Foundation models have significantly enhanced 2D task performance, and recent works like Bridge3D have successfully applied these models to improve 3D scene understanding through knowledge distillation, marking considerable advancements. Nonetheless, challenges such as the misalignment between 2D and 3D representations and the persistent long-tail distribution in 3D datasets still restrict the effectiveness of knowledge distillation from 2D to 3D using foundation models. To tackle these issues, we introduce a novel SAM-guided tokenization method that seamlessly aligns 3D transformer structures with region-level knowledge distillation, replacing the traditional KNN-based tokenization techniques. Additionally, we implement a group-balanced re-weighting strategy to effectively address the long-tail problem in knowledge distillation. Furthermore, inspired by the recent success of masked feature prediction, our framework incorporates a two-stage masked token prediction process in which the student model predicts both the global embeddings and token-wise local embeddings derived from the teacher models trained in the first stage. Our methodology has been validated across multiple datasets, including SUN RGB-D, ScanNet, and S3DIS, for tasks like 3D object detection and semantic segmentation. The results demonstrate significant improvements over current state-of-the-art self-supervised methods, establishing new benchmarks in this field.

Current learning-based 3-D object detection accuracy is heavily impacted by the annotation quality. It is still a challenge to expect an overall high detection accuracy for all classes under different scenarios given the dataset sparsity. To mitigate this challenge, this article proposes a novel method called semi-supervised learning and progressive distillation (SPD), which uses semi-supervised learning (SSL) and knowledge distillation to improve label efficiency. The SPD uses two big backbones to hand the unlabeled/labeled input data augmented by the periodic IO augmentation (PA). Then the backbones are compressed using progressive distillation (PD). Precisely, PA periodically shifts the data augmentation operations between the input and output of the big backbone, aiming to improve the network's generalization of the unseen and unlabeled data. Using the big backbone can benefit from large-scale augmented data better than the small one. And two backbones are trained by the data scale and ratio-sensitive loss (data-loss). It solves the over-flat caused by the large-scale unlabeled data from PA and helps the big backbone prevent overfitting on the limited-scale labeled data. Hence, using the PA and data loss during SSL training dramatically improves the label efficiency. Next, the trained big backbone set as the teacher CNN is progressively distilled to obtain a small student model, referenced as PD. PD mitigates the problem that student CNN performance degrades when the gap between the student and the teacher is oversized. Extensive experiments are conducted on the indoor datasets SUN RGB-D and ScanNetV2 and outdoor dataset KITTI. Using only 50% labeled data and a 27% smaller model size, SPD performs 0.32 higher than the fully supervised VoteNet [1] which is adopted as our backbone. Besides, using only 2% labeled data, compared to the other fully supervised backbone PV-RCNN [2], SPD accomplishes a similar accuracy (84.1 and 84.83) and 30% less inference time.

This research proposes a novel semi-supervised learning framework FourStr (Four-Stream formed by two two-stream models) that focuses on the improvement of fusion and labeling efficiency for 3D multi-sensor detector. FourStr adopts a multi-sensor single-stage detector named adaptive fusion network (AFNet) as the backbone and trains it through the semi-supervision learning (SSL) strategy Stereo Fusion. Note that multi-sensor AFNet and SSL Stereo Fusion can benefit each other. On the one hand, the Four-stream composed of two AFNets naturally provides rich inputs and large models for SSL Stereo Fusion. While other SSL works have to use massive augmentation to obtain rich inputs, and deepen and widen the network for large models. On the other hand, by the novel three fusion stages and Loss Pruning, Stereo Fusion improves the fusion and labeling efficiency for AFNet. Finally, extensive experiments demonstrate that FourStr performs excellently on outdoor dataset (KITTI and Waymo Open Dataset) and indoor dataset (SUN RGB-D), especially for the small contour objects. And compared to the fully-supervised methods, FourStr achieves similar accuracy with only 2% labeled data on KITTI (or with 50% labeled data on SUN RGB-D).

Foundation models have achieved remarkable results in 2D and language tasks like image segmentation, object detection, and visual-language understanding. However, their potential to enrich 3D scene representation learning is largely untapped due to the existence of the domain gap. In this work, we propose an innovative methodology called Bridge3D to address this gap by pre-training 3D models using features, semantic masks, and captions sourced from foundation models. Specifically, our method employs semantic masks from foundation models to guide the masking and reconstruction process for the masked autoencoder, enabling more focused attention on foreground representations. Moreover, we bridge the 3D-text gap at the scene level using image captioning foundation models, thereby facilitating scene-level knowledge distillation. We further extend this bridging effort by introducing an innovative object-level knowledge distillation method that harnesses highly accurate object-level masks and semantic text data from foundation models. Our methodology significantly surpasses the performance of existing state-of-the-art methods in 3D object detection and semantic segmentation tasks. For instance, on the ScanNet dataset, Bridge3D improves the baseline by a notable margin of 6.3%. Code will be available at: https://github.com/Zhimin-C/Bridge3D

Recent state-of-the-art method FlexMatch firstly demonstrated that correctly estimating learning status is crucial for semi-supervised learning (SSL). However, the estimation method proposed by FlexMatch does not take into account imbalanced data, which is the common case for 3D semi-supervised learning. To address this problem, we practically demonstrate that unlabeled data class-level confidence can represent the learning status in the 3D imbalanced dataset. Based on this finding, we present a novel class-level confidence based 3D SSL method. Firstly, a dynamic thresholding strategy is proposed to utilize more unlabeled data, especially for low learning status classes. Then, a re-sampling strategy is designed to avoid biasing toward high learning status classes, which dynamically changes the sampling probability of each class. To show the effectiveness of our method in 3D SSL tasks, we conduct extensive experiments on 3D SSL classification and detection tasks. Our method significantly outperforms state-of-the-art counterparts for both 3D SSL classification and detection tasks in all datasets.

One of the major components of the neural network, the feature pyramid plays a vital part in perception tasks, like object detection in autonomous driving. But it is a challenge to fuse multi-level and multi-sensor feature pyramids for object detection. This paper proposes a simple yet effective framework namedMuTrans(MultipleTransformers) to fuse feature pyramid in single-stream 2D detector or two-stream 3D detector. The MuTrans based on encoder-decoder focuses on the significant features via multiple Transformers. MuTrans encoder uses three innovative self-attention mechanisms:Spatial-wiseBoxAlign attention (SB) for low-level spatial locations,Context-wiseAffinity attention (CA) for high-level context information, and high-level attention for multi-level features. Then MuTrans decoder processes these significant proposals including the RoI and context affinity. Besides, theLow andHigh-levelFusion (LHF) in the encoder reduces the number of computational parameters. And the Pre-LN [1] is utilized to accelerate the training convergence. LHF and Pre-LN are proven to reduce self-attention’s computational complexity and slow training convergence. Our result demonstrates the higher detection accuracy of MuTrans than that of the baseline method, particularly in small object detection. MuTrans demonstrates a 2.1 higher detection accuracy onAPSindex in small object detection on MS-COCO 2017 with ResNeXt-101 backbone, a 2.18 higher 3D detection accuracy (moderate difficulty) for small object-pedestrian on KITTI, and 6.85 higher RC index (Town05 Long) on CARLA urban driving simulator platform.

Recently significant progress has been made in 3D detection. However, it is still challenging to detect small contour objects under complex scenes. This paper proposes a novel Attention-based Multi-phase Multi-task Fusion (AMMF) that uses point-level, RoI-level, and multi-task fusions to complement the disadvantages of LiDAR and camera, to solve this challenge. First, at the feature extraction phase, AMMF uses the Low and High-level Fusion with Matching Attention (LHF-MA) and efficient FPN (eFPN) to perform point-level fusion for cross sensors and single sensor, respectively. Instead of merging each level and using expensive 3D CNN like other methods, LHF-MA fuses low-level spatial location and high-level contextual feature of 2D CNN customized feature extractors and ignores the fusion of middle levels, reducing the computational cost. Then, at the proposal generation phase, Progressive Proposal Fusion (PPF) with learned attention map is used to perform coarse-to-fine RoI-level fusion, instead of only combining coarse-grained features at high-level of network. PPF using progressively increasing IoU thresholds could avoid overfitting and improve the performance. Note that the matching attentions and learned attention maps are utilized to weigh the priority of different sensors. Moreover, to solve the sparseness of point-wise fusion between LiDAR BEV and RGB image, AMMF uses multi-task fusion that generates pseudo-LiDAR from camera by depth estimation task, to guide this point-wise fusion. Finally, AMMF performs excellently for detecting small contour objects like pedestrians, cyclists, and distant cars. On the KITTI, AMMF finishes 3.62% improvements in the moderate instance for pedestrians. It achieves a 2.21% improvement in the$>$50 instance of LEVEL$_{-}$2 level for vehicle on the Waymo Open Dataset. And AMMF is further verified on our customized dataset consisting of challenging scenarios like strong illumination and heavy shadow cases.

The feature pyramid, which is a vital component of the convolutional neural networks, plays a significant role in several perception tasks, including object detection for autonomous driving. However, how to better fuse multi-level and multi-sensor feature pyramids is still a significant challenge, especially for object detection. This paper presents a FocusTR (Focusing on the valuable features by multiple Transformers), which is a simple yet effective architecture, to fuse feature pyramid for the single-stream 2D detector and two-stream 3D detector. Specifically, FocusTR encompasses several novel self-attention mechanisms, including the spatial-wise boxAlign attention (SB) for low-level spatial locations, context-wise affinity attention (CA) for high-level context information, and level-wise attention for the multi-level feature. To alleviate self-attention's computational complexity and slow training convergence, Fo-cusTR introduces a low and high-level fusion (LHF) to reduce the computational parameters, and the Pre- Ln [1]to accelerate the training convergence.

Assistive robots can be found in hospitals and rehabilitation clinics, where they help patients maintain a positive disposition. Our proposed robotic mobility solution combines state of the art hardware and software to provide a safer, more independent, and more productive lifestyle for people with some of the most severe disabilities. New hardware includes, a retractable roof, manipulator arm, a hard backpack, a number of sensors that collect environmental data and processors that generate 3D maps for a hands-free human-machine interface.The proposed new system receives input from the user via head tracking or voice command, and displays information through augmented reality into the user’s field of view. The software algorithm will use a novel cycle of self-learning artificial intelligence that achieves autonomous navigation while avoiding collisions with stationary and dynamic objects. The prototype will be assembled and tested over the next three years and a publicly available version could be ready two years thereafter.

Assistive navigation for blind or visually impaired (BVI) individuals is of significance to extend their mobility and safety in traveling, enhancing their employment opportunities and fostering personal fulfillment. Conventional research is mainly based on robotic navigation approaches through localization, mapping, and path planning frameworks. They require heavy manual annotation of semantic information in maps and its alignment with sensor mapping. Inspired by the fact that we human beings naturally rely on language instruction inquiry and visual scene understanding to navigate in an unfamiliar environment, this paper proposes a novel vision-language model-based approach for BVI navigation. It does not need heavy-labeled indoor maps and provides a Safe and Efficient E-Wayfinding (SeeWay) assistive solution for BVI individuals. The system consists of a scene-graph map construction module, a navigation path generation module for global path inference by vision-language navigation (VLN), and a navigation with obstacle avoidance module for real-time local navigation. The SeeWay system was deployed on portable iPhone devices with cloud computing assistance for the VLN model inference. The field tests show the effectiveness of the VLN global path finding and local path re-planning. Experiments and quantitative results reveal that heuristic-style instruction outperforms direction/detailed-style instructions for VLN success rate (SR), and the SR decreases as the navigation length increases.

Traditional manual wheelchairs have a fixed seat with no movement or angle adjustment, which can seriously affect the user's comfort and greatly limit user experience. However, the electric wheelchair relies on strong intelligence and automatic features; it can not only realize the multidegree freedom adjustment of the human body and the seat but also has a rich and powerful man–machine control interface, which greatly facilitates and improves the user experience. This study upgraded a Permobil C400-powered wheelchair with multisensor data fusion technology to enrich its terrain recognition, tipping stability, and comfortability prediction. The tipping stability modeling of the wheelchair dummy system is carried out using multibody dynamics and vibration mechanics to obtain the tipping stability limit and the comfort evaluation of the wheelchair vibration acceleration on the human body during travel. Based on the elevation mapping method, the wheelchair can estimate the terrain from the local point of view at any point in time. At the same time, the RGB-D depth camera is connected to the robot operating system (ROS) system, and the open-source algorithm package RTAB-MAP is used to complete the MAP construction and collect the 3-D point-cloud terrain data. Then, the real 3-D terrain files are generated through the point-cloud stitching technology for stability simulation of the wheelchair–human system. The tipping stability and comfort indexes of the wheelchair–human system when passing over different physical terrains can be obtained. The experimental results show that the IMU data located on the human chest agree well with the simulation analysis data and are suitable for a variety of complex real-terrain conditions, verifying the accuracy of the wheelchair–human system dynamics model and the feasibility of the simulation analysis process. Thus, this modeling and simulation method can predict wheelchair stability and user comfortability well and ensure a high-performance experience.

This paper proposes a real-time multi-scale semantic segmentation network (MsNet). MsNet is a combination of our novel multi-scale fusion with matching attention model (MFMA) as the decoding network and the network searched by asymptotic neural architecture search (ANAS) or MobileNetV3 as the encoding network. The MFMA not only extracts low level spatial features from multi-scale inputs but also decodes the contextual features extracted by ANAS. Specifically, considering the advantages and disadvantages of the addition fusion and concatenation fusion, we design multi-scale fusion (MF) that balances speed and accuracy. Then we creatively design two matching attention mechanisms (MA), including matching attention with low calculation (MALC) mechanism and matching attention with strong global context modeling (MASG) mechanism, to match varying resolutions and information of features at different levels of a network. Besides, the ANAS performs the deep neural network search by employing an asymptotic method and provide an efficient encoding network for MsNet, releasing researchers from those tedious mechanical trials. Through extensive experiments, we prove that MFMA, which can be applied to numerous recognition tasks, possesses excellent decoding ability. And we demonstrate the effectiveness and necessity of implementing the “matching” attention mechanism. Finally, the proposed two versions, MsNet_ANAS and MsNet_M achieve a new state-of-the-art trade-off between accuracy and speed on the CamVid and Cityscapes datasets. More remarkably, on the Nvidia Tesla V100 GPU, our MsNet_ANAS achieves 74.1% mIoU with the speed of 184.2 FPS on the CamVid while 72.9% mIoU with the speed of 119.9 FPS on the Cityscapes.

Accurate localization on the autonomous driving cars is essential for autonomy and driving safety, especially for complex urban streets and search-and-rescue subterranean environments where high-accurate GPS is not available. However current odometry estimation may introduce the drifting problems in long-term navigation without robust global localization. The main challenges involve scene divergence under the interference of dynamic environments and effective perception of observation and object layout variance from different viewpoints. To tackle these challenges, we present PSE-Match, a viewpoint-free place recognition method based on parallel semantic analysis of isolated semantic attributes from 3D point-cloud models. Compared with the original point cloud, the observed variance of semantic attributes is smaller. PSE-Match incorporates a divergence place learning network to capture different semantic attributes parallelly through the spherical harmonics domain. Using both existing benchmark datasets and two in-field collected datasets, our experiments show that the proposed method achieves above 70% average recall with top one retrieval and above 95% average recall with top ten retrieval cases. And PSE-Match has also demonstrated an obvious generalization ability with limited training dataset.

This paper presents a novel metric inspection robot system using a deep neural network to detect and measure surface flaws (i.e., crack and spalling) on concrete structures performed by a wall-climbing robot. The system consists of four modules: robotics data collection module to obtain RGB-D images and IMU measurement, visual-inertial SLAM module to generate pose coupled key-frames with depth information, InspectionNet module to classify each pixel into three classes (back-ground, crack and spalling), and 3D registration and map fusion module to register the flaw patch into registered 3D model overlaid and highlighted with detected flaws for spatial-contextual visualization. The system enables the metric model of each surface flaw patch with pixel-level accuracy and determines its location in 3D space that is significant for structural health assessment and monitoring. The InspectionNet achieves an average accuracy of 87.64% for crack and spalling inspection. We also demonstrate our InspectionNet is robust to view angle, scale and illumination variation. Finally, we design a metric voxel volume map to highlight the flaw in 3D model and provide location and metric information.

This paper presents a new holistic vision-based mobile assistive navigation system to help blind and visually impaired people with indoor independent travel. The system detects dynamic obstacles and adjusts path planning in real-time to improve navigation safety. First, we develop an indoor map editor to parse geometric information from architectural models and generate a semantic map consisting of a global 2D traversable grid map layer and context-aware layers. By leveraging the visual positioning service (VPS) within the Google Tango device, we design a map alignment algorithm to bridge the visual area description file (ADF) and semantic map to achieve semantic localization. Using the on-board RGB-D camera, we develop an efficient obstacle detection and avoidance approach based on a time-stamped map Kalman filter (TSM-KF) algorithm. A multi-modal human-machine interface (HMI) is designed with speech-audio interaction and robust haptic interaction through an electronic SmartCane. Finally, field experiments by blindfolded and blind subjects demonstrate that the proposed system provides an effective tool to help blind individuals with indoor navigation and wayfinding.

We present a novel collaborative mapping and autonomous parking system for semi-structured multi-story parking garages, based on cooperative 3-D LiDAR point cloud registration and Bayesian probabilistic updating. First, an inertial-enhanced (IE) generalized iterative closest point (G-ICP) approach is presented to perform high accuracy registration for LiDAR odometry, which is loosely coupled with inertial measurement unit using multi-state extended Kalman filter fusion. Second, the IE G-ICP is utilized to reconstruct the 3-D point cloud model for each vehicle, and then the individual model maps are merged and updated into a global probabilistic 2-D grid map. A collaborative multiple layer semantic map is constructed to support autonomous parking. Finally, we propose a collaborative navigation approach for path planning when there are multiple vehicles in the parking garage through vehicle-to-vehicle communication. A global path planner is designed to explore the minimum cost path based on the semantic map, and local motion planning is performed using a random exploring algorithm for obstacle avoidance and path smoothing. Our pilot experimental evaluation provides a proof of concept for indoor autonomous parking by collaborative perception, map merging, and updating methodologies.

This paper proposes a novel assistive navigation system based on simultaneous localization and mapping (SLAM) and semantic path planning to help visually impaired users navigate in indoor environments. The system integrates multiple wearable sensors and feedback devices including a RGB-D sensor and an inertial measurement unit (IMU) on the waist, a head mounted camera, a microphone and an earplug/speaker. We develop a visual odometry algorithm based on RGB-D data to estimate the user's position and orientation, and refine the orientation error using the IMU. We employ the head mounted camera to recognize the door numbers and the RGB-D sensor to detect major landmarks such as corridor corners. By matching the detected landmarks against the corresponding features on the digitalized floor map, the system localizes the user, and provides verbal instruction to guide the user to the desired destination. The software modules of our system are implemented in Robotics Operating System (ROS). The prototype of the proposed assistive navigation system is evaluated by blindfolded sight persons. The field tests confirm the feasibility of the proposed algorithms and the system prototype.

This paper provides a framework for context-aware navigation services for vision impaired people. Integrating advanced intelligence into navigation requires knowledge of the semantic properties of the objects around the user's environment. This interaction is required to enhance communication about objects and places to improve travel decisions. Our intelligent system is a human-in-the-loop cyber-physical system that interprets ubiquitous semantic entities by interacting with the physical world and the cyber domain, viz., 1) visual cues and distance sensing of material objects as line-of-sight interaction to interpret location-context information, and 2) data (tweets) from social media as event-based interaction to interpret situational vibes. The case study elaborates our proposed localization methods (viz., topological, landmark, metric, crowdsourced, and sound localization) for applications in way finding, way confirmation, user tracking, socialization, and situation alerts. Our pilot evaluation provides a proof of concept for an assistive navigation system.