Consistent Mapping Research Articles

Robust Collaborative Visual-Inertial SLAM for Mobile Augmented Reality.

Achieving precise real-time localization and ensuring robustness are critical challenges in multi-user mobile AR applications. Leveraging collaborative information to augment tracking accuracy on lightweight devices and fortify overall system robustness emerges as a crucial necessity. In this paper, we propose a robust centralized collaborative rnulti-agent VI-SLAM system for mobile AR interaction and server-side efficient consistent mapping. The system deploys a lightweight VIO frontend on mobile devices for real-time tracking, and a backend running on a remote server to update multiple submaps. When overlapping areas between submaps across agents are detected, the system performs submap fusion to establish a globally consistent map. Additionally, we propose a map registration and fusion strategy based on covisibility areas for online registration and fusion in multi-agent scenarios. To improve the tracking accuracy of the frontend on agent, we introduce a strategy for updating the global map to the local map at a moderate frequency between the camera-rate pose estimation of the frontend VIO and the low-frequency global map optimization, using a tightly coupled strategy to achieve consistency of the multi-agent frontend poses estimation in the global map. The effectiveness of the proposed method is further confirmed by executing backend mapping on the server and deploying VIO frontends on multiple mobile devices for AR demostration. Additionally, we discuss the scalability of the proposed system by analyzing network traffic, synchronization frequency, and other factors at both the agent and server ends.

Reconstructing urban vegetation evolution in China using multimodal deep learning and 30-years Landsat archive

To retrospect the impact of urbanization on vegetation, consistent and reliable maps that provide urban vegetation estimates are required. In this study, we make the first attempt to employ deep learning for urban vegetation mapping of 2120 cities in China using a 30-years Landsat archive (UV-30). We present a multimodal deep learning (MDL) model to combine features from 5 categories, namely reflective bands, remote sensing indices, texture variables, topographical variables, and time labels. We normalized elevation data locally to remove the inter-city differences, while retaining topographic details. We used time labels to account for spectral variability over time and Landsat sensors. The performance of the MDL models was compared with a vanilla deep neural network, random forest, and support vector regression. Results indicated that the fusion of different categories of features improved the prediction accuracy. The MDL performed best in most validation cities, and our results were superior to the fractional maps aggregated from high-resolution datasets in homogeneous urban landscapes. We further analyzed urban vegetation dynamics across the urban core-new town-outskirt gradients of different geographical regions and city sizes. We found a gradually diminishing de-greening trend in urban China, especially in urban cores and new towns. Nevertheless, we observed a more pronounced degradation of urban vegetation in the south regions of China compared to the north, and smaller cities show no evident urban vegetation increase. The suggested framework is valuable for urban vegetation assessments and facilitates a thorough understanding and management of urban ecosystems.

ZnO thin films with stable, tunable electrical and optical properties deposited by atomic layer deposition using Et2Zn:NEtMe2 precursor

ZnO thin films were deposited via atomic layer deposition (ALD), using Zn precursor, DEZDMEA (Et2Zn:NEtMe2) and H2O as a reactant. The design of DEZDMEA aimed to achieve enhanced stability and reduced decomposition compared to traditional DEZ (Et2Zn) precursor, as verified through NMR analysis, leading to the ZnO thin films with no-detectable impurities. The growth characteristics of ZnO thin films with the DEZDMEA were systematically evaluated, and high growth rate and the self-limiting reaction were observed. The comparable ZnO growth rate using DEZDMEA with those using DEZ was shown due to the selective removal of one Et and NEtMe2 ligands during the surface adsorption of DEZDMEA as established by DFT calculations. Additionally, increasing the precursor pulse and reducing the H2O pulse time resulted in decreased resistivity of the ZnO thin films, due to the modulation of n-type defects. This approach provided a consistent contour map for achieving stable, reproducible, and tailored resistivity depending on the DEZDMEA pulse time, and H2O pulse time with its vapor pressure. Moreover, the optical band gap could be modulated as well. The utilization of DEZDMEA is anticipated to be a robust and effective method in ALD for Zn-related deposition, promising substantial advantages for commercial applications.

Advanced mapping and ablation technologies to target pulmonary veins gaps and residual antral potentials during repeat paroxysmal AF ablation: acute and long-term follow-up

Abstract Background A comprehensive understanding of pulmonary vein (PV) reconnection and residual potentials (RAP) within the antral scar during repeat atrial fibrillation (AF) ablation, facilitated by advanced mapping and ablation technologies, is currently lacking. Optimizing radiofrequency (RF) deliveries after failed PV isolation (PVI) may enhance ablation efficiency. Purpose This study aims to assess both the acute and long-term outcomes focusing on PV gaps (PVGs) and RAPs, utilizing advanced mapping and ablation technologies during repeat PV ablation in patients (pts) with paroxysmal AF. Methods Repeat AF-CHARISMA study, a prospective, single-arm cohort study, included consecutive pts from 13 centers undergoing repeat ablation of paroxysmal AF. The Lumipoint (Boston Scientific) map-analysis tool was used sequentially on each PV component. All patients were followed-up for at least 12 months after the procedure. The ablation endpoint was PVI and electrical quiescence in the antral region. Long-term outcome was the recurrence of AT/AF over follow-up. Results One-hundred pts were included (69% after failed RF procedure, 31% after failed cryoablation procedure). A total of 276 PVGs (192 after RF ablation, 2.8±1.8/pt; 84 after cryoablation, 2.7±1.9/pt, p=0.665 vs RF) and 44 RAPs (43 after RF ablation, 0.6±1.1/pt; 1 after cryoablation, 0.03±0.2/pt, p=0.002 vs RF) were detected. In the majority of cases (75%), only PVGs were present whereas, in 21%, both PVGs and RAPs were detected (2 pts had only RAPs). In 2 pts no PVG nor RAP were identified. PVGs were most common at anterior sites (103, 37.3%), followed by posterior sites (75, 27.2%) and carina sites (49, 17.8%). PVG location sites did not differ between prior ablation approach. Only 1 pt after cryoablation showed a RAP (at carina site), whereas the 44 RAPs after RF were fairly distributed across different sites. Acute procedural success was 100%, with all PVs successfully isolated and RAPs completely abolished in all study pts. After the blanking period, over a follow-up of 480±211 days, 18 pts (18%) suffered an AT/AF recurrence (time to recurrence: 245±80 days). Previous ablation approach (HR=0.66, 95%CI: 0.2 to 2.0 for previous cryoablation, p=0.46), number of PVGs (1.05, 0.8 to 1.3, p=0.67) or number of RAPs (0.41, 0.1 to 1.2, p=0.12) were not associated to AT/AF recurrence. No major complications or adverse events occurred. Conclusion A structured ablation workflow for repeat catheter ablation, incorporating consistent high-density mapping and improved diagnostic tools to guide the procedure, demonstrated safety and efficacy in treating paroxysmal AF irrespectively of previous ablation approach. Residual potentials within the antral scar were more frequently detected after RF ablation compared to cryoablation.

Efficient Multi-floor Indoor Mapping with A Versatile Rotating LiDAR System

Abstract. Light Detection And Ranging (LiDAR) technology has provided an impactful way to capture 3D environmental map. However, consistent mapping in sensing-degenerated and perceptually-limited scenes (e.g. multi-story buildings) or under high dynamic sensor motion (e.g. rotating platform) remains a significant challenge. To this end, an efficient multi-floor indoor mapping system is proposed in this paper utilizing a versatile rotating LiDAR platform. In the front end, measurements from motor, IMU and LiDAR are tightly integrated to track the fast motion state of system, based on iterative Error State Kalman Filter (ESKF). Then linear and planar features are extracted from the point cloud map voxelized with adaptive resolutions. A sliding-window-based batch optimization is performed to simultaneously optimize the states of local frames with reference to the map consistency. In the experiments, we investigated the influence of rotating speed on the mapping performance as well as the superiority of rotating mechanism when compared to standard LiDAR setup. Moreover, comparative studies with one of the SOTA work, FAST-LIO2, have shown the competitive mapping results in multi-floor indoor environments.

Dynamic Coastal Mapping Using Sentinel-1 and Sentinel-2 Data Through Digital Earth Africa

Abstract. Coastal erosion poses a continuous threat to ecosystems, infrastructure, and property. To address these challenges and mitigate the effects of coastal changes, effective and current monitoring is essential. It is particularly important to monitor coastlines and coastal changes in Africa, where a significant portion of the population resides in coastal regions. While optical satellite imagery has been used for large-scale annual coastlines and change monitoring for Africa, its availability and quality are largely limited by the presence of cloud and cloud shadow. In comparison, using radar satellite observations such as Sentinel-1 data can provide consistent coastal mapping and change detection regardless of cloud presence. This paper outlines a fully automated supervised machine learning workflow using Sentinel-1 data and training samples extracted from Sentinel-2 data. It also explores the performance of the workflow for different coastal morphology types across the African coast. The workflow has proved to perform better and produced results that were visually more consistent with Sentinel-2 data compared to thresholding methods. While challenges exist to distinguish between land and water over smooth sandy beaches and rough near-shore water surfaces, our workflow provides an alternative method for coastal change mapping where optical satellites provide insufficient observations free from clouds. Python code of the proposed methodology has been made publicly available.

A LiDAR SLAM Algorithm Considering Dynamic Extraction of Feature Points in Underground Coal Mine

Abstract. To address the absence of GNSS signals in underground coal mines and the susceptibility of mainstream LiDAR SLAM to degradation due to insufficient feature constraints, this paper proposes a tightly-coupled SLAM algorithm incorporating LiDAR and IMU for use in such environments. The paper initially introduces a dynamic feature point extraction strategy, where the number of corner feature points can be dynamically adjusted based on the occurrence of degradation in the underground coal mine environment. This approach constructs a constraint matrix with rich and robust feature information, enhancing pose estimation accuracy in environments with inadequate feature constraints, mitigating degradation effects, and reducing global cumulative error. Subsequently, the consistent building of the underground coal mine environment is achieved through back-end factor graph optimization. Finally, to validate the effectiveness of the method, experimental analysis is conducted in an underground coal mine. The results demonstrate that LeGO_LOAM exhibits poor robustness in underground coal mines and fails to construct a globally consistent pose estimation and map. Conversely, the pose estimation error of the proposed method in this paper is 50.93% lower in the plane direction and 42.13% lower in the elevation direction compared to LIO_SAM. This underscores the method's significance as crucial technical support for the intelligent perception and positioning of robots in underground coal mines.

P09.02.A A WINDOW TO THE BRAIN: MONITORING THE HUMAN BRAIN USING FUNCTIONAL ULTRASOUND (FUS)-IMAGING IN FREELY MOVING SUBJECTS WITH AN ACOUSTICALLY FAVORABLE SKULL IMPLANT

Abstract BACKGROUND To this day, the mainstay of human neuroimaging remains (functional) Magnetic Resonance Imaging (fMRI), where subjects are asked to lie completely still and silent in a noisy scanner. Many key questions in clinical care and neuroscience remain poorly studied because of fMRI’s inability to facilitate natural behavior such as locomotion or even verbal speech. What is more, due to the cost and size of an (f)MRI-machine, monitoring of morphological (i.e. tumor regrowth) or functional changes in the human brain, pre- and post-tumor resection, is often spread out over intervals of many months, leaving patients in suspense about possible tumor regrowth.With the advent of ultrafast ultrasound imaging, a new high-resolution, depth-resolved and mobile neuro-imaging technique called functional Ultrasound (fUS) has emerged. In the last decade, fUS has been applied successfully in awake neurosurgical contexts to map out hemodynamics-based functional brain activity as well as tumor tissue at a mesoscopic scale. However, given the significant aberration of ultrasound signal through human skull bone, fUS is not available outside of OR as of yet. As a first step outside of the OR, we aimed to apply fUS in freely moving subjects with a skull bone defect covered by PEEK, an acoustically favorable material. MATERIAL AND METHODS Two subjects with PEEK-implants were repeatedly imaged over a period of >1 year. One of the subjects received a PEEK-implant after infection of the autologous bone, post-resection of an astrocytoma. Each subject received a custom-fitted 3D-printed helmet based on their MRI to fixate the probe and to attach the geometry needed for optical tracking. Using an optical tracking camera, the position of the Power Doppler Image (PDI) relative to CT/(f)MRI-volumes of the subject could be determined in real-time. Images were acquired using an experimental research system (Vantage-256, Verasonics) interfaced with a 9L-D linear array (GE, 5.3 MHz). All scans were acquired with a PRF of 800 Hz. The PDIs and angle compounded beamformed frames were stored for offline processing, consisting of 3D-vascular reconstructions and functional mapping. Functional tasks focused on the sensorimotor area of the lip. For measurements outside of the lab, the acquisition system was transformed into a mobile cart, which could be pushed by the subject itself. RESULTS Trans-PEEK imaging resulted in PDIs with highly detailed vasculature of healthy brain tissue and the resection cavity, up to several centimeters in-depth, reproduced reliably over a period of >12 months. In sedentary, standing and walking conditions, we were able to produce consistent functional maps of the human brain. CONCLUSION Our current work demonstrates how fUS has the potential to become a powerful window to the brain, both in terms of vascular monitoring of tumor regrowth, as well as functional brain mapping, outside of the operating room.

OR-LIM: Observability-aware robust LiDAR-inertial-mapping under high dynamic sensor motion

Light Detection And Ranging (LiDAR) technology has provided an impactful way to capture 3D data. However, consistent mapping in sensing-degenerated and perceptually-limited scenes (e.g. multi-story buildings) or under high dynamic sensor motion (e.g. rotating platform) remains a significant challenge. In this paper, we present OR-LIM, a novel observability-aware LiDAR-inertial-mapping system. Essentially, it combines a robust real-time LiDAR-inertial-odometry (LIO) module with an efficient surfel-map-smoothing (SMS) module that seamlessly optimizes the sensor poses and scene geometry at the same time. To improve robustness, the planar surfels are hierarchically generated and grown from point cloud maps to provide reliable correspondences for fixed-lag optimization. Moreover, the normals of surfels are analyzed for the observability evaluation of each frame. To maintain global consistency, a factor graph is utilized integrating the information from IMU propagation, LIO as well as the SMS. The system is extensively tested on the datasets collected by a low-cost multi-beam LiDAR (MBL) mounted on a rotating platform. The experiments with various settings of sensor motion, conducted on complex multi-story buildings and large-scale outdoor scenes, demonstrate the superior performance of our system over multiple state-of-the-art methods. The improvement of point accuracy reaches 3.39–13.6 % with an average 8.71 % outdoor and correspondingly 1.89–15.88 % with 9.09 % indoor, with reference to the collected Terrestrial Laser Scanning (TLS) map.

Mapping South Africa's canopy-forming kelp forests using low-cost, high-resolution Sentinel-2 imagery

Globally, kelp forests play a vital role in providing ecological, social and economic services. They contribute to the livelihoods and resilience of many coastal communities. To date, a gap exists in contemporary, comprehensive and consistent maps of South African kelp forests, since they were last mapped partially in 2006 by covering 900 km of South Africa's coastline. This study addresses this gap by producing a comprehensive remotely sensed high-resolution 2016 map of canopy forming coastal kelp forests. Using the 10m resolution Sentinel-2 A and B spectral bands 4 (RED) and 8 (NIR), an expert derived Normalized Difference Vegetation Index (NDVI) threshold was used to detect surface protruding kelp forests along 1300 km coastline. The results were validated with expert verification and using Google Earth, data from historic maps, remote sensing and biomass estimation of South African kelp resources. Our study demonstrates the aptitude of Sentinel-2 for detecting canopy-forming kelps at low cost along South Africa's west and south coast. We report the encountered drawbacks in mapping an expansive coastline with varied sea conditions including wave activity, tides and turbidity. Comprehensive ground validation, long-term monitoring sites and the use of time series datasets remain a gap towards improving the calibration of remote sensing methods and long-term monitoring of South African kelp forests. Our results provide a foundation for continuous monitoring of South African kelp forests as well as exploring other unmapped kelp/similar habitats to inform research and environmental management.

Hindcasting and updating Landsat-based forest structure mapping across years to support forest management and planning

Forest vegetation mapping that integrates forest inventory data with multispectral remote sensing data provides valuable geospatial products for public land management agencies, but resource managers may require rapid updating of maps as new imagery becomes available (updating) or retrospective mapping for times prior to forest inventory plot measurement (hindcasting). While forest attribute mapping using Landsat multispectral imagery is common, the accuracy of applying models outside of reference epoch to support long-term forest monitoring is not normally quantified. We examine whether a Landsat-based mapping approach can support robust, temporally consistent multivariate mapping of forest structure and composition data in support of forest management planning and landscape analysis. Specifically, we ask: how accurate forest attribute mapping was when hindcasting or updating outside of a period of time when forest inventory plot data were available (reference epoch)? In the western Cascade Mountains of Oregon and California, USA, we used the gradient nearest neighbor approach to annually impute USDA Forest Inventory and Analysis (FIA) plot data (2001–2016) to all 30-m forested pixels based on temporally smoothed Landsat multispectral imagery (1986–2021), including basal area, canopy cover, quadratic mean diameter of dominant trees, stand height, and the density of large diameter trees. We made extrapolations from models fit to a 10-year reference epoch to both earlier periods (2001–2006 hindcast) and to later period (2011–2016 update) and quantified prediction accuracies relative to models based on the full data (2001–2016). To evaluate the influence of spatial scale on hindcasting and updating, we compared full and extrapolation model predictions at pixel-level (0.09 ha) and hexagon-level (780 ha).At the plot-level, we found no strong differences between the full and extrapolation model predictions for R2 and mean error nor among predicted vs. observed regression coefficients. At the pixel-level, average differences due to hindcasting and updating were near zero, though differences varied up to 20 % across pixels. At the hexagon-level, the range in map differences was small (+/- 5 %), but hindcasting resulted in lesser forest attribute predictions. We observed greater variability in pixel-level and hexagon-level prediction differences when hindcasting or updating was temporally further away from the reference period. Using 2001 hindcast and 2016 updated maps as a case study, we found that with hindcasting and updating map differences were spatially aggregated across the study region. Our results support Landsat-based hindcasting and updating of forest attribute mapping beyond the time period covered by forest plot data. Our results suggest aggregating data to coarse spatial resolutions may minimize differences due to hindcasting and updating. Further research is needed to identify the key drivers for prediction differences to improve the accuracy of both hindcasting and updating as a basis for forest monitoring.

Mapping UK Pain Datasets to the OMOP Common Data Model: Lessons Learned

ObjectiveChronic Pain is defined as pain lasting for three months or longer and a report by the British Pain Society in 2016 found that 28 million adults live with this condition. Chronic pain research is problematic as it is poorly represented in health data and often in silos. Alleviate is the Advanced Pain Discovery Platform (APDP) Data Hub which is removing barriers to access to data. ApproachThe aim for Alleviate is to transform pain data sets within the UK to be Findable, Accessible, Interoperable and Reusable (FAIR) and be more discoverable for researchers via Health Data Research UK’s Cohort Discovery Tool. The tool enables identification of potential cohorts in real-time from datasets Alleviate provides. The onboarding requires datasets to be transformed to the open and widely used Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM). ResultsWe collaboratively developed CaRROT tools and used our mapping expertise to transform datasets to OMOP-CDM. The tools have improved efficient mapping of clinical and research data, covering ~10 million individuals’ records. Subsequently we learned valuable lessons on how to accurately and consistently map across data sets to support federated discovery. We also identified a lack of standard vocabulary representations for pain specific data. ConclusionsMapping data to OMOP is not straightforward. Attention to detail is required to ensure that consistent mapping decisions are made to maximise data reusability and interoperability. ImplicationsIn sharing the lessons learned of mapping datasets to OMOP we aim to help improve interoperability of disparate data sources.

Data Quality Implications in Research when Transitioning to a New Electronic Medical Record

The province of Alberta has been implementing a new hospital based electronic medical record (EMR) over the past several years based on the Epic system. With this implementation comes challenges related to consistent mapping of data to ensure consistency of data over time before and after implementation, even for standardized and routinely used elements. In a research study evaluating potential disparities in access to Emergency Department (ED) care, particularly in wait times and patient boarding (the time spent in the ED awaiting transfer to an inpatient bed), between mental health (MH) and non-MH patients during the COVID-19 pandemic. An interrupted time series analysis was used to evaluate the impact of the pandemic on boarding time. While there initially appeared to be a large decrease in boarding time near the start of the pandemic, it was later determined to be a data quality issue corresponding to the rollout of the new EMR at one hospital, despite the data coming from a secondary standardized dataset that is used for provincial and national reporting and expected to be reliable. This presentation will highlight how the issue was identified, what steps were taken to confirm the underlying cause, and how it was corrected to ensure accurate results in this research study.

(How) do universities listen? Evidence from institutional websites of the world's top universities

Introduction: this study investigates the implementation of organizational listening methods in Higher Education through a comparative analysis of institutional websites from top global universities and leading Lithuanian institutions. Emphasizing the role of these websites as pivotal digital platforms, the research aims to discern the extent to which universities facilitate two-way communication with stakeholders through digital platforms. Methodology: Websites were selected based on world university rankings, encompassing institutions from each continent. Websites were selected based on global university rankings, spanning institutions across continents. Qualitative content analysis employed predefined and emergent categories to evaluate interactivity and organizational listening features on these platforms. Findings reveal consistent stakeholder mapping but significant disparities in communication tools and channels, which impacts real-time, asynchronous, and symmetric engagement effectiveness. Communication structures range from integrated governance roles to fragmented responsibilities, influencing stakeholder accessibility and institutional transparency. Discussion: implications for organizational communication practices are discussed, highlighting strategies to enhance stakeholder engagement via institutional websites. The study underscores the pivotal role of communication management teams in fostering transparency and responsiveness. Conclusions: advocating for leveraging technological advancements, conclusions propose transforming websites into proactive platforms for organizational listening. Recommendations emphasize developing tailored communication strategies to optimize engagement and effectiveness in Higher Education contexts.

Tomographic projection optimization for volumetric additive manufacturing with general band constraint Lp-norm minimization

Tomographic volumetric additive manufacturing is a rapidly growing fabrication technology that enables rapid production of 3D objects through a single build step. In this process, the design of projections directly impacts geometric resolution, material properties, and manufacturing yield of the final printed part. Herein, we identify the hidden equivalent operations of three major existing projection optimization schemes and reformulate them into a general loss function where the optimization behavior can be systematically studied, and unique capabilities of the individual schemes can coalesce. The loss function formulation proposed in this study unified the optimization for binary and greyscale targets and generalized problem relaxation strategies with local tolerancing and weighting. Additionally, this formulation offers control on error sparsity and consistent dose response mapping throughout initialization, optimization, and evaluation. A parameter-sweep analysis in this study guides users in tuning optimization parameters for application-specific goals.

CRefNet: Learning Consistent Reflectance Estimation With a Decoder-Sharing Transformer.

We present CRefNet, a hybrid transformer-convolutional deep neural network for consistent reflectance estimation in intrinsic image decomposition. Estimating consistent reflectance is particularly challenging when the same material appears differently due to changes in illumination. Our method achieves enhanced global reflectance consistency via a novel transformer module that converts image features to reflectance features. At the same time, this module also exploits long-range data interactions. We introduce reflectance reconstruction as a novel auxiliary task that shares a common decoder with the reflectance estimation task, and which substantially improves the quality of reconstructed reflectance maps. Finally, we improve local reflectance consistency via a new rectified gradient filter that effectively suppresses small variations in predictions without any overhead at inference time. Our experiments show that our contributions enable CRefNet to predict highly consistent reflectance maps and to outperform the state of the art by 10% WHDR.

BA-CLM: A Globally Consistent 3D LiDAR Mapping Based on Bundle Adjustment Cost Factors.

Constructing a globally consistent high-precision map is essential for the application of mobile robots. Existing optimization-based mapping methods typically constrain robot states in pose space during the graph optimization process, without directly optimizing the structure of the scene, thereby causing the map to be inconsistent. To address the above issues, this paper presents a three-dimensional (3D) LiDAR mapping framework (i.e., BA-CLM) based on LiDAR bundle adjustment (LBA) cost factors. We propose a multivariate LBA cost factor, which is built from a multi-resolution voxel map, to uniformly constrain the robot poses within a submap. The framework proposed in this paper applies the LBA cost factors for both local and global map optimization. Experimental results on several public 3D LiDAR datasets and a self-collected 32-line LiDAR dataset demonstrate that the proposed method achieves accurate trajectory estimation and consistent mapping.

Addressing the challenges of loop detection in agricultural environments

AbstractWhile visual Simultaneous Localization and Mapping systems are well studied and achieve impressive results in indoor and urban settings, natural, outdoor, and open‐field environments are much less explored and still present relevant research challenges. Visual navigation and local mapping have shown a relatively good performance in open‐field environments. However, globally consistent mapping and long‐term localization still depend on the robustness of loop detection and closure, for which the literature is scarce. In this work, we propose a novel method to pave the way towards robust loop detection in open fields, particularly in agricultural settings, based on local feature search and stereo geometric refinement, with a final stage of relative pose estimation. Our method consistently achieves good loop detections, with a median error of 15 cm. We aim to characterize open fields as a novel environment for loop detection, understanding the limitations and problems that arise when dealing with them. Code is available at: https://github.com/CIFASIS/StereoLoopDetector

Like teacher, like pupil: Transferring backdoors via feature-based knowledge distillation

With the widespread adoption of edge computing, compressing deep neural networks (DNNs) via knowledge distillation (KD) has emerged as a popular technique for resource-limited scenarios. Among various KD methods, feature-based KD, which leverages the feature representations from intermediate layers of the teacher model to supervise the training of the student model, has shown superior performance and enjoyed wide application. However, users often overlook potential backdoor threats when using knowledge distillation (KD) to extract knowledge. To address the issue, this paper mainly contributes to three aspects: (1) we try the first step of exploring the security risks in feature-based KD, where implanted backdoors in teacher models can survive and transfer to student models. (2) We propose a backdoor attack method targeting feature distillation, achieved by encoding backdoor knowledge into specific neuron activation layers. Specifically, we optimize triggers to induce consistent feature map values in the teacher model and transfer the backdoor knowledge to student models through these features. We also design an adaptive defense method against this attack. (3) Extensive experiments on four common datasets and six sets of different teacher and student models validate that our attack outperforms the state-of-the-art (SOTA) baselines, with an average attack success rate of (∼×1.5). Additionally, we discuss effective defense methods against such backdoor attacks.

Dynamic selection for reconstructing instance-dependent noisy labels

As an inevitable issue in annotating large-scale datasets, instance-dependent label noise (IDN) can cause serious overfitting in neural networks. To combat IDN, label reconstruction methods have been developed with noise transition matrices or DNNs to simulate the transition from clean labels to noisy labels. Nevertheless, the absence of correct supervisions will lead to learning wrong noise transitions. This motivates us to select samples with clean labels to fetch the correct supervisions. However, the difficulty in obtaining prior knowledge of the noise rate prohibits the use of existing sample selection methods. To this end, we propose a dynamic sample selection method, namely Identity Mapping (IdMap), to overcome this limitation. Inspired by the feature-dependent characteristic of IDN, we first introduce the extracted instance features and pseudo-ground-truth labels to reconstruct noisy labels. A partial identity mapping between two labels is then established and samples with consistent identity mapping output are selected as clean data to update the classifier. Extensive experiments on both artificial and real-world noisy datasets demonstrate the superiority of IdMap compared with other state-of-the-art methods.