap on the timeline, each exhibiting varied durations. Given these challenges, effectively exploiting the audio-visual relations and the temporal features encoded at various granularities becomes crucial. To address these challenges, we introduce a novel CCNet, comprising two core modules: the Cross-Modal Consistency Collaboration (CMCC) and the Multi-Temporal Granularity Collaboration (MTGC). Specifically, the CMCC module contains two branches: a cross-modal interaction branch and a temporal consistency-gated branch. The former branch facilitates the aggregation of consistent event semantics across modalities through the encoding of audio-visual relations, while the latter branch guides one modality's focus to pivotal event-relevant temporal areas as discerned in the other modality. The MTGC module includes a coarse-to-fine collaboration block and a fine-to-coarse collaboration block, providing bidirectional support among coarse- and fine-grained temporal features. Extensive experiments on the UnAV-100 dataset validate our module design, resulting in a new state-of-the-art performance in dense audio-visual event localization. The code is available at https://github.com/zzhhfut/CCNet-AAAI2025.
Multispectral point cloud (MPC) captures 3D spatial-spectral information from the observed scene, which can be used for scene understanding and has a wide range of applications. However, most of the existing classification methods were extensively tested on indoor datasets, and when applied to outdoor datasets they still face problems including sparse labeled targets, differences in land-covers scales, and long-tailed distributions. To address the above issues, an enhanced classification method based on adaptive multi-scale fusion for MPCs with long-tailed distributions is proposed. In the training set generation stage, a grid-balanced sampling strategy is designed to reliably generate training samples from sparse labeled datasets. In the feature learning stage, a multi-scale feature fusion module is proposed to fuse shallow features of land-covers at different scales, addressing the issue of losing fine features due to scale variations in land-covers. In the classification stage, an adaptive hybrid loss module is devised to utilize multi-classification heads with adaptive weights to balance the learning ability of different classes, improving the classification performance of small classes due to various-scales and long-tailed distributions in land-covers. Experimental results on three MPC datasets demonstrate the effectiveness of the proposed method compared with the state-of-the-art methods.