Research on 3D Vision-Language Models (3D-VLMs) is gaining increasing attention, which is crucial for developing embodied AI within 3D scenes, such as visual navigation and embodied question answering. Due to the high density of visual features, especially in large 3D scenes, accurately locating task-relevant visual information is challenging. Existing works attempt to segment all objects and consider their features as scene representations. However, these task-agnostic object features include much redundant information and missing details for the task-relevant area. To tackle these problems, we propose LSceneLLM, an adaptive framework that automatically identifies task-relevant areas by leveraging LLM's visual preference for different tasks, followed by a plug-and-play scene magnifier module to capture fine-grained details in focused areas. Specifically, a dense token selector examines the attention map of LLM to identify visual preferences for the instruction input. It then magnifies fine-grained details of the focusing area. An adaptive self-attention module is leveraged to fuse the coarse-grained and selected fine-grained visual information. To comprehensively evaluate the large scene understanding ability of 3D-VLMs, we further introduce a cross-room understanding benchmark, XR-Scene, which contains a series of large scene understanding tasks including XR-QA, XR-EmbodiedPlanning, and XR-SceneCaption. Experiments show that our method surpasses existing methods on both large scene understanding and existing scene understanding benchmarks. Plunging our scene magnifier module into the existing 3D-VLMs also brings significant improvement.
Recently, several studies have combined Gaussian Splatting to obtain scene representations with language embeddings for open-vocabulary 3D scene understanding. While these methods perform well, they essentially require very dense multi-view inputs, limiting their applicability in real-world scenarios. In this work, we propose SparseLGS to address the challenge of 3D scene understanding with pose-free and sparse view input images. Our method leverages a learning-based dense stereo model to handle pose-free and sparse inputs, and a three-step region matching approach to address the multi-view semantic inconsistency problem, which is especially important for sparse inputs. Different from directly learning high-dimensional CLIP features, we extract low-dimensional information and build bijections to avoid excessive learning and storage costs. We introduce a reconstruction loss during semantic training to improve Gaussian positions and shapes. To the best of our knowledge, we are the first to address the 3D semantic field problem with sparse pose-free inputs. Experimental results show that SparseLGS achieves comparable quality when reconstructing semantic fields with fewer inputs (3-4 views) compared to previous SOTA methods with dense input. Besides, when using the same sparse input, SparseLGS leads significantly in quality and heavily improves the computation speed (5$\times$speedup). Project page: https://ustc3dv.github.io/SparseLGS
3D semantic maps have played an increasingly important role in high-precision robot localization and scene understanding. However, real-time construction of semantic maps requires mobile edge devices with extremely high computing power, which are expensive and limit the widespread application of semantic mapping. In order to address this limitation, inspired by cloud-edge collaborative computing and the high transmission efficiency of semantic communication, this paper proposes a method to achieve real-time semantic mapping tasks with limited-resource mobile devices. Specifically, we design an encoding-decoding semantic communication framework for real-time semantic mapping tasks under limited-resource situations. In addition, considering the impact of different channel conditions on communication, this paper designs a module based on the attention mechanism to achieve stable data transmission under various channel conditions. In terms of simulation experiments, based on the TUM dataset, it was verified that the system has an error of less than 0.1% compared to the groundtruth in mapping and localization accuracy and is superior to some novel semantic communication algorithms in real-time performance and channel adaptation. Besides, we implement a prototype system to verify the effectiveness of the proposed framework and designed module in real indoor scenarios. The results show that our system can complete real-time semantic mapping tasks for common indoor objects (chairs, computers, people, etc.) with a limited-resource device, and the mapping update time is less than 1 second.
Dynamic scene understanding remains a persistent challenge in robotic applications. Early dynamic mapping methods focused on mitigating the negative influence of short-term dynamic objects on camera motion estimation by masking or tracking specific categories, which often fall short in adapting to long-term scene changes. Recent efforts address object association in long-term dynamic environments using neural networks trained on synthetic datasets, but they still rely on predefined object shapes and categories. Other methods incorporate visual, geometric, or semantic heuristics for the association but often lack robustness. In this work, we introduce BYE, a class-agnostic, per-scene point cloud encoder that removes the need for predefined categories, shape priors, or extensive association datasets. Trained on only a single sequence of exploration data, BYE can efficiently perform object association in dynamically changing scenes. We further propose an ensembling scheme combining the semantic strengths of Vision Language Models (VLMs) with the scene-specific expertise of BYE, achieving a 7% improvement and a 95% success rate in object association tasks. Code and dataset are available at https://byencoder.github.io.
TL;DR: Gaussian Splatting is a widely adopted approach for 3D scene representation that offers efficient, high-quality 3D reconstruction and rendering. A major reason for the success of 3DGS is its simplicity of representing a scene with a set of Gaussians, which makes it easy to interpret and adapt. To enhance scene understanding beyond the visual representation, approaches have been developed that extend 3D Gaussian Splatting with semantic vision-language features, especially allowing for open-set tasks. In this setting, the language features of 3D Gaussian Splatting are often aggregated from multiple 2D views. Existing works address this aggregation problem using cumbersome techniques that lead to high computational cost and training time. In this work, we show that the sophisticated techniques for language-grounded 3D Gaussian Splatting are simply unnecessary. Instead, we apply Occam's razor to the task at hand and perform weighted multi-view feature aggregation using the weights derived from the standard rendering process, followed by a simple heuristic-based noisy Gaussian filtration. Doing so offers us state-of-the-art results with a speed-up of two orders of magnitude. We showcase our results in two commonly used benchmark datasets: LERF and 3D-OVS. Our simple approach allows us to perform reasoning directly in the language features, without any compression whatsoever. Such modeling in turn offers easy scene manipulation, unlike the existing methods -- which we illustrate using an application of object insertion in the scene. Furthermore, we provide a thorough discussion regarding the significance of our contributions within the context of the current literature. Project Page: https://insait-institute.github.io/OccamLGS/
The idea of 3D reconstruction as scene understanding is foundational in computer vision. Reconstructing 3D scenes from 2D visual observations requires strong priors to disambiguate structure. Much work has been focused on the anthropocentric, which, characterized by smooth surfaces, coherent normals, and regular edges, allows for the integration of strong geometric inductive biases. Here, we consider a more challenging problem where such assumptions do not hold: the reconstruction of natural scenes containing trees, bushes, boulders, and animals. While numerous works have attempted to tackle the problem of reconstructing animals in the wild, they have focused solely on the animal, neglecting environmental context. This limits their usefulness for analysis tasks, as animals exist inherently within the 3D world, and information is lost when environmental factors are disregarded. We propose a method to reconstruct natural scenes from single images. We base our approach on recent advances leveraging the strong world priors ingrained in Large Language Models and train an autoregressive model to decode a CLIP embedding into a structured compositional scene representation, encompassing both animals and the wild (RAW). To enable this, we propose a synthetic dataset comprising one million images and thousands of assets. Our approach, having been trained solely on synthetic data, generalizes to the task of reconstructing animals and their environments in real-world images. We will release our dataset and code to encourage future research at https://raw.is.tue.mpg.de/
Injecting semantics into 3D Gaussian Splatting (3DGS) has recently garnered significant attention. While current approaches typically distill 3D semantic features from 2D foundational models (e.g., CLIP and SAM) to facilitate novel view segmentation and semantic understanding, their heavy reliance on 2D supervision can undermine cross-view semantic consistency and necessitate complex data preparation processes, therefore hindering view-consistent scene understanding. In this work, we present FreeGS, an unsupervised semantic-embedded 3DGS framework that achieves view-consistent 3D scene understanding without the need for 2D labels. Instead of directly learning semantic features, we introduce the IDentity-coupled Semantic Field (IDSF) into 3DGS, which captures both semantic representations and view-consistent instance indices for each Gaussian. We optimize IDSF with a two-step alternating strategy: semantics help to extract coherent instances in 3D space, while the resulting instances regularize the injection of stable semantics from 2D space. Additionally, we adopt a 2D-3D joint contrastive loss to enhance the complementarity between view-consistent 3D geometry and rich semantics during the bootstrapping process, enabling FreeGS to uniformly perform tasks such as novel-view semantic segmentation, object selection, and 3D object detection. Extensive experiments on LERF-Mask, 3D-OVS, and ScanNet datasets demonstrate that FreeGS performs comparably to state-of-the-art methods while avoiding the complex data preprocessing workload. Our code is publicly available at https://github.com/wb014/FreeGS.
While numerous recent benchmarks focus on evaluating generic Vision-Language Models (VLMs), they fall short in addressing the unique demands of geospatial applications. Generic VLM benchmarks are not designed to handle the complexities of geospatial data, which is critical for applications such as environmental monitoring, urban planning, and disaster management. Some of the unique challenges in geospatial domain include temporal analysis for changes, counting objects in large quantities, detecting tiny objects, and understanding relationships between entities occurring in Remote Sensing imagery. To address this gap in the geospatial domain, we present GEOBench-VLM, a comprehensive benchmark specifically designed to evaluate VLMs on geospatial tasks, including scene understanding, object counting, localization, fine-grained categorization, and temporal analysis. Our benchmark features over 10,000 manually verified instructions and covers a diverse set of variations in visual conditions, object type, and scale. We evaluate several state-of-the-art VLMs to assess their accuracy within the geospatial context. The results indicate that although existing VLMs demonstrate potential, they face challenges when dealing with geospatial-specific examples, highlighting the room for further improvements. Specifically, the best-performing GPT4o achieves only 40\% accuracy on MCQs, which is only double the random guess performance. Our benchmark is publicly available at https://github.com/The-AI-Alliance/GEO-Bench-VLM .
3D scene understanding has become an essential area of research with applications in autonomous driving, robotics, and augmented reality. Recently, 3D Gaussian Splatting (3DGS) has emerged as a powerful approach, combining explicit modeling with neural adaptability to provide efficient and detailed scene representations. However, three major challenges remain in leveraging 3DGS for scene understanding: 1) an imbalance between appearance and semantics, where dense Gaussian usage for fine-grained texture modeling does not align with the minimal requirements for semantic attributes; 2) inconsistencies between appearance and semantics, as purely appearance-based Gaussians often misrepresent object boundaries; and 3) reliance on top-down instance segmentation methods, which struggle with uneven category distributions, leading to over- or under-segmentation. In this work, we propose InstanceGaussian, a method that jointly learns appearance and semantic features while adaptively aggregating instances. Our contributions include: i) a novel Semantic-Scaffold-GS representation balancing appearance and semantics to improve feature representations and boundary delineation; ii) a progressive appearance-semantic joint training strategy to enhance stability and segmentation accuracy; and iii) a bottom-up, category-agnostic instance aggregation approach that addresses segmentation challenges through farthest point sampling and connected component analysis. Our approach achieves state-of-the-art performance in category-agnostic, open-vocabulary 3D point-level segmentation, highlighting the effectiveness of the proposed representation and training strategies. Project page: https://lhj-git.github.io/InstanceGaussian/
Most of current computer vision-based advanced driver assistance systems (ADAS) perform detection and tracking of objects quite successfully under regular conditions. However, under adverse weather and changing lighting conditions, and in complex situations with many overlapping objects, these systems are not completely reliable. The spectral reflectance of the different objects in a driving scene beyond the visible spectrum can offer additional information to increase the reliability of these systems, especially under challenging driving conditions. Furthermore, this information may be significant enough to develop vision systems that allow for a better understanding and interpretation of the whole driving scene. In this work we explore the use of snapshot, video-rate hyperspectral imaging (HSI) cameras in ADAS on the assumption that the near infrared (NIR) spectral reflectance of different materials can help to better segment the objects in real driving scenarios. To do this, we have used the HSI-Drive 1.1 dataset to perform various experiments on spectral classification algorithms. However, the information retrieval of hyperspectral recordings in natural outdoor scenarios is challenging, mainly because of deficient colour constancy and other inherent shortcomings of current snapshot HSI technology, which poses some limitations to the development of pure spectral classifiers. In consequence, in this work we analyze to what extent the spatial features codified by standard, tiny fully convolutional network (FCN) models can improve the performance of HSI segmentation systems for ADAS applications. The abstract above is truncated due to submission limits. For the full abstract, please refer to the published article.
Simulating stylized human-scene interactions (HSI) in physical environments is a challenging yet fascinating task. Prior works emphasize long-term execution but fall short in achieving both diverse style and physical plausibility. To tackle this challenge, we introduce a novel hierarchical framework named SIMS that seamlessly bridges highlevel script-driven intent with a low-level control policy, enabling more expressive and diverse human-scene interactions. Specifically, we employ Large Language Models with Retrieval-Augmented Generation (RAG) to generate coherent and diverse long-form scripts, providing a rich foundation for motion planning. A versatile multicondition physics-based control policy is also developed, which leverages text embeddings from the generated scripts to encode stylistic cues, simultaneously perceiving environmental geometries and accomplishing task goals. By integrating the retrieval-augmented script generation with the multi-condition controller, our approach provides a unified solution for generating stylized HSI motions. We further introduce a comprehensive planning dataset produced by RAG and a stylized motion dataset featuring diverse locomotions and interactions. Extensive experiments demonstrate SIMS's effectiveness in executing various tasks and generalizing across different scenarios, significantly outperforming previous methods.
Quantifying the gap between synthetic and real-world imagery is essential for improving both transformer-based models - that rely on large volumes of data - and datasets, especially in underexplored domains like aerial scene understanding where the potential impact is significant. This paper introduces a novel methodology for scene complexity assessment using Multi-Model Consensus Metric (MMCM) and depth-based structural metrics, enabling a robust evaluation of perceptual and structural disparities between domains. Our experimental analysis, utilizing real-world (Dronescapes) and synthetic (Skyscenes) datasets, demonstrates that real-world scenes generally exhibit higher consensus among state-of-the-art vision transformers, while synthetic scenes show greater variability and challenge model adaptability. The results underline the inherent complexities and domain gaps, emphasizing the need for enhanced simulation fidelity and model generalization. This work provides critical insights into the interplay between domain characteristics and model performance, offering a pathway for improved domain adaptation strategies in aerial scene understanding.
The increasing number of actions in the real world makes it difficult for traditional deep-learning models to recognize unseen actions. Recently, pretrained contrastive image-based visual-language (I-VL) models have been adapted for efficient “zero-shot” scene understanding. Pairing such models with transformers to implement temporal modeling has been rewarding for zero-shot action recognition (ZSAR). However, the significance of modeling the local spatial context of objects and action environments remains unexplored. In this work, we propose a ZSAR framework called LoCATe-GAT, comprising a novel Local Context-Aggregating Temporal transformer (LoCATe) and a Graph Attention Network (GAT). Specifically, image and text encodings extracted from a pretrained I-VL model are used as inputs for LoCATe-GAT. Motivated by the observation that object-centric and environmental contexts drive both distinguishability and functional similarity between actions, LoCATe captures multi-scale local context using dilated convolutional layers during temporal modeling. Furthermore, the proposed GAT models semantic relationships between classes and achieves a strong synergy with the video embeddings produced by LoCATe. Extensive experiments on four widely-used benchmarks – UCF101, HMDB51, ActivityNet, and Kinetics – show we achieve state-of-the-art results. Specifically, we obtain relative gains of 3.8% and 4.8% on these datasets in conventional and 16.6% on UCF101in generalized ZSAR settings. For large-scale datasets like ActivityNet and Kinetics, our method achieves a relative gain of 31.8% and 27.9%, respectively, over the previous methods. Additionally, we gain 25.3% and 18.4% on UCF101 and HMDB51 as per the recent “TruZe” evaluation protocol.
Recent advances in multimodal models have demonstrated impressive capabilities in object recognition and scene understanding. However, these models often struggle with precise spatial localization - a critical capability for real-world applications. Inspired by how humans use grid-based references like chess boards and maps, we propose introducing explicit visual position encoding through a simple grid overlay approach. By adding a 9x9 black grid pattern onto input images, our method provides visual spatial guidance analogous to how positional encoding works in transformers, but in an explicit, visual form. Experiments on the COCO 2017 dataset demonstrate that our grid-based approach achieves significant improvements in localization accuracy, with a 107.4% increase in IoU (from 0.27 to 0.56) and a 194.4% improvement in GIoU (from 0.18 to 0.53) compared to baseline performance. Through attention visualization analysis, we show how this visual position encoding helps models better ground spatial relationships. Our method's simplicity and effectiveness make it particularly valuable for applications requiring accurate spatial reasoning, such as robotic manipulation, medical imaging, and autonomous navigation.
Object detection and semantic segmentation are both scene understanding tasks yet they differ in data structure and information level. Object detection requires box coordinates for object instances while semantic segmentation requires pixel-wise class labels. Making use of one task's information to train the other would be beneficial for multi-task partially supervised learning where each training example is annotated only for a single task, having the potential to expand training sets with different-task datasets. This paper studies various weak losses for partially annotated data in combination with existing supervised losses. We propose Box-for-Mask and Mask-for-Box strategies, and their combination BoMBo, to distil necessary information from one task annotations to train the other. Ablation studies and experimental results on VOC and COCO datasets show favorable results for the proposed idea. Source code and data splits can be found at https://github.com/lhoangan/multas.
We present the updated version of the HSI-Drive dataset aimed at developing automated driving systems (ADS) using hyperspectral imaging (HSI). The v2.0 version includes new annotated images from videos recorded during winter and fall in real driving scenarios. Added to the spring and summer images included in the previous v1.1 version, the new dataset contains 752 images covering the four seasons. In this paper, we show the improvements achieved over previously published results obtained on the v1.1 dataset, showcasing the enhanced performance of models trained on the new v2.0 dataset. We also show the progress made in comprehensive scene understanding by experimenting with more capable image segmentation models. These models include new segmentation categories aimed at the identification of essential road safety objects such as the presence of vehicles and road signs, as well as highly vulnerable groups like pedestrians and cyclists. In addition, we provide evidence of the performance and robustness of the models when applied to segmenting HSI video sequences captured in various environments and conditions. Finally, for a correct assessment of the results described in this work, the constraints imposed by the processing platforms that can sensibly be deployed in vehicles for ADS must be taken into account. Thus, and although implementation details are out of the scope of this paper, we focus our research on the development of computationally efficient, lightweight ML models that can eventually operate at high throughput rates. The dataset and some examples of segmented videos are available in https://ipaccess.ehu.eus/HSI-Drive/.
Existing conditional Denoising Diffusion Probabilistic Models (DDPMs) with a Noise-Conditional Framework (NCF) remain challenging for 3D scene understanding tasks, as the complex geometric details in scenes increase the difficulty of fitting the gradients of the data distribution (the scores) from semantic labels. This also results in longer training and inference time for DDPMs compared to non-DDPMs. From a different perspective, we delve deeply into the model paradigm dominated by the Conditional Network. In this paper, we propose an end-to-end robust semantic Segmentation Network based on a Conditional-Noise Framework (CNF) of DDPMs, named CDSegNet. Specifically, CDSegNet models the Noise Network (NN) as a learnable noise-feature generator. This enables the Conditional Network (CN) to understand 3D scene semantics under multi-level feature perturbations, enhancing the generalization in unseen scenes. Meanwhile, benefiting from the noise system of DDPMs, CDSegNet exhibits strong noise and sparsity robustness in experiments. Moreover, thanks to CNF, CDSegNet can generate the semantic labels in a single-step inference like non-DDPMs, due to avoiding directly fitting the scores from semantic labels in the dominant network of CDSegNet. On public indoor and outdoor benchmarks, CDSegNet significantly outperforms existing methods, achieving state-of-the-art performance.
Open-vocabulary 3D scene understanding is indispensable for embodied agents. Recent works leverage pretrained vision-language models (VLMs) for object segmentation and project them to point clouds to build 3D maps. Despite progress, a point cloud is a set of unordered coordinates that requires substantial storage space and does not directly convey occupancy information or spatial relation, making existing methods inefficient for downstream tasks, e.g., path planning and complex text-based object retrieval. To address these issues, we propose Octree-Graph, a novel scene representation for open-vocabulary 3D scene understanding. Specifically, a Chronological Group-wise Segment Merging (CGSM) strategy and an Instance Feature Aggregation (IFA) algorithm are first designed to get 3D instances and corresponding semantic features. Subsequently, an adaptive-octree structure is developed that stores semantics and depicts the occupancy of an object adjustably according to its shape. Finally, the Octree-Graph is constructed where each adaptive-octree acts as a graph node, and edges describe the spatial relations among nodes. Extensive experiments on various tasks are conducted on several widely-used datasets, demonstrating the versatility and effectiveness of our method.
Recently, Vision Language Models (VLMs) have experienced significant advancements, yet these models still face challenges in spatial hierarchical reasoning within indoor scenes. In this study, we introduce ROOT, a VLM-based system designed to enhance the analysis of indoor scenes. Specifically, we first develop an iterative object perception algorithm using GPT-4V to detect object entities within indoor scenes. This is followed by employing vision foundation models to acquire additional meta-information about the scene, such as bounding boxes. Building on this foundational data, we propose a specialized VLM, SceneVLM, which is capable of generating spatial hierarchical scene graphs and providing distance information for objects within indoor environments. This information enhances our understanding of the spatial arrangement of indoor scenes. To train our SceneVLM, we collect over 610,000 images from various public indoor datasets and implement a scene data generation pipeline with a semi-automated technique to establish relationships and estimate distances among indoor objects. By utilizing this enriched data, we conduct various training recipes and finish SceneVLM. Our experiments demonstrate that \rootname facilitates indoor scene understanding and proves effective in diverse downstream applications, such as 3D scene generation and embodied AI. The code will be released at \url{https://github.com/harrytea/ROOT}.
Urban scene reconstruction is crucial for real-world autonomous driving simulators. Although existing methods have achieved photorealistic reconstruction, they mostly focus on pinhole cameras and neglect fisheye cameras. In fact, how to effectively simulate fisheye cameras in driving scene remains an unsolved problem. In this work, we propose UniGaussian, a novel approach that learns a unified 3D Gaussian representation from multiple camera models for urban scene reconstruction in autonomous driving. Our contributions are two-fold. First, we propose a new differentiable rendering method that distorts 3D Gaussians using a series of affine transformations tailored to fisheye camera models. This addresses the compatibility issue of 3D Gaussian splatting with fisheye cameras, which is hindered by light ray distortion caused by lenses or mirrors. Besides, our method maintains real-time rendering while ensuring differentiability. Second, built on the differentiable rendering method, we design a new framework that learns a unified Gaussian representation from multiple camera models. By applying affine transformations to adapt different camera models and regularizing the shared Gaussians with supervision from different modalities, our framework learns a unified 3D Gaussian representation with input data from multiple sources and achieves holistic driving scene understanding. As a result, our approach models multiple sensors (pinhole and fisheye cameras) and modalities (depth, semantic, normal and LiDAR point clouds). Our experiments show that our method achieves superior rendering quality and fast rendering speed for driving scene simulation.