Poor Precision Research Articles

English Grammar Auto-Correction Robot based on Grammatical Error Generation Model

The traditional English grammar error correction system has problems such as poor error recognition precision and error correction success rate to be improved. Therefore, in this study, data augmentation techniques were used to transform and process error correcting English texts, and a rule-based and shallow neural network-based English text grammar correction model was constructed. The experimental results showed that the grammar error generation (GEG), rule-based (RB), classification-based (CB), and recurrent neural network (RNN) models achieved accuracy rates of 93.92%, 82.17%, 79.41%, and 88.09% in correcting grammar errors on 1702641 test sentences in the One Billion word corpus, respectively. The experimental results showed that the English grammar error correction model designed in this study had a strong error correction ability, but the computational efficiency was low. The research results significantly improved the accuracy and generalization ability of English grammar correction, optimized learning costs, and brought positive impacts to educational applications, providing strong support for the development of intelligent English grammar correction.

Stereolithography-assisted sodium alginate-collagen hydrogel scaffold with molded internal channels

Fabricating internal vascular networks within a hydrogel scaffold is essential for facilitating the supply of nutrients, oxygen, and metabolism exchange required by the encapsulated cells. The challenges in current hydrogel scaffold fabrication involve the difficulty of building adequate internal channels, poor scaffold geometry precision, and low cell viability caused by the fabrication process and polymer material properties. Stereolithography (SLA) stands out as a 3D printing technique distinguished by its superior production efficiency, advanced precision, and remarkable resolution in crafting intricate custom geometries. These attributes establish it as an innovative approach for templates in scaffold fabrication, potentially surpassing the fused deposition modeling (FDM)-based template strategy. Meanwhile, it exerts less shear stress on the cells compared to the direct bioprinting process. This novel strategy enables the fabrication of hydrogel vascular structure within the precision of 500 µm in both channel diameter and wall thickness. In this paper, various sodium alginate and collagen (SA-Col) composite hydrogels with varying collagen concentrations have been investigated to identify the optimal ratio for fabricating hydrogel scaffolds with channels.

Background adaptive PosMarker based on online generation and detection for locating watermarked regions in photographs

Robust watermarking technology can embed invisible messages in screens to trace the source of unauthorized screen photographs. Locating the four vertices of the embedded region in the photograph is necessary, as existing watermarking methods require geometric correction of the embedded region before revealing the message. Existing localization methods suffer from a performance trade-off: either causing unaesthetic visual quality by embedding visible markers or achieving poor localization precision, leading to message extraction failure. To address this issue, we propose a background adaptive position marker, PosMarker, based on the gray level co-occurrence matrix and the noise visibility function. Besides, we propose an online generation scheme that employs a learnable generator to cooperate with the detector, allowing joint optimization between the two. This simultaneously improves both visual quality and detection precision. Extensive experiments demonstrate the superior localization precision of our PosMarker-based method compared to others.

Dissociable effects of mild COVID-19 on short- and long-term memories.

Recent studies have highlighted the presence of cognitive deficits following COVID-19 that persist beyond acute infection, regardless of the initial disease severity. Impairments in short- and long-term memory are among the core deficits reported by patients and observed in objective tests of memory performance. We aimed to extend previous studies by examining performance in a task that allows us to directly compare and contrast memories at different timescales. More specifically, we assessed both short- and long-term memories for contextual-spatial associations encoded during a common session and probed at different durations using an equivalent task in non-hospitalized individuals recovering from mild COVID-19 compared to healthy controls. The approach equated all aspects of memory materials and response demands, isolating performance changes resulting only from memory timescales and thus allowing us to quantify the impact of COVID-19 on cognition. In addition to providing measures of accuracy and response times, the task also provided a sensitive continuous readout of the precision of memory representations, specifically by examining the resolution with which spatial locations were retained in memory. The results demonstrated selective impairment of long-term memory performance in individuals recovering from mild COVID-19 infection. Short-term memory performance remained comparable to healthy controls. Specifically, poor precision of long-term memory representations was demonstrated, which improved with days since diagnosis. No such relationship was observed for short-term memory performance. Our findings reveal a specific impairment to the precision of spatial-contextual long-term memory representations in individuals recovering from mild COVID-19 and demonstrate evidence of recovery in long-term memory over time. Further, the experimental design provides a carefully controlled and sensitive framework to assess memory across different durations with the potential to provide more detailed phenotyping of memory deficits associated with COVID-19 in general.

Stacking Approach for Classification of Multiclass Dry Beans

Background: The most frequently cultivated edible legume crop in the world is Dry Beans (DB), which exhibit high genetic diversity. The impact of seed quality on crop production is undeniable. Seed classification is essential for production and marketing to provide building blocks for sustainable agricultural systems. With regard to big production numbers, traditional methods for classifying seed quality have flaws, including complicated processes, poor precision and sluggish inspection. Rapid and high throughput solutions are provided by automatic categorization algorithms based on machine learning and computer vision. Modern automatic classification models have made significant strides, yet there is still room for improvement by adding new methods. Since crop production is in the form of population rather than a single variation, the main goal of this study is to offer a technique for getting homogeneous dry bean variants. Although numerous intelligent models have been presented, most rely on a single classifier, which makes them unable to handle noisy and unbalanced data and can cause overfitting. Methods: To reduce bias and variance and avoid overfitting a single classifier-based model, this study provides an ensemble-based prediction model combining pertinent attributes and a simple stacking ensemble technique, Xtreme Stacking Prediction of Dry Beams (X-SPDB). The forecast is made using the proposed X-SPDB, which incorporates several assumptions. Result: Comparisons are made between the proposed X-SPDB’s performance and simple Decision Tree, SVM, Random Forest, Naive Bayes, SVM, Logistic Regression and SVM with XGBoost.

Model-based hybrid control of combined active–passive vertical zero-gravity system

Ground-based full-scale physical experiments on spacecraft have become an essential part of ground simulations. However, achieving a vertical translational degree of freedom for zero-gravity simulation has become a key technological bottleneck for six-degree-of-freedom ground simulations. The combined active–passive vertical zero-gravity system discussed in this paper can achieve zero-gravity simulation in the vertical direction; however, it exhibits deficiencies such as inaccurate modeling, low control precision, and significant control delays. Addressing the issue of inaccurate modeling, this study uses interpretable ODE neural networks to identify the dynamic model of the system. For resolving the problems of poor control precision and significant delays, this study proposes a hybrid control algorithm based on the ODE dynamic model. The controller includes nominal model feedforward, PID feedback, and learning-based feedforward components. Finally, comparative experiments are designed to validate the proposed model-based hybrid control algorithm. Compared to PID-Feedforward controller, the hybrid controller exhibits a higher response speed, while compared to model-based feedforward controller, it exhibits a better control accuracy. Therefore, the proposed hybrid control algorithm based on the ODE dynamic model effectively enhances the control response speed and improves the control precision of the system.

High-accuracy measurement of 36SF5 + signal using an ultrahigh-resolution isotope ratio mass spectrometer.

The Δ36S standard deviation measured in a conventional isotope ratio mass spectrometer such as MAT 253 is ca 0.1‰ to 0.3‰. At this precision, it is difficult to resolve the origin of non-mass-dependent sulfur isotope fractionation in tropospheric sulfate aerosol and in Martian meteorites or small deviations from the canonical mass-dependent fractionation laws. Interfering ions with m/z at 131 of 36SF5 + are suggested by the community as the cause of the poor precision, but the exact ion species has not been identified or confirmed. Here we examined the potential interfering ions by using a Thermo Scientific ultrahigh-resolution isotope ratio mass spectrometer to measure SF6 working gas and SF6 gases converted from IAEA-S1/2/3 Ag2S reference materials. We found that there are two resolvable peaks to the right of the 36SF5 + peak when a new filament was installed, which are 186WF4 2+ followed by 12C3F5 +. However, only the 12C3F5 + interference peak was observed after more than three days of filament use. 12C3F5 + is generated inside the instrument during the ionization process. Avoiding the interfering signals, we were able to achieve a Δ36S standard deviation of 0.046‰ (n = 8) for SF6 zero-enrichment and 0.069‰ (n = 8) for overall measurement start from silver sulfide IAEA-S1. Aging the filament with SF6 gas can avoid the interference of 186WF4 2+. Minimizing the presence of carbon-bearing compounds and avoiding the interfering signals of 12C3F5 + from 36SF5 +, we can improve Δ36S measurement accuracy and precision, which helps to open new territories for research using quadruple sulfur isotope composition.

External Validation and Further Exploration of Fall Prediction Models Based on Questionnaires and Daily-Life Trunk Accelerometry

Ambulatory measurements of trunk accelerations can provide valuable insight into the amount and quality of daily life activities. Such information has been used to create models to identify individuals at high risk of falls. However, external validation of such prediction models is lacking, yet crucial for clinical implementation. We externally validated 3 previously described fall prediction models. Complete questionnaires and 1-week trunk acceleration data were obtained from 263 community-dwelling people (mean age 71.8 years, 68.1% female). To validate models, we first used the coefficients and optimal cutoffs from the original cohort, then recalibrated the original models, as well as optimized parameters based on our new cohort. Among all participants, 39.9% experienced falls during a 6-month follow-up. All models showed poor precision (0.20-0.49), poor sensitivity (0.32-0.58), and good specificity (0.45-0.89). Calibration of the original models had limited effect on model performance. Using coefficients and cutoffs optimized on the external cohort also had limited benefits. Lastly, the odds ratios in our cohort were different from those in the original cohort, which indicated that gait characteristics, except for the index of harmonicity ML (medial-lateral direction), were not statistically associated with falls. Fall risk prediction in our cohort was not as effective as in the original cohort. Recalibration as well as optimized model parameters resulted in a limited increase in accuracy. Fall prediction models are highly specific to the cohort studied. This highlights the need for large representative cohorts, preferably with an external validation cohort.

Optimizing the Electrical Discharge Machining Process Parameters of the Nimonic C263 Superalloy: A Sustainable Approach

Engineers continue to be concerned about electrical discharge-machined components’ high energy consumption, machining debris, and poor dimensional precision. The aim of this research is to propose a hybrid neuro-genetic approach to improve the machinability of the electrical discharge machining (EDM) of the Nimonic C263 superalloy. This approach focuses on reducing the energy consumption and negative environmental impacts. The material removal rate (MRR), electrode wear ratio (EWR), specific energy consumption (SEC), surface roughness (Ra), machining debris (db), and circularity (C) are examined as a function of machining parameters such as the voltage (V), pulse on time (Ton), current (I), duty factor (τ), and electrode type. By employing the VIKOR method, all the responses are transformed into a distinctive VIKOR index (VI). Neuro-genetic methods (a hybrid VIKOR-based ANN-GA) can further enhance the best possible result from the VIKOR index. During this step, the hybrid technique (VIKOR-based ANN-GA) is used to estimate an overall improvement of 9.87% in the response, and an experiment is conducted to confirm this condition of optimal machining. This work is competent enough to provide aeroengineers with an energy-efficient, satisfying workplace by lowering the machining costs and increasing productivity.

A planar compound eye based microsystem for high precision 3D perception

Three-dimensional (3D) panoramic vision system plays a fundamental role in the biological perception of external information, and naturally becomes a key system for embodied intelligence to interact with the outside world. A binocular vision system with rotating eyeball has long baseline, large volume and weak sensitivity to motion. A compound eye system has small volume, high sensitivity to motion but poor precision. Here, a planar compound eye microsystem for high precision 3D perception is proposed by combining semiconductor manufacturing process and biological compound eye structure. Using a semiconductor planar image sensor as the sensing unit, a space-coded planar sub-eye array is designed and its sub field of view (FOV) is dynamically mapped to the image sensor. It solves the problem that a traditional vision system cannot simultaneously accommodate wide FOV with long focal length and high sensitivity to motion with high resolution. The parallax among different sub-eyes enables the system to accurately perceive and dynamically track the 3D position of the target in the range of 10 m and within the FOV of 120 ° in a single compound eye. This system is of great significance in the fields of intelligent robot and intelligent perception.

TWOFOLD SUBAREA MODEL FOR ESTIMATING COMMUTER PROPORTION IN 10 METROPOLITAN AREAS

The metropolitan area is a major contributor to national GDP. The metropolitan area is a center of attraction for many people who come to earn income as commuters. Commuters are people who carry out work activities in the center of the metropolitan area, which are carried out by residents who live in suburban areas around the center of the metropolitan area and commute regularly every day. The availability of commuter statistics from surveys for presentation level down to the smallest administrative level, such as regencies/municipalities, is unreliable. This happens because this level of presentation has poor precision due to insufficient samples due to the Statistics Indonesia survey design for making estimates at the national and provincial levels. It can be done using small area estimation (SAE) to meet increasing data needs, but existing SAE models can often estimate only at one level. To meet data requests more effectively, a model is needed that can estimate several small areas simultaneously. In SAE, one of the SAE models that can do this is the twofold subarea model. The twofold subarea model produces estimates of the proportion of commuters with good precision at the subarea level (regencies/municipalities) and area level (metropolitan area), with the RRMSE percentage value of the estimated proportion of commuters being below 25% for all regions. The results of this research can be used to present commuter data at the regencies/municipalities level and metropolitan area level where there is a lack of samples and become a new opportunity for Statistics Indonesia to increase statistical production in small areas, which is more effective compared to other SAE methods which have so far been used only to estimate one area level.

#2812 Continuous blood pressure monitoring with a finger cuff is unreliable in chronic hemodialysis patients

Abstract Background and Aims Intradialytic hypotension (IDH) is the most common complication of hemodialysis (HD), which is often related to ultrafiltration. Due to the high prevalence of cardiovascular disease (CVD), compensatory mechanisms often fail. In clinical practice, BP is measured 2x/hour, while additional readings are only performed when symptoms arise. However, since IDH defined by a nadir systolic BP (SBP)<90 mmHg or <100 mmHg (depending on pre-dialysis SBP), regardless of symptoms, is most strongly associated with mortality, current BP measurement frequency may severely underestimate the true incidence of IDH. Continuous BP-monitoring may offer a solution to this issue. The Clearsight device (CSD) is a non-invasive continuous BP-monitoring system that measures arterial pressure and waveforms through a finger cuff by photoplethysmography. Yet, a compromised peripheral circulation may hamper its reliability in HD patients. This study aims to assess the agreement between intradialytic CSD-monitoring and standard oscillometric brachial arterial pressure (oBAP) assessment. Methods 40 Chronic HD patients were cross-over randomized to 4 modalities, each for 2 weeks: [standard HD (S-HD), cool HD (C-HD), low volume hemodiafiltration (LV-HDF) and high-volume HDF (HV-HDF)]. During one session of each second week, BP and heart rate (HR) were continuously recorded with the CSD (BMEYE/Edwards Lifesciences) and every 15 minutes with a brachial arterial cuff (contralateral of the fistula or graft, if applicable). Simultaneous measurements of SBP, diastolic BP (DBP) and HR were compared with descriptive statistics (mean difference [bias, BAP minus CSD], standard deviation (SD) and range) and Bland-Altman analysis. Bias was defined as the mean difference between oBAP and CSD, and 95% limits of agreement (LA) as the mean difference ±1.96 SD. Post-hoc, patients were stratified into tertiles of unexplained missing CSD-readings, to explore whether specific patient characteristics explain failed readings. Results 75% of patients was male, with a mean age 69.7 (±13.5 years) and median dialysis vintage 3.0 years (IQR 1.0-5.8). Diabetes mellitus and CVD were present in 48% and 73%, resp. Of the 2720 paired BP-measurements, 42% CSD and 6% of oBAP-readings were missing, leaving 36% of absent CSD-readings unexplained. Bland-Altman analysis, comparing oBAP and CSD-measurements, revealed a mean bias ± LA of 13.3 ± 44.5 mmHg for SBP, a mean bias ± LA of 13.3 ± 25.5 for DBP, and a mean bias ± LA of −0.03 ± 22.9 for HR. Surprisingly, the group with a high (>38%) proportion of unexplained missing CSD-values was significantly younger compared to the low (<21%) and intermediate group (21-38%) (p = 0.015), whereas no significant differences were observed with respect to dialysis vintage, diabetes mellitus or CVD. Conclusions 1) Compared to the oBAP method, finger cuff-monitoring shows poor precision and accuracy in BP-measurements, but good accuracy and poor precision in HR-readings; 2) 1/3 of readings failed for unknown reasons. Thus, CSD-monitoring is unsuitable for intradialytic hemodynamic monitoring in chronic HD patients.

Extraction of laser stripe centerlines from translucent optical components using a multi-scale attention deep neural network

The precise extraction of laser stripe centerlines is critical for line-laser 3D scanning systems. However, conventional methods relying on threshold segmentation and morphological operations face significant challenges when confronted with pervasive optical phenomena, including specular reflection, scattering, and bleeding, which are commonly observed in translucent optical components. These methods typically require complex preprocessing procedures and often yield poor precision in centerline extraction. In this paper, we introduce a novel learning-based approach, complemented by a meticulously curated dataset, explicitly designed to address these challenges. Our proposed method leverages a multi-scale attention U-Net-like architecture, initially tasked with the segmentation of laser stripes from the complex background environment. Subsequently, it employs the Steger algorithm for the precise extraction of laser stripe centerlines. The experimental results, obtained by comprehensively evaluating real-world captured images, clearly demonstrate the effectiveness of our deep neural network combined with the Steger algorithm. This combined approach exhibits exceptional accuracy even when challenged by the interferences from specular reflection, scattering, and bleeding artifacts. Specifically, our method achieves a mean intersection over union (mIoU) of 84.71% for the laser stripe detection task, accompanied by a mean square error (MSE) of 10.371 pixels. Also, the average execution time for the centerline extraction task is notably efficient at 0.125 s.

Assessment of anxiety in children with neurodevelopment disorders: Rasch analysis of the Spence Children's Anxiety Scale.

Anxiety is common in neurodevelopmental disorders (NDD). The parent version of the Spence Children's Anxiety Scale (SCAS-P) is a widely used measure to assess anxiety across a broad range of childhood populations. However, assessment of the measurement properties of the SCAS-P in NDDs have been limited. The present study aimed to assess the psychometric properties of the SCAS-P in children with attention-deficit/hyperactivity disorder (ADHD) and autism spectrum disorder (ASD) using Rasch Measurement Theory. Data from the Province of Ontario Neurodevelopmental Disorders Network Registry were used in the analysis. Children (ages 6-13 years old) with a primary diagnosis of ADHD (n=146) or ASD (n=104) were administered the SCAS-P. Rasch Measurement Theory was used to assess measurement properties of the SCAS-P, including unidimensionality and item-level fit, category ordering, item targeting, person separation index and reliability and differential item functioning. The SCAS-P fit well to the Rasch model in both ADHD and ASD, including unidimensionality, satisfactory category ordering and goodness-of-fit. However, item-person measures showed poor precision at lower levels of anxiety. Some items showed differential item functioning, including items within the obsessive-compulsive, panic/agoraphobia and physical injury fears domains, suggesting that the presentation of anxiety may differ between ADHD and ASD. Overall, the results generally support the use of the SCAS-P to screen and monitor anxiety symptoms in children with ADHD and ASD. Future studies would benefit from examination of more severely anxious NDD cohort, including those with clinically diagnosed anxiety.

Precision and accuracy of transcutaneous CO2 monitoring in infants born at 32-36 weeks of pregnancy on respiratory support

Background Respiratory disorders in premature neonates often require respiratory support. Continous transcutaneous monitoring is an available non-invasive option to monitor CO2 pressure, substituting the need for blood gas analysis as the gold standard evaluation in practice. Most studies have been conducted on very and extremely preterm neonates, but rarely in late and moderately preterm neonates. Objective To determine the precision and accuracy of transcutaneous CO2 pressure measuring devices compared to arterial blood gas analysis in neonates of 32-36 weeks gestational age who received respiratory support. Methods This diagnostic, cross-sectional study was conducted on 35 late and moderately preterm neonates of 32–36 weeks gestation who received cardiopulmonary resuscitation (CPR) in the Neonatology Unit at Rumah Sakit Cipto Mangunkusumo, Jakarta. Subjects were monitored with a transcutaneous CO2 monitor and blood gas analysis (BGA). CO2 pressure measurements were made three times from the two devices. Data were analyzed using Spearman’s correlation and Bland-Altman tests to determine the precision and accuracy of transcutaneous monitoring by comparing its mean difference (MD) to BGA as the gold standard measurement. Results Spearman’s analysis revealed a significant positive correlation between BGA and transcutaneous CO2 monitoring (P<0.001). However, the Bland - Altman test revealed a level of agreement between measuring devices was -14.46 to 6.9, with mean difference of -3.78; indicating poor precision of the transcutaneous evaluation regardless its high accuracy compared to its gold standard. Conclusion The transcutaneous CO2 monitoring device has low precision, but a strong positive correlation to BGA; underlining its high accuracy in practice. Transcutaneous CO2 monitoring cannot replace BGA, the gold standard examination.

A Soft Collaborative Robot for Contact-based Intuitive Human Drag Teaching.

Soft material-based robots, known for their safety and compliance, are expected to play an irreplaceable role in human-robot collaboration. However, this expectation is far from real industrial applications due to their complex programmability and poor motion precision, brought by the super elasticity and large hysteresis of soft materials. Here, a soft collaborative robot (Soft Co-bot) with intuitive and easy programming by contact-based drag teaching, and also with exceptional motion repeatability (< 0.30% of body length) and ultra-low hysteresis (< 2.0%) is reported. Such an unprecedented capability is achieved by a biomimetic antagonistic design within a pneumatic soft robot, in which cables are threaded to servo motors through tension sensors to form a self-sensing system, thus providing both precise actuation and dragging-aware collaboration. Hence, the Soft Co-bots can be first taught by human drag and then precisely repeat various tasks on their own, such as electronics assembling, machine tool installation, etc. The proposed Soft Co-bots exhibit a high potential for safe and intuitive human-robot collaboration in unstructured environments, promoting the immediate practical application of soft robots.

Predictive Model Control and Intelligent Agent are Used To Optimize The Precision Control of an Electro-Pneumatic Actuator-Based Robotic Bottle Capper

The inferior corking performance observed in bottled drinks from breweries is attributed to poor precision. To address this issue, an enhanced precision control system for an electro-pneumatic actuator-based bottle capper is introduced, utilizing model predictive control and intelligent agent technologies. This involves characterizing the electro-pneumatic actuator system, developing a conventional model for the robotic bottle capper, and designing a rule base to improve precision in the capping mechanism, thereby boosting production capacity per unit time. To achieve this, AI is trained to design a rule-based model ensuring optimal efficacy of the capping mechanism, thus enhancing the brewery industry's production capacity and revenue generation. A SIMULINK model is developed to demonstrate the improved precision control of the electro-pneumatic actuator robotic bottle capper using model predictive control and intelligent agent. Results show a significant increase in corking precision from 94.4% with conventional methods to 99.9% when intelligent agents are incorporated into the system. This translates to a 5.5% improvement in corking precision, with production capacity increasing from 27,000 crates using conventional methods to 28,570 with model predictive control. Moreover, with the integration of intelligence, the production capacity further rises to 342,500 bottles.

A photographic method to identify reservoir geohazards induced by rock mass deterioration of hydro-fluctuation belt

Potential geohazards triggered by the rock mass deterioration (RMD) of the hydro-fluctuation belt (HFB) in the Three Gorges Reservoir area (TGRA) severely threaten the lives and property of people, it is necessary and urgent to timely identify this kind of potential geohazard. Aiming at this issue, several typical evolution modes of potential RMD-induced geohazards were generalized, and an unmanned aerial vehicle (UAV) photographic method integrating the orthophotography, three-dimensional multi-angle oblique photography and high-precision digital elevation model (DEM) was proposed to early recognize the newly formed potential RMD-induced geohazards. By resolving the problem of poor precision in the orthophotographic images of high steep bank slopes, the proposed method improved the recognition reliability of potential reservoir geohazards in the HFBs. Taking the bank slopes from Zigui County to Badong County as the study area, 116 new potential RMD-induced geohazard sites were identified by the presented approach, including 64 and 52 sites on the left and right bank, respectively. Through field verifications, 17 and 19 sites on the left and right bank were found to be consistent with existing hazards, indicating the effectiveness of the proposed approach in avoiding the false and missing recognitions of potential RMD-induced reservoir geohazards at current levels of knowledge and technology. The study can provide a valuable reference for the early identification of RMD-induced geohazards in reservoir areas.

Task-Driven Causal Feature Distillation: Towards Trustworthy Risk Prediction

Since artificial intelligence has seen tremendous recent successes in many areas, it has sparked great interest in its potential for trustworthy and interpretable risk prediction. However, most models lack causal reasoning and struggle with class imbalance, leading to poor precision and recall. To address this, we propose a Task-Driven Causal Feature Distillation model (TDCFD) to transform original feature values into causal feature attributions for the specific risk prediction task. The causal feature attribution helps describe how much contribution the value of this feature can make to the risk prediction result. After the causal feature distillation, a deep neural network is applied to produce trustworthy prediction results with causal interpretability and high precision/recall. We evaluate the performance of our TDCFD method on several synthetic and real datasets, and the results demonstrate its superiority over the state-of-the-art methods regarding precision, recall, interpretability, and causality.

Not all points are balanced: Class balanced single-stage outdoor multi-class 3D object detector from point clouds

Outdoor 3D object detection is a hot topic in autonomous driving. The mainstream pure point cloud method is down-sampling through different task-oriented strategies to retain representative foreground points. Although such strategies are conducive to finding instances, these methods still suffer from two issues: class points imbalance during down-sampling stages, and foreground/background points imbalance in the final retained point clouds. The former imbalance results in poor precision for small objects; and the latter ignores background points, leading to a false positive phenomenon. To tackle the unbalanced phenomenon, we propose a simple yet effective balanced 3D detector, termed CB-SSD, including two balanced strategies: class balance strategy (CBS) and foreground/background balance strategy (FBBS). It is important to note that we do not alter the distribution of point clouds. Instead, we guide the model’s attention towards different classes equally. CB-SSD shows better precision on small objects, reducing false positives where foreground points and background points are similar. Considering both speed and accuracy, CB-SSD achieves state-of-the-art based on pure point clouds (single-stage) on KITTI and ONCE datasets. On KITTI, CB-SSD attains a multi-class accuracy of 72.92 mAP with 81 FPS.