Complex Scenes Research Articles

Assessing Habitat Quality on Synergetic Land-Cover Dataset Across the Greater Mekong Subregion over the Last Four Decades

In the face of rapid infrastructure expansion and escalating anthropogenic activities, it becomes imperative to prioritize the examination of long-term transformations in land cover and ecological quality within the Greater Mekong Subregion (GMS). We developed an ecological evaluation system integrating the land cover data assimilation framework (LCDAF) with the InVEST model to accomplish this goal. The LCDAF compensates for the disadvantages of weather interference, difficulty in recognizing complex scenes, and poor generalization in remote sensing image classification, and also adds temporal continuity that other fusion methods do not have. The synthesized land cover dataset demonstrates superior overall accuracy compared to five existing global products. This enhanced dataset provides a robust foundation for comprehensive analysis and decision making within the ecological evaluation system. We implemented a rigorous and quantitative assessment of changes in land cover and habitat quality spanning 1980 to 2020. The land cover analysis unveiled a noteworthy trend that surfaced in the dynamic interplay between forested areas and croplands, highlighting simultaneous processes of forest restoration and agricultural expansion, albeit at varying rates. Further analysis of habitat quality showed that the GMS generally sustained a moderate level with a slight downward trend observed over the period. Significantly, Laos attained the highest ranking in habitat quality, succeeded by Myanmar, China, Cambodia, Vietnam, and Thailand. In human factors, land use intensity and landscape fragmentation emerge as contributors with detrimental effects on habitat quality. Substantial progress was achieved in implementing forestland conservation measures, exemplified in regions such as Cambodia and Guangxi Province of China, where these endeavors proved effective in mitigating habitat degradation. Despite these positive endeavors, the GMS’s overall habitat quality did not significantly improve. It emphasizes the enduring challenges confronted by the region in terms of ecological management and habitat conservation.

Semantic Fusion Algorithm of 2D LiDAR and Camera Based on Contour and Inverse Projection

Common single-line 2D LiDAR sensors and cameras have become core components in the field of robotic perception due to their low cost, compact size, and practicality. However, during the data fusion process, the randomness and complexity of real industrial scenes pose challenges. Traditional calibration methods for LiDAR and cameras often rely on precise targets and can accumulate errors, leading to significant limitations. Additionally, the semantic fusion of LiDAR and camera data typically requires extensive projection calculations, complex clustering algorithms, or sophisticated data fusion techniques, resulting in low real-time performance when handling large volumes of data points in dynamic environments. To address these issues, this paper proposes a semantic fusion algorithm for LiDAR and camera data based on contour and inverse projection. The method has two remarkable features: (1) Combined with the ellipse extraction algorithm of the arc support line segment, a LiDAR and camera calibration algorithm based on various regular shapes of an environmental target is proposed, which improves the adaptability of the calibration algorithm to the environment. (2) This paper proposes a semantic segmentation algorithm based on the inverse projection of target contours. It is specifically designed to be versatile and applicable to both linear and arc features, significantly broadening the range of features that can be utilized in various tasks. This flexibility is a key advantage, as it allows the method to adapt to a wider variety of real-world scenarios where both types of features are commonly encountered. Compared with existing LiDAR point cloud semantic segmentation methods, this algorithm eliminates the need for complex clustering algorithms, data fusion techniques, and extensive laser point reprojection calculations. When handling a large number of laser points, the proposed method requires only one or two inverse projections of the contour to filter the range of laser points that intersect with specific targets. This approach enhances both the accuracy of point cloud searches and the speed of semantic processing. Finally, the validity of the semantic fusion algorithm is proven by field experiments.

RDAS-MODEL: an efficient model of road drivable area segmentation for autonomous driving

Abstract The drivable area segmentation technique plays a crucial role in ensuring the safe driving of autonomous driving vehicles in complex environments. However, in realistic drivable area segmentation scenarios, the presence of complex and high uncertainty factors such as lighting differences and weather makes accurate drivable area segmentation still challenging. To reduce the impact of these uncertainties on the real scene, and improve the accuracy of the drivable area segmentation, this study proposes an efficient road drivable area segmentation model (RDAS-MODEL). Firstly, a multi-scale feature adaptation module (MFAM) is designed and integrated into the backbone network, which aims to reduce the uncertainty of the feature map, and thus effectively improve the accuracy of the drivable area segmentation. Next, the high and low feature interactive pyramid network (HLFIPN) is introduced, which can enhance the interaction between shallow and deep features, thus improving the segmentation of multi-scale objects in complex scenes. Then, the redesigned shared convolutional segmentation head (SCSH) can reduce the number of parameters and computation, and also ensure the real-time performance of the model, which further improves the segmentation performance. Finally, the adaptive training sample selection (ATSS) strategy is utilized to balance positive and negative samples, effectively minimizing object misdetection and omission. Experimental results based on the BDD100K dataset indicate that the mean Average Precision (mAP)mask@50 and mAPmask@50:95 of the proposed model on the drivable area are 95.2% and 77.2%, respectively, and the real-time inference speed is 102 frames per second (FPS), which significantly exceeds the performance of some advanced methods. Experiments are also conducted on the UAS dataset to further demonstrate the generalization ability and effectiveness of the method. In addition, the qualitative results of experiments on real campus scenes suggest that the proposed method is well adapted to real scenes and has a high segmentation accuracy. The results of a large number of experiments conducted on two public datasets and self-captured images all show that the proposed method can effectively improve the segmentation accuracy of the drivable area, providing a new idea for accurately segmenting the drivable area.

SCRM-Net: Self-Supervised Deep Clustering Feature Representation for Urban 3D Mesh Semantic Segmentation

Semantic urban 3D meshes obtained by deep learning networks have been widely applied in urban analytics. Typically, a large amount of labeled samples are required to train a deep learning network to extract discriminative features for the semantic segmentation of urban 3D mesh. However, it is labor intensive and time consuming to obtain enough labeled samples due to the complexity of urban 3D scenes. To obtain discriminative features without extensive labeled data, we propose a novel self-supervised deep clustering feature learning network, named SCRM-Net. The proposed SCRM-Net consists of two mutually self-supervised branches: one branch utilizes autoencoder to learn intrinsic feature representations of urban 3D Mesh, while the other applies GCN to capture the structural relationships between them. During the semantic segmentation process, only a limited proportion of the labeled samples is required to fine-tune the pretrained encoder of SCRM-Net for discriminative feature extraction and to train the segmentation head consisting of two edge convolution layers. Extensive comparative experiments demonstrate the effectiveness of our approach and show its competitiveness against the state-of-the-art semantic segmentation methods.

Multimodal MEG and microstructure-MRI investigations of the human hippocampal scene network.

Although several studies have demonstrated that perceptual discrimination of complex scenes relies on an extended hippocampal posteromedial system, we currently have limited insight into the specific functional and structural properties of this system in humans. Here, combining electrophysiological (magnetoencephalography, MEG) and advanced microstructural (multi-shell diffusion MRI, dMRI; quantitative magnetisation transfer, qMT) imaging in healthy human adults (30 female/10 male), we show that both theta power modulation of the hippocampus, and fibre restriction/hindrance (reflecting axon packing/myelination) of the fornix (a major input/output pathway of the hippocampus), were independently related to scene, but not face, perceptual discrimination accuracy. Conversely, microstructural features of the inferior longitudinal fasciculus (a long-range occipito-anterotemporal tract) correlated with face, but not scene, perceptual discrimination accuracy. Our results provide new mechanistic insight into the neurocognitive systems underpinning complex scene discrimination, providing novel support for the idea of multiple processing streams within the human medial temporal lobe.Significance Statement In contrast to theories positing segregated cortical areas for perception and memory, the specialized representations of the hippocampus may support both the perception and memory of visual scenes. To investigate, we utilised the unique window into hippocampal electrophysiological activity offered by magnetoencephalography (MEG). We found hippocampal theta activity modulations in the hippocampus and posteromedial cortex during scene, versus face and shape-size, perceptual odd-one-out discrimination, the magnitude of which correlated with scene, but not face or shape-size, discrimination accuracy. Moreover, multimodal white matter imaging revealed that specific microstructural features of the fornix - the major hippocampal output tract - independently predicted scene discrimination performance. Our multimodal MEG-microstructure study provides novel evidence that the hippocampus and connected structures conjointly support online scene processing.

Enhancing Low-Light High-Dynamic-Range Image from Industrial Cameras Using Dynamic Weighting and Pyramid Fusion

In order to solve the problem of imaging quality of industrial cameras for low-light and large dynamic scenes in many fields, such as smart city and target recognition, this study focuses on overcoming two core challenges: first, the loss of image details due to the significant difference in light distribution in complex scenes, and second, the coexistence of dark and light areas under the constraints of the limited dynamic range of a camera. To this end, we propose a low-light high-dynamic-range image enhancement method based on dynamic weights and pyramid fusion. In order to verify the effectiveness of the method, experimental data covering full-time scenes are acquired based on an image acquisition platform built in the laboratory, and a comprehensive evaluation system combining subjective visual assessment and objective indicators is constructed. The experimental results show that, in a multi-temporal fusion task, this study’s method performs well in multiple key indicators such as information entropy (EN), average gradient (AG), edge intensity (EI), and spatial frequency (SF), making it especially suitable for imaging in low-light and high-dynamic-range environments. Specifically in localized low-light high-dynamic-range regions, compared with the best-performing comparison method, the information entropy indexes of this study’s method are improved by 4.88% and 6.09%, which fully verifies its advantages in detail restoration. The research results provide a technical solution with all-day adaptive capability for low-cost and lightweight surveillance equipment, such as intelligent transportation systems and remote sensing security systems, which has broad application prospects.

Multi-Scale Feature Fusion Based on Difference Enhancement for Remote Sensing Image Change Detection

Remote sensing image change detection is a core task of remote sensing image analysis; its purpose is to identify and quantify land cover changes in different periods. However, when the existing methods deal with complex features and subtle changes in buildings, vegetation, water bodies, roads, and other ground objects, there are often problems of false detection and missing detection, which affect the detection accuracy. To improve the accuracy of change detection, a multi-scale feature fusion network based on difference enhancement (FEDNet) is proposed. The FEDNet consists of a difference enhancement module (DEM) and a multi-scale feature fusion module (MFM). By summing the variation features of two-phase remote sensing images, the DEM enhances pixel-level differences, captures subtle changes, and aggregates features. The MFM fully integrates the multi-stage deep semantic information, which enables better extraction of changing features in complex scenes. Experiments on the LEVIR-CD, CLCD, WHU, NJDS, and GBCNR datasets show that the FEDNet significantly improves the detection efficiency of changes in buildings, cities, and vegetation. In terms of F1 value, IoU (Intersection over Union), precision, and recall rate, the FEDNet is superior to existing methods, which verifies its excellent performance.

AAB-YOLO: An Improved YOLOv11 Network for Apple Detection in Natural Environments

Apple detection in natural environments is crucial for advancing agricultural automation. However, orchards often employ bagging techniques to protect apples from pests and improve quality, which introduces significant detection challenges due to the varied appearance and occlusion of apples caused by bags. Additionally, the complex and variable natural backgrounds further complicate the detection process. To address these multifaceted challenges, this study introduces AAB-YOLO, a lightweight apple detection model based on an improved YOLOv11 framework. AAB-YOLO incorporates ADown modules to reduce model complexity, the C3k2_ContextGuided module for enhanced understanding of complex scenes, and the Detect_SEAM module for improved handling of occluded apples. Furthermore, the Inner_EIoU loss function is employed to boost detection accuracy and efficiency. The experimental results demonstrate significant improvements: mAP@50 increases from 0.917 to 0.921, precision rises from 0.948 to 0.951, and recall improves by 1.04%, while the model’s parameter count and computational complexity are reduced by 37.7% and 38.1%, respectively. By achieving lightweight performance while maintaining high accuracy, AAB-YOLO effectively meets the real-time apple detection needs in natural environments, overcoming the challenges posed by orchard bagging techniques and complex backgrounds.

LGINet: Autonomous vehicle trajectory prediction using lane graphs and imitative learning policy

Trajectory prediction is a challenging problem in autonomous driving, due to changes in scene structure and multimodal distributions of future trajectories. This study proposes a novel trajectory prediction model (LGINet). The central innovation of LGINet is the Context-Aware Graph Attention (CAGA) layer, which achieves a balance between handling scene complexity and the ability to capture information, improving prediction performance. Unlike conventional attention mechanisms, CAGA uses Sparsemax to focus on key neighbors, reducing noise and enhancing computational efficiency. The multi-head attention mechanism enhances the ability of the model to capture interactions across both local and long-distance spatial scales. Experiments with the nuScenes dataset show that the proposed LGINet model outperforms existing prediction methods. Specifically, the MinADE 5 value of LGINet is 1.23, the MR 5,2 value is 0.49, and the MinFDE 5 value is 2.39, demonstrating its effectiveness in generating diverse and accurate future trajectories for autonomous driving applications. Qualitative comparisons confirm that LGINet generates diverse trajectories aligned with scene structures and agent motion, highlighting its potential for autonomous driving.

MM-CamObj: A Comprehensive Multimodal Dataset for Camouflaged Object Scenarios

Large visual-language models (LVLMs) have achieved great success in multiple applications. However, they still encounter challenges in complex scenes, especially those involving camouflaged objects. This is primarily due to the lack of samples related to camouflaged scenes in the training dataset. To mitigate this issue, we construct the MM-CamObj dataset for the first time, comprising two subsets: CamObj-Align and CamObj-Instruct. Specifically, CamObj-Align contains 11,363 image-text pairs, and it is designed for VL alignment and injecting rich knowledge of camouflaged scenes into LVLMs. CamObj-Instruct is collected for fine-tuning the LVLMs with improved instruction-following capabilities, and it includes 11,363 images and 68,849 conversations with diverse instructions. Based on the MM-CamObj dataset, we propose the CamObj-Llava, an LVLM specifically designed for addressing tasks in camouflaged scenes. To facilitate our model's effective acquisition of knowledge about camouflaged objects and scenes, we introduce a curriculum learning strategy with six distinct modes. Additionally, we construct the CamObj-Bench to evaluate the existing LVLMs' capabilities of understanding, recognition, localization and count in camouflage scenes. This benchmark includes 600 images and 7 tasks, with a total of 9,449 questions. Extensive experiments are conducted on the CamObj-Bench with CamObj-Llava, 8 existing open-source and 3 closed-source LVLMs. Surprisingly, the results indicate that our model achieves a 25.84% improvement in 4 out of 7 tasks compared to GPT-4o.

High-Fidelity Polarimetric Implicit 3D Reconstruction with View-Dependent Physical Representation

Neural implicit methods have made remarkable progress in 3D reconstruction. However, previous methods often assume view-independent properties of target objects, which fails to accurately reconstruct objects with challenging characteristics, such as transparency and high reflectivity. To address this limitation, we propose a polarimetric implicit 3D reconstruction method that integrates geometric and polarization information, enabling the production of high-quality meshes in complex scenes. For high-fidelity surface reconstruction, we introduce a view-dependent physical representation that thoroughly analyzes the subtle physical properties of reflections. The reconstruction process is further enhanced by a simple yet effective view-dependent detection algorithm and optimized using the principles of ray tracing and polarization. Experimental results demonstrate the superior performance of the proposed method in both real and synthetic scenarios.

Energy-Guided Optimization for Personalized Image Editing with Pretrained Text-to-Image Diffusion Models

The rapid advancement of pretrained text-driven diffusion models has significantly enriched applications in image generation and editing. However, as the demand for personalized content editing increases, new challenges emerge especially when dealing with arbitrary objects and complex scenes. Existing methods usually mistakes mask as the object shape prior, which struggle to achieve a seamless integration result. The mostly used inversion noise initialization also hinders the identity consistency towards the target object. To address these challenges, we propose a novel training-free framework that formulates personalized content editing as the optimization of edited images in the latent space, using diffusion models as the energy function guidance conditioned by reference text-image pairs. A coarse-to-fine strategy is proposed that employs text energy guidance at the early stage to achieve a natural transition toward the target class and uses point-to-point feature-level image energy guidance to perform fine-grained appearance alignment with the target object. Additionally, we introduce the latent space content composition to enhance overall identity consistency with the target. Extensive experiments demonstrate that our method excels in object replacement even with a large domain gap, highlighting its potential for high-quality, personalized image editing.

EGSRAL:An Enhanced 3D Gaussian Splatting Based Renderer with Automated Labeling for Large-Scale Driving Scene

3D Gaussian Splatting (3D GS) has gained popularity due to its faster rendering speed and high-quality novel view synthesis. Some researchers have explored using 3D GS for reconstructing driving scenes. However, these methods often rely on various types of data, such as depth maps, 3D bounding boxes, and trajectories of moving objects. Additionally, the lack of annotations for synthesized images limits their direct application in downstream tasks. To address these issues, we propose EGSRAL, a 3D GS-based method that relies solely on training images without extra annotations. EGSRAL enhances 3D GS's capability to model both dynamic objects and static backgrounds and introduces a novel adaptor for auto labeling, generating corresponding annotations based on existing annotations. We also propose a grouping strategy for vanilla 3D GS to address perspective issues in rendering large-scale, complex scenes. Our method achieves state-of-the-art performance on multiple datasets without any extra annotation. For example, the PSNR metric reaches 29.04 on the nuScenes dataset. Moreover, our automated labeling can significantly improve the performance of 2D/3D detection tasks.

Normal-NeRF: Ambiguity-Robust Normal Estimation for Highly Reflective Scenes

Neural Radiance Fields (NeRF) often struggle with reconstructing and rendering highly reflective scenes. Recent advancements have developed various reflection-aware appearance models to enhance NeRF's capability to render specular reflections. However, the robust reconstruction of highly reflective scenes is still hindered by the inherent shape ambiguity on specular surfaces. Existing methods typically rely on additional geometry priors to regularize the shape prediction, but this can lead to oversmoothed geometry in complex scenes. Observing the critical role of surface normals in parameterizing reflections, we introduce a transmittance-gradient-based normal estimation technique that remains robust even under ambiguous shape conditions. Furthermore, we propose a dual activated densities module that effectively bridges the gap between smooth surface normals and sharp object boundaries. Combined with a reflection-aware appearance model, our proposed method achieves robust reconstruction and high-fidelity rendering of scenes featuring both highly specular reflections and intricate geometric structures. Extensive experiments demonstrate that our method outperforms existing state-of-the-art methods on various datasets.

Foundation Model Driven Appearance Extraction for Robust Multiple Object Tracking

Multiple Object Tracking (MOT) is a fundamental task in computer vision. Existing methods utilize motion information or appearance information to perform object tracking. However, these algorithms still struggle with special circumstances, such as occlusion and blurring in complex scenes. Inspired by the fact that people can pinpoint objects through verbal descriptions, we explore performing long-term robust tracking using semantic features of objects. Motivated by the success of the multimodal foundation model in text-image alignment, we reconsider the appearance feature extraction module in MOT and propose a Foundation model Driven multi-object tracker (FDTracker). Specifically, we propose a two-stage trained appearance feature extractor. In the first stage, using a single image of the object as input, the model could capture the attributes of objects with the assistance of natural language instructions. In the second stage, using a sequence of images of objects as input, the model learns how to use these attributes to distinguish between different objects and connect the same object at different times. Finally, for coordinating appearance and motion information, we propose a reasonable combined strategy, which better facilitates trajectory assignment and reconnection. Extensive experiments on benchmarks demonstrate the robustness of FDTracker.

SceneX: Procedural Controllable Large-Scale Scene Generation

Developing comprehensive explicit world models is crucial for understanding and simulating real-world scenarios. Recently, Procedural Controllable Generation (PCG) has gained significant attention in large-scale scene generation by enabling the creation of scalable, high-quality assets. However, PCG faces challenges such as limited modular diversity, high expertise requirements, and challenges in managing the diverse elements and structures in complex scenes. In this paper, we introduce a large-scale scene generation framework, SceneX, which can automatically produce high-quality procedural models according to designers' textual descriptions. Specifically, the proposed method comprises two components, PCGHub and PCGPlanner. The former encompasses an extensive collection of accessible procedural assets and thousands of hand-craft API documents to perform as a standard protocol for PCG controller. The latter aims to generate executable actions for Blender to produce controllable and precise 3D assets guided by the user's instructions. Extensive experiments demonstrated the capability of our method in controllable large-scale scene generation, including nature scenes and unbounded cities, as well as scene editing such as asset placement and season translation.

Alignment-Free RGB-T Salient Object Detection: A Large-Scale Dataset and Progressive Correlation Network

Alignment-free RGB-Thermal (RGB-T) salient object detection (SOD) aims to achieve robust performance in complex scenes by directly leveraging the complementary information from unaligned visible-thermal image pairs, without requiring manual alignment. However, the labor-intensive process of collecting and annotating image pairs limits the scale of existing benchmarks, hindering the advancement of alignment-free RGB-T SOD. In this paper, we construct a large-scale and high-diversity unaligned RGB-T SOD dataset named UVT20K, comprising 20,000 image pairs, 407 scenes, and 1256 object categories. All samples are collected from real-world scenarios with various challenges, such as low illumination, image clutter, complex salient objects, and so on. To support the exploration for further research, each sample in UVT20K is annotated with a comprehensive set of ground truths, including saliency masks, scribbles, boundaries, and challenge attributes. In addition, we propose a Progressive Correlation Network (PCNet), which models inter- and intra-modal correlations on the basis of explicit alignment to achieve accurate predictions in unaligned image pairs. Extensive experiments conducted on two unaligned three weakly aligned three aligned datasets demonstrate the effectiveness of our method.

Style Nursing with Spatial and Semantic Guidance for Zero-Shot Traffic Scene Style Transfer

Recent advances in text-to-image diffusion models have shown an outstanding ability in zero-shot style transfer. However, existing methods often struggle to balance preserving the semantic content of the input image and faithfully transferring the target style in line with the edit prompt. Especially when applied to complex traffic scenes with diverse objects, layouts, and stylistic variations, current diffusion models tend to exhibit Style Neglection, i.e., failing to generate the required style in the prompt. To address this issue, we propose Style Nursing, which directs the model to focus on style subject tokens in the text prompt and excites their corresponding visual activations. Moreover, we introduce Spatial and Semantic Guidance to guide the preservation of content after editing, which utilizes spatial features from the DDIM sampling process together with attention maps from the semantic reconstruction. To evaluate the performance of zero-shot style transfer methods in traffic scenes, we present STREET-6K, a new benchmark dataset comprising 6000 images showcasing diverse traffic scenes and style transfer variations, accompanied by comprehensive annotations and evaluation metrics. Our approach beats state-of-the-art image translation methods in comprehensive quantitative metrics and human evaluations on traffic scene image synthesis while seamlessly generalizing to various other types of images without training or fine-tuning. Further experiments on detection and segmentation tasks show that fine-tuning perception models on our synthesized images improves Recall and mean Intersection over Union (mIoU) by over 10% and 3% respectively in rarely-seen traffic scenes.

HybridReg: Robust 3D Point Cloud Registration with Hybrid Motions

Scene-level point cloud registration is very challenging when considering dynamic foregrounds. Existing indoor datasets mostly assume rigid motions, so the trained models cannot robustly handle scenes with non-rigid motions. On the other hand, non-rigid datasets are mainly object-level, so the trained models cannot generalize well to complex scenes. This paper presents HybridReg, a new approach to 3D point cloud registration, learning uncertainty mask to account for hybrid motions: rigid for backgrounds and non-rigid/rigid for instance-level foregrounds. First, we build a scene-level 3D registration dataset, namely HybridMatch, designed specifically with strategies to arrange diverse deforming foregrounds in a controllable manner. Second, we account for different motion types and formulate a mask-learning module to alleviate the interference of deforming outliers. Third, we exploit a simple yet effective negative log-likelihood loss to adopt uncertainty to guide the feature extraction and correlation computation. To our best knowledge, HybridReg is the first work that exploits hybrid motions for robust point cloud registration. Extensive experiments show HybridReg's strengths, leading it to achieve state-of-the-art performance on both widely-used indoor and outdoor datasets.

BEV-TSR: Text-Scene Retrieval in BEV Space for Autonomous Driving

The rapid development of the autonomous driving industry has led to a significant accumulation of autonomous driving data. Consequently, there comes a growing demand for retrieving data to provide specialized optimization. However, directly applying previous image retrieval methods faces several challenges, such as the lack of global feature representation and inadequate text retrieval ability for complex driving scenes. To address these issues, firstly, we propose the BEV-TSR framework which leverages descriptive text as an input to retrieve corresponding scenes in the Bird’s Eye View (BEV) space. Then to facilitate complex scene retrieval with extensive text descriptions, we employ a large language model (LLM) to extract the semantic features of the text inputs and incorporate knowledge graph embeddings to enhance the semantic richness of the language embedding. To achieve feature alignment between the BEV feature and language embedding, we propose Shared Cross-modal Embedding with a set of shared learnable embeddings to bridge the gap between these two modalities, and employ a caption generation task to further enhance the alignment. Furthermore, there lack of well-formed retrieval datasets for effective evaluation. To this end, we establish a multi-level retrieval dataset, nuScenes-Retrieval, based on the widely adopted nuScenes dataset. Experimental results on the multi-level nuScenes-Retrieval show that BEV-TSR achieves state-of-the-art performance, e.g., 85.78% and 87.66% top-1 accuracy on scene-to-test and text-to-scene retrieval respectively.