End-to-end autonomous driving demonstrates strong planning capabilities with large-scale data but still struggles in complex, rare scenarios due to limited commonsense. In contrast, Large Vision-Language Models (LVLMs) excel in scene understanding and reasoning. The path forward lies in merging the strengths of both approaches. Previous methods using LVLMs to predict trajectories or control signals yield suboptimal results, as LVLMs are not well-suited for precise numerical predictions. This paper presents Senna, an autonomous driving system combining an LVLM (Senna-VLM) with an end-to-end model (Senna-E2E). Senna decouples high-level planning from low-level trajectory prediction. Senna-VLM generates planning decisions in natural language, while Senna-E2E predicts precise trajectories. Senna-VLM utilizes a multi-image encoding approach and multi-view prompts for efficient scene understanding. Besides, we introduce planning-oriented QAs alongside a three-stage training strategy, which enhances Senna-VLM's planning performance while preserving commonsense. Extensive experiments on two datasets show that Senna achieves state-of-the-art planning performance. Notably, with pre-training on a large-scale dataset DriveX and fine-tuning on nuScenes, Senna significantly reduces average planning error by 27.12% and collision rate by 33.33% over model without pre-training. We believe Senna's cross-scenario generalization and transferability are essential for achieving fully autonomous driving. Code and models will be released at https://github.com/hustvl/Senna.
Surgical scene segmentation is a fundamental task for robotic-assisted laparoscopic surgery understanding. It often contains various anatomical structures and surgical instruments, where similar local textures and fine-grained structures make the segmentation a difficult task. Vision-specific transformer method is a promising way for surgical scene understanding. However, there are still two main challenges. Firstly, the absence of inner-patch information fusion leads to poor segmentation performance. Secondly, the specific characteristics of anatomy and instruments are not specifically modeled. To tackle the above challenges, we propose a novel Transformer-based framework with an Asymmetric Feature Enhancement module (TAFE), which enhances local information and then actively fuses the improved feature pyramid into the embeddings from transformer encoders by a multi-scale interaction attention strategy. The proposed method outperforms the SOTA methods in several different surgical segmentation tasks and additionally proves its ability of fine-grained structure recognition. Code is available at https://github.com/cyuan-sjtu/ViT-asym.
Multi-modal fusion has played a vital role in multi-modal scene understanding. Most existing methods focus on cross-modal fusion involving two modalities, often overlooking more complex multi-modal fusion, which is essential for real-world applications like autonomous driving, where visible, depth, event, LiDAR, etc., are used. Besides, few attempts for multi-modal fusion, \emph{e.g.}, simple concatenation, cross-modal attention, and token selection, cannot well dig into the intrinsic shared and specific details of multiple modalities. To tackle the challenge, in this paper, we propose a Part-Whole Relational Fusion (PWRF) framework. For the first time, this framework treats multi-modal fusion as part-whole relational fusion. It routes multiple individual part-level modalities to a fused whole-level modality using the part-whole relational routing ability of Capsule Networks (CapsNets). Through this part-whole routing, our PWRF generates modal-shared and modal-specific semantics from the whole-level modal capsules and the routing coefficients, respectively. On top of that, modal-shared and modal-specific details can be employed to solve the issue of multi-modal scene understanding, including synthetic multi-modal segmentation and visible-depth-thermal salient object detection in this paper. Experiments on several datasets demonstrate the superiority of the proposed PWRF framework for multi-modal scene understanding. The source code has been released on https://github.com/liuyi1989/PWRF.
The performance of neural networks scales with both their size and the amount of data they have been trained on. This is shown in both language and image generation. However, this requires scaling-friendly network architectures as well as large-scale datasets. Even though scaling-friendly architectures like transformers have emerged for 3D vision tasks, the GPT-moment of 3D vision remains distant due to the lack of training data. In this paper, we introduce ARKit LabelMaker, the first large-scale, real-world 3D dataset with dense semantic annotations. Specifically, we complement ARKitScenes dataset with dense semantic annotations that are automatically generated at scale. To this end, we extend LabelMaker, a recent automatic annotation pipeline, to serve the needs of large-scale pre-training. This involves extending the pipeline with cutting-edge segmentation models as well as making it robust to the challenges of large-scale processing. Further, we push forward the state-of-the-art performance on ScanNet and ScanNet200 dataset with prevalent 3D semantic segmentation models, demonstrating the efficacy of our generated dataset.
3D visual grounding is crucial for robots, requiring integration of natural language and 3D scene understanding. Traditional methods depending on supervised learning with 3D point clouds are limited by scarce datasets. Recently zero-shot methods leveraging LLMs have been proposed to address the data issue. While effective, these methods only use object-centric information, limiting their ability to handle complex queries. In this work, we present VLM-Grounder, a novel framework using vision-language models (VLMs) for zero-shot 3D visual grounding based solely on 2D images. VLM-Grounder dynamically stitches image sequences, employs a grounding and feedback scheme to find the target object, and uses a multi-view ensemble projection to accurately estimate 3D bounding boxes. Experiments on ScanRefer and Nr3D datasets show VLM-Grounder outperforms previous zero-shot methods, achieving 51.6% [email omitted] on ScanRefer and 48.0% Acc on Nr3D, without relying on 3D geometry or object priors. Codes are available at https://github.com/OpenRobotLab/VLM-Grounder .
Low-light image enhancement (LLIE) is essential for numerous computer vision tasks, including object detection, tracking, segmentation, and scene understanding. Despite substantial research on improving low-quality images captured in underexposed conditions, clear vision remains critical for autonomous vehicles, which often struggle with low-light scenarios, signifying the need for continuous research. However, paired datasets for LLIE are scarce, particularly for street scenes, limiting the development of robust LLIE methods. Despite using advanced transformers and/or diffusion-based models, current LLIE methods struggle in real-world low-light conditions and lack training on street-scene datasets, limiting their effectiveness for autonomous vehicles. To bridge these gaps, we introduce a new dataset LoLI-Street (Low-Light Images of Streets) with 33k paired low-light and well-exposed images from street scenes in developed cities, covering 19k object classes for object detection. LoLI-Street dataset also features 1,000 real low-light test images for testing LLIE models under real-life conditions. Furthermore, we propose a transformer and diffusion-based LLIE model named "TriFuse". Leveraging the LoLI-Street dataset, we train and evaluate our TriFuse and SOTA models to benchmark on our dataset. Comparing various models, our dataset's generalization feasibility is evident in testing across different mainstream datasets by significantly enhancing images and object detection for practical applications in autonomous driving and surveillance systems. The complete code and dataset is available on https://github.com/tanvirnwu/TriFuse.
3D scene understanding is crucial for facilitating seamless interaction between digital devices and the physical world. Real-time capturing and processing of the 3D scene are essential for achieving this seamless integration. While existing approaches typically separate acquisition and processing for each frame, the advent of resolution-scalable 3D sensors offers an opportunity to overcome this paradigm and fully leverage the otherwise wasted acquisition time to initiate processing. In this study, we introduce VX-S3DIS, a novel point cloud dataset accurately simulating the behavior of a resolution-scalable 3D sensor. Additionally, we present RESSCAL3D++, an important improvement over our prior work, RESSCAL3D, by incorporating an update module and processing strategy. By applying our method to the new dataset, we practically demonstrate the potential of joint acquisition and semantic segmentation of 3D point clouds. Our resolution-scalable approach significantly reduces scalability costs from 2% to just 0.2% in mIoU while achieving impressive speed-ups of 15.6 to 63.9% compared to the non-scalable baseline. Furthermore, our scalable approach enables early predictions, with the first one occurring after only 7% of the total inference time of the baseline. The new VX-S3DIS dataset is available at https://github.com/remcoroyen/vx-s3dis.
The task of Visual Sound Source Localization (VSSL) involves identifying the location of sound sources in visual scenes, integrating audio-visual data for enhanced scene understanding. Despite advancements in state-of-the-art (SOTA) models, we observe three critical flaws: i) The evaluation of the models is mainly focused in sounds produced by objects that are visible in the image, ii) The evaluation often assumes a prior knowledge of the size of the sounding object, and iii) No universal threshold for localization in real-world scenarios is established, as previous approaches only consider positive examples without accounting for both positive and negative cases. In this paper, we introduce a novel test set and metrics designed to complete the current standard evaluation of VSSL models by testing them in scenarios where none of the objects in the image corresponds to the audio input, i.e. a negative audio. We consider three types of negative audio: silence, noise and offscreen. Our analysis reveals that numerous SOTA models fail to appropriately adjust their predictions based on audio input, suggesting that these models may not be leveraging audio information as intended. Additionally, we provide a comprehensive analysis of the range of maximum values in the estimated audio-visual similarity maps, in both positive and negative audio cases, and show that most of the models are not discriminative enough, making them unfit to choose a universal threshold appropriate to perform sound localization without any a priori information of the sounding object, that is, object size and visibility.
The autoregressive world model exhibits robust generalization capabilities in vectorized scene understanding but encounters difficulties in deriving actions due to insufficient uncertainty modeling and self-delusion. In this paper, we explore the feasibility of deriving decisions from an autoregressive world model by addressing these challenges through the formulation of multiple probabilistic hypotheses. We propose LatentDriver, a framework models the environment's next states and the ego vehicle's possible actions as a mixture distribution, from which a deterministic control signal is then derived. By incorporating mixture modeling, the stochastic nature of decisionmaking is captured. Additionally, the self-delusion problem is mitigated by providing intermediate actions sampled from a distribution to the world model. Experimental results on the recently released close-loop benchmark Waymax demonstrate that LatentDriver surpasses state-of-the-art reinforcement learning and imitation learning methods, achieving expert-level performance. The code and models will be made available at https://github.com/Sephirex-X/LatentDriver.
The Automated Audio Captioning (AAC) task asks models to generate natural language descriptions of an audio input. Evaluating these machine-generated audio captions is a complex task that requires considering diverse factors, among them, auditory scene understanding, sound-object inference, temporal coherence, and the environmental context of the scene. While current methods focus on specific aspects, they often fail to provide an overall score that aligns well with human judgment. In this work, we propose CLAIR-A, a simple and flexible method that leverages the zero-shot capabilities of large language models (LLMs) to evaluate candidate audio captions by directly asking LLMs for a semantic distance score. In our evaluations, CLAIR-A better predicts human judgements of quality compared to traditional metrics, with a 5.8% relative accuracy improvement compared to the domain-specific FENSE metric and up to 11% over the best general-purpose measure on the Clotho-Eval dataset. Moreover, CLAIR-A offers more transparency by allowing the language model to explain the reasoning behind its scores, with these explanations rated up to 30% better by human evaluators than those provided by baseline methods. CLAIR-A is made publicly available at https://github.com/DavidMChan/clair-a.
We propose Hier-SLAM, a semantic 3D Gaussian Splatting SLAM method featuring a novel hierarchical categorical representation, which enables accurate global 3D semantic mapping, scaling-up capability, and explicit semantic label prediction in the 3D world. The parameter usage in semantic SLAM systems increases significantly with the growing complexity of the environment, making it particularly challenging and costly for scene understanding. To address this problem, we introduce a novel hierarchical representation that encodes semantic information in a compact form into 3D Gaussian Splatting, leveraging the capabilities of large language models (LLMs). We further introduce a novel semantic loss designed to optimize hierarchical semantic information through both inter-level and cross-level optimization. Furthermore, we enhance the whole SLAM system, resulting in improved tracking and mapping performance. Our \MethodName{} outperforms existing dense SLAM methods in both mapping and tracking accuracy, while achieving a 2x operation speed-up. Additionally, it achieves on-par semantic rendering performance compared to existing methods while significantly reducing storage and training time requirements. Rendering FPS impressively reaches 2,000 with semantic information and 3,000 without it. Most notably, it showcases the capability of handling the complex real-world scene with more than 500 semantic classes, highlighting its valuable scaling-up capability. The open-source code is available at https://github.com/LeeBY68/Hier-SLAM
Infrared and visible image fusion aims to combine complementary information from both modalities to provide a more comprehensive scene understanding. However, due to the significant differences between the two modalities, preserving key features during the fusion process remains a challenge. To address this issue, we propose a dual-branch feature decomposition fusion network (DAF-Net) with domain adaptive, which introduces Multi-Kernel Maximum Mean Discrepancy (MK-MMD) into the base encoder and designs a hybrid kernel function suitable for infrared and visible image fusion. The base encoder built on the Restormer network captures global structural information while the detail encoder based on Invertible Neural Networks (INN) focuses on extracting detail texture information. By incorporating MK-MMD, the DAF-Net effectively aligns the latent feature spaces of visible and infrared images, thereby improving the quality of the fused images. Experimental results demonstrate that the proposed method outperforms existing techniques across multiple datasets, significantly enhancing both visual quality and fusion performance. The related Python code is available at https://github.com/xujian000/DAF-Net.
Re-identification (ReID) is a critical challenge in computer vision, predominantly studied in the context of pedestrians and vehicles. However, robust object-instance ReID, which has significant implications for tasks such as autonomous exploration, long-term perception, and scene understanding, remains underexplored. In this work, we address this gap by proposing a novel dual-path object-instance re-identification transformer architecture that integrates multimodal RGB and depth information. By leveraging depth data, we demonstrate improvements in ReID across scenes that are cluttered or have varying illumination conditions. Additionally, we develop a ReID-based localization framework that enables accurate camera localization and pose identification across different viewpoints. We validate our methods using two custom-built RGB-D datasets, as well as multiple sequences from the open-source TUM RGB-D datasets. Our approach demonstrates significant improvements in both object instance ReID (mAP of 75.18) and localization accuracy (success rate of 83% on TUM-RGBD), highlighting the essential role of object ReID in advancing robotic perception. Our models, frameworks, and datasets have been made publicly available.
This work addresses the task of zero-shot monocular depth estimation. A recent advance in this field has been the idea of utilising Text-to-Image foundation models, such as Stable Diffusion. Foundation models provide a rich and generic image representation, and therefore, little training data is required to reformulate them as a depth estimation model that predicts highly-detailed depth maps and has good generalisation capabilities. However, the realisation of this idea has so far led to approaches which are, unfortunately, highly inefficient at test-time due to the underlying iterative denoising process. In this work, we propose a different realisation of this idea and present PrimeDepth, a method that is highly efficient at test time while keeping, or even enhancing, the positive aspects of diffusion-based approaches. Our key idea is to extract from Stable Diffusion a rich, but frozen, image representation by running a single denoising step. This representation, we term preimage, is then fed into a refiner network with an architectural inductive bias, before entering the downstream task. We validate experimentally that PrimeDepth is two orders of magnitude faster than the leading diffusion-based method, Marigold, while being more robust for challenging scenarios and quantitatively marginally superior. Thereby, we reduce the gap to the currently leading data-driven approach, Depth Anything, which is still quantitatively superior, but predicts less detailed depth maps and requires 20 times more labelled data. Due to the complementary nature of our approach, even a simple averaging between PrimeDepth and Depth Anything predictions can improve upon both methods and sets a new state-of-the-art in zero-shot monocular depth estimation. In future, data-driven approaches may also benefit from integrating our preimage.
Nighttime camera-based depth estimation is a highly challenging task, especially for autonomous driving applications, where accurate depth perception is essential for ensuring safe navigation. We aim to improve the reliability of perception systems at night time, where models trained on daytime data often fail in the absence of precise but costly LiDAR sensors. In this work, we introduce Light Enhanced Depth (LED), a novel cost-effective approach that significantly improves depth estimation in low-light environments by harnessing a pattern projected by high definition headlights available in modern vehicles. LED leads to significant performance boosts across multiple depth-estimation architectures (encoder-decoder, Adabins, DepthFormer) both on synthetic and real datasets. Furthermore, increased performances beyond illuminated areas reveal a holistic enhancement in scene understanding. Finally, we release the Nighttime Synthetic Drive Dataset, a new synthetic and photo-realistic nighttime dataset, which comprises 49,990 comprehensively annotated images.
3D point clouds enhanced the robot's ability to perceive the geometrical information of the environments, making it possible for many downstream tasks such as grasp pose detection and scene understanding. The performance of these tasks, though, heavily relies on the quality of data input, as incomplete can lead to poor results and failure cases. Recent training loss functions designed for deep learning-based point cloud completion, such as Chamfer distance (CD) and its variants (\eg HyperCD ), imply a good gradient weighting scheme can significantly boost performance. However, these CD-based loss functions usually require data-related parameter tuning, which can be time-consuming for data-extensive tasks. To address this issue, we aim to find a family of weighted training losses ({\em weighted CD}) that requires no parameter tuning. To this end, we propose a search scheme, {\em Loss Distillation via Gradient Matching}, to find good candidate loss functions by mimicking the learning behavior in backpropagation between HyperCD and weighted CD. Once this is done, we propose a novel bilevel optimization formula to train the backbone network based on the weighted CD loss. We observe that: (1) with proper weighted functions, the weighted CD can always achieve similar performance to HyperCD, and (2) the Landau weighted CD, namely {\em Landau CD}, can outperform HyperCD for point cloud completion and lead to new state-of-the-art results on several benchmark datasets. {\it Our demo code is available at \url{https://github.com/Zhang-VISLab/IROS2024-LossDistillationWeightedCD}.}
3D scene reconstruction from stereo endoscopic video data is crucial for advancing surgical interventions. In this work, we present an online framework for online, dense 3D scene reconstruction and tracking, aimed at enhancing surgical scene understanding and assisting interventions. Our method dynamically extends a canonical scene representation using Gaussian splatting, while modeling tissue deformations through a sparse set of control points. We introduce an efficient online fitting algorithm that optimizes the scene parameters, enabling consistent tracking and accurate reconstruction. Through experiments on the StereoMIS dataset, we demonstrate the effectiveness of our approach, outperforming state-of-the-art tracking methods and achieving comparable performance to offline reconstruction techniques. Our work enables various downstream applications thus contributing to advancing the capabilities of surgical assistance systems.
Scene observation from multiple perspectives would bring a more comprehensive visual experience. However, in the context of acquiring multiple views in the dark, the highly correlated views are seriously alienated, making it challenging to improve scene understanding with auxiliary views. Recent single image-based enhancement methods may not be able to provide consistently desirable restoration performance for all views due to the ignorance of potential feature correspondence among different views. To alleviate this issue, we make the first attempt to investigate multi-view low-light image enhancement. First, we construct a new dataset called Multi-View Low-light Triplets (MVLT), including 1,860 pairs of triple images with large illumination ranges and wide noise distribution. Each triplet is equipped with three different viewpoints towards the same scene. Second, we propose a deep multi-view enhancement framework based on the Recurrent Collaborative Network (RCNet). Specifically, in order to benefit from similar texture correspondence across different views, we design the recurrent feature enhancement, alignment and fusion (ReEAF) module, in which intra-view feature enhancement (Intra-view EN) followed by inter-view feature alignment and fusion (Inter-view AF) is performed to model the intra-view and inter-view feature propagation sequentially via multi-view collaboration. In addition, two different modules from enhancement to alignment (E2A) and from alignment to enhancement (A2E) are developed to enable the interactions between Intra-view EN and Inter-view AF, which explicitly utilize attentive feature weighting and sampling for enhancement and alignment, respectively. Experimental results demonstrate that our RCNet significantly outperforms other state-of-the-art methods. All of our dataset, code, and model will be available at https://github.com/hluo29/RCNet.
Panoptic 3D reconstruction from a monocular video is a fundamental perceptual task in robotic scene understanding. However, existing efforts suffer from inefficiency in terms of inference speed and accuracy, limiting their practical applicability. We present EPRecon, an efficient real-time panoptic 3D reconstruction framework. Current volumetric-based reconstruction methods usually utilize multi-view depth map fusion to obtain scene depth priors, which is time-consuming and poses challenges to real-time scene reconstruction. To end this, we propose a lightweight module to directly estimate scene depth priors in a 3D volume for reconstruction quality improvement by generating occupancy probabilities of all voxels. In addition, to infer richer panoptic features from occupied voxels, EPRecon extracts panoptic features from both voxel features and corresponding image features, obtaining more detailed and comprehensive instance-level semantic information and achieving more accurate segmentation results. Experimental results on the ScanNetV2 dataset demonstrate the superiority of EPRecon over current state-of-the-art methods in terms of both panoptic 3D reconstruction quality and real-time inference. Code is available at https://github.com/zhen6618/EPRecon.
Shadows are created when light encounters obstacles, resulting in regions of reduced illumination. In computer vision, detecting, removing, and generating shadows are critical tasks for improving scene understanding, enhancing image quality, ensuring visual consistency in video editing, and optimizing virtual environments. This paper offers a comprehensive survey and evaluation benchmark on shadow detection, removal, and generation in both images and videos, focusing on the deep learning approaches of the past decade. It covers key aspects such as tasks, deep models, datasets, evaluation metrics, and comparative results under consistent experimental settings. Our main contributions include a thorough survey of shadow analysis, the standardization of experimental comparisons, an exploration of the relationships between model size, speed, and performance, a cross-dataset generalization study, the identification of open challenges and future research directions, and the provision of publicly available resources to support further research in this field.