One-stage Model Research Articles

Two-stage hybrid model for hydrological series prediction based on a new method of partitioning datasets

• A new method of dividing datasets based on PSO is proposed. • BNN incorporates Langevin dynamics in the parameter update step (LDBNN). • Integration of LDBNN, CEEMDAN, VMD and the proposed method yield the hybrid model. • The new data dividing method performs better than the traditional dividing method. • The hybrid model yields excellent point and interval predictions. Obtaining accurate point estimates and reliable interval prediction results for rainfall and runoff series is important to aid in water resource decision-making and planning management in a changing environment. In this paper, we propose a two-stage hybrid model (P-CVEE-LDBNN). The model uses the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and variational mode decomposition (VMD) methods for data preprocessing and a Bayesian neural network (LDBNN) based on Langevin dynamics as a prediction model. Then, a new dataset partitioning method based on a particle swarm optimization algorithm, which is different from the traditional dataset partitioning method, is applied. A one-stage hybrid model (T-CEE-LDBNN) that integrates the CEEMDAN method and the traditional method of partitioning datasets, a one-stage hybrid model (P-CEE-LDBNN) that integrates the CEEMDAN method and the new method of partitioning datasets, and the two-stage hybrid P-CVEE-LDBNN model were compared. These models were applied to monthly runoff and monthly precipitation series from seven hydrological and meteorological stations in the Yellow River basin. The mean absolute error (MAE), absolute root mean squared error (RMSE), Nash–Sutcliffe efficiency (NSE), and correlation coefficient (R) were used to evaluate the predictive ability of the models. The containment ratio (CR), the average bandwidth (IW), and the average asymmetry degree (S) were used to evaluate the interval prediction results of the models. The results show that (1) the subset obtained with the new method of data partitioning considering statistical properties is more favorable for model prediction. (2) The second decomposition approach based on VMD is a reliable method that can significantly improve the prediction accuracy of the final model. (3) Compared with traditional neural networks that can only obtain deterministic point prediction results, Bayesian methods can provide intervals with prediction results, making the results of the two-stage hybrid P-CVEE-LDBNN model highly reliable.

A novel detection method for hot spots of photovoltaic (PV) panels using improved anchors and prediction heads of YOLOv5 network

Accurate classification and detection of hot spots of photovoltaic (PV) panels can help guide operation and maintenance decisions, improve the power generation efficiency of the PV system, and ensure power stations’ safe and stable operation. Considering that, in this paper, the hot spots of PV panels collected on site are taken as the research object, and their formation mechanism is studied. Based on this, the morphological characteristics possessed by the hot spots of PV panels are classified into circular, linear, and array ones. A novel method for detecting hot spots of PV panels based on improved anchors and prediction heads of the YOLOv5 (AP-YOLOv5) network is proposed. Besides, to improve the detection precision of the YOLOv5 network at different scales in hot spots of PV panels, the K-means clustering algorithm is employed to cluster the length–width ratio of the data annotation frame, and a group of the anchors with smaller values is added so as to realize the detection of small targets by optimizing the cluster number. Apart from that, the corresponding prediction heads are constructed for the new anchor parameters to improve the detection precision concerning hot spots of PV panels. Furthermore, the model is verified by training experimental data and comparing test set results. The results showed that compared with other one-stage object detection models, the mean average precision (mAP) of the proposed network can achieve 87.8%, while the average recall rate is 89.0%, and the F1 score reaches 88.9%. In addition, the precision of this model is better than that of other models while maintaining a high frame rate; the frames per second (FPS) is as high as 98.6, thus laying a foundation for developing rapid detection tool of hot spots of PV panels, improving the safety of power station operation, and providing a method for intelligent operation and maintenance of PV power stations.

Higher Arm Versus Lower Arm Systolic Blood Pressure and Cardiovascular Outcomes: a Meta-Analysis of Individual Participant Data From the INTERPRESS-IPD Collaboration.

BACKGROUND:Guidelines recommend measuring blood pressure (BP) in both arms, adopting the higher arm readings for diagnosis and management. Data to support this recommendation are lacking. We evaluated associations of higher and lower arm systolic BPs with diagnostic and treatment thresholds, and prognosis in hypertension, using data from the Inter-arm Blood Pressure Difference—Individual Participant Data Collaboration.METHODS:One-stage multivariable Cox regression models, stratified by study, were used to examine associations of higher or lower reading arm BPs with cardiovascular mortality, all-cause mortality, and cardiovascular events, in individual participant data meta-analyses pooled from 23 cohorts. Cardiovascular events were modelled for Framingham and atherosclerotic cardiovascular disease risk scores. Model fit was compared throughout using Akaike information criteria. Proportions reclassified across guideline recommended intervention thresholds were also compared.RESULTS:We analyzed 53 172 participants: mean age 60 years; 48% female. Higher arm BP, compared with lower arm, reclassified 12% of participants at either 130 or 140 mm Hg systolic BP thresholds (both P<0.001). Higher arm BP models fitted better for all-cause mortality, cardiovascular mortality, and cardiovascular events (all P<0.001). Higher arm BP models better predicted cardiovascular events with Framingham and atherosclerotic cardiovascular disease risk scores (both P<0.001) and reclassified 4.6% and 3.5% of participants respectively to higher risk categories compared with lower arm BPs).CONCLUSIONS:Using BP from higher instead of lower reading arms reclassified 12% of people over thresholds used to diagnose hypertension. All prediction models performed better when using the higher arm BP. Both arms should be measured for accurate diagnosis and management of hypertension.REGISTRATION:URL: https://www.clinicaltrials.gov; Unique identifier: CRD42015031227.

X-Ray Small Target Security Inspection Based on TB-YOLOv5

Security inspection is extremely important for the safety of public places. In this research, we are trying to propose a novel algorithm and investigated theoretically in the X-ray dataset, which can optimize the relative low detection accuracy and the latent omission detection of smaller objects when using You Only Look Once version 5 (YOLOv5). For one side, the transform detection network is selected to be added at the bottom layer of backbone structure to avoid the loss of useful information during sequential calculation. On another side, we attempt to adjust the existing PANet structural elements of the model, including their connections and other related parameters to improve the detection performance. We integrate an efficient BiFPN with the CA mechanism, which can enhance feature extraction, and named it attention-BiFPN. Experimental consequences demonstrate that the detection accuracy of the proposed model, which we name “TB-YOLOv5,” has obvious advantages in check performance compared with the mainstream one-stage object detection models. Meanwhile, compared with YOLOv5, the data results display an improvement of up to 14.9%, and the average precision at 0.5 IOU even reached 23.4% higher in the region of small object detection. Our purpose was to explore the potential of changing a popular detection algorithm such as YOLO to address specific tasks and provide insights on how specialized adjustments can influence the detection of small objects. Our work can supply an effective method of enhancing the performance of X-ray security inspection and show promising potential for deep learning in related fields.

O duration da vantagem competitiva gera retornos de longo prazo no mercado de ações?

ABSTRACT The purpose of this article was to develop a new indicator to estimate the aggregate long-term expected return on stocks. There is not a widely used method to model directly the aggregated expected return of the stock market. Most current methods use indirect approaches. We developed a new indicator that does not need an econometric model to generate expected returns and provides an estimate of the long-term expected returns. The proposed methodology can be used to develop an indicator of future returns of the stock market similar to the yield-to-maturity used for bonds. We used a restricted one-stage constant-growth model - a variant of the residual income model (RIM) - whose main input is the duration of companies’ competitive advantage and cyclical adjusted real return on invested capital (ROIC) with a 10-year average. We used a new methodology to develop an indicator of the long-term expected return on the equity market at the aggregate level, considering the duration of the competitive advantage of companies. Our results showed a strong correlation between the estimated implied return on equity (IRE) of current stock prices and realized returns of the 10-year real total return of the index.

Learning Latent Multi-Criteria Ratings From User Reviews for Recommendations

Multi-criteria recommender systems have been increasingly useful for helping consumers identify the most relevant items based on different dimensions of user experiences and highlighting their most valued features. Therefore, researchers have proposed various multi-criteria models to improve recommendation performance. However, most of the existing methods utilize only multi-criteria ratings explicitly provided by the users. Note that explicit multi-criteria ratings are sparse and have the problem of missing values. User reviews, on the other hand, contain richer information of user experiences and reveal multi-dimensional user preferences. Therefore, we propose to use latent multi-criteria ratings generated from user reviews, as opposed to explicit multi-criteria ratings, to provide recommendations and capture latent complex heterogeneous user preferences. Specifically, we propose two novel models for the latent multi-criteria rating generation process: the one-stage model LatentMC-1S that utilizes document hashing method to directly compute latent ratings and the two-stage model LatentMC-2S that uses GRU and Gumbel-Softmax for indirect rating generation. Extensive experiments show that the proposed latent multi-criteria rating approaches outperform explicit ratings across different datasets and performance measures. We also show that latent multi-criteria ratings could be used for imputing missing explicit multi-criteria ratings and thus further improving multi-criteria recommender systems.

Deep visual social distancing monitoring to combat COVID-19: A comprehensive survey.

Since the start of the COVID-19 pandemic, social distancing (SD) has played an essential role in controlling and slowing down the spread of the virus in smart cities. To ensure the respect of SD in public areas, visual SD monitoring (VSDM) provides promising opportunities by (i) controlling and analyzing the physical distance between pedestrians in real-time, (ii) detecting SD violations among the crowds, and (iii) tracking and reporting individuals violating SD norms. To the authors’ best knowledge, this paper proposes the first comprehensive survey of VSDM frameworks and identifies their challenges and future perspectives. Typically, we review existing contributions by presenting the background of VSDM, describing evaluation metrics, and discussing SD datasets. Then, VSDM techniques are carefully reviewed after dividing them into two main categories: hand-crafted feature-based and deep-learning-based methods. A significant focus is paid to convolutional neural networks (CNN)-based methodologies as most of the frameworks have used either one-stage, two-stage, or multi-stage CNN models. A comparative study is also conducted to identify their pros and cons. Thereafter, a critical analysis is performed to highlight the issues and impediments that hold back the expansion of VSDM systems. Finally, future directions attracting significant research and development are derived.

Integrated one-stage models considering undesirable outputs and weighting preference in slacks-based measure of efficiency and superefficiency

In the evaluation scenario with undesirable outputs, the SBM-Undesirable model is usually used together with the SuperSBM-Undesirable model to fully differentiate efficient and inefficient decision-making units (DMUs). The existing SuperSBM-Undesirable model has two main forms, but both have problems in implementation. One form may be infeasible under the variable returns to scale (VRS); Another form has an inappropriate definition for undesirable output efficiency. These two problems are seriously ignored in the existing applications. In this paper, we first propose an improved SuperSBM-Undesirable model with strongly efficient projections, which is feasible under the constant returns to scale (CRS) or VRS technology. Then, by integrating the SBM-Undesirable and the improved SuperSBM-Undesirable models, we focus on proposing a concise, precise and practical one-stage model considering undesirable outputs to differentiate all DMUs and determine their strongly efficient projections simultaneously. The proposed one-stage model only contains essential decision variables and constraints, thereby effectively conserving computational time for large-scale practical applications. Further, by introducing the weighting preference of inputs and outputs, we construct the one-stage model with weighting preference to differentiate the importance of different indicators. Finally, through several data experiments, the superiority of our models in computational results and computational scale are verified.

A deep learning based system for real-time detection and sorting of earthworm cocoons

Vermicompost, created by earthworms after eating and digesting organic waste, plays an important role as an organic fertiliser in sustainable agriculture. In this study, a deep learning-based smart system was developed to separate earthworm cocoons used in the production of vermicompost from the compost and return it to production. In the first stage of the study, a dataset containing 1000 images of cocoons was created. The cocoons in each image were labeled and training was performed using a deep learning architecture, one-stage and two-stage models. The models were trained over 2000 epochs with a learning rate of 0.01. From the experimental results, faster R-CNN with ResNet50-FPN model detected the earthworm cocoons better compared to other models. The best performance was obtained by this model with an average precision (AP) of 0.89. In the other stage of the study, the cocoons detected by the software were separated from the compost using a specially designed conveyor belt system. In this process, the detected cocoons are separated from the compost using 10 pneumatic valves that spray air at the separation point. The study is the first of its kind that enables earthworm cocoons to be returned to production with the use of a real-time intelligent system. It also contributes to the literature on small object detection using deep learning.

Are farmers technically efficient in growing sorghum crops?: evidence from western part of Ethiopia Gudeya Bila district

Agriculture provides the majority of Ethiopian rural households with their principal source of income, yet it performs poorly. There is a rise in food insecurity as well as a decline in productivity as a result of this. Even if sorghum production in Ethiopia is increasing rapidly, it requires an efficient level of output to ensure high levels of productivity and profit. Hence the goal of this study was to examine the technical efficiency of sorghum production and its determinants in the Gudeya Bila area in western Ethiopia, utilizing primary data obtained through semi-structured questionnaires from 203 randomly selected households. The study utilized one-stage stochastic frontier production model to investigate the technical efficiency and its determinants. The mean technical efficiency of the homes was 45.64 percent, according to the results of the stochastic frontier of the parametric approach. These results suggest that farmers in the research area are technically inefficient in sorghum by 56.36 percent on average. Weeding frequency, farm size, and cell phone use were also key factors of technical efficiency in a one-stage stochastic frontier approach. As a result, the study reveals that by enhancing technological efficiency, it may be possible to increase production to the level of potential output. Ensure mobile information service, raise knowledge about intensive land use, subsidize chemical inputs, and expand educational possibilities in the research region are some of the numerous strategies to improve technical efficiency.

Deep learning for component fault detection in electricity transmission lines

Component fault detection and inventory are one of the most significant bottlenecks facing the electricity transmission and distribution utility establishments especially in developing countries for delivery of efficient services to the customers and to ensure proper asset audit and management for network optimization and load forecasting. For lack of technology and data, insecurity, the complexity associated with traditional methods, untimeliness, and general human cost, electricity assets monitoring, and management have remained a big problem in many developing countries. In view of this, we explored the use of oblique UAV imagery with high spatial resolution and fine-tuned deep Convolutional Neural Networks (CNNs) for automatic faulty component inspection and inventory in an Electric power transmission network (EPTN). This study investigated the capability of the Single Shot Multibox Detector (SSD), a one-stage object detection model on the electric transmission power line imagery to localize, detect and classify faults. Our proposed neural network model is a CNN based on a multiscale layer feature pyramid network (FPN) using aerial image patches and ground truth to localise and detect faults through a one-phase procedure. The SSD Rest50 architecture variation performed the best with a mean Average Precision (mAP) of 89.61%. All the developed SSD-based models achieve a high precision rate and low recall rate in detecting faulty components, thus achieving acceptable balance levels of F1-score and representation. We have established in this paper that combined use of UAV imagery and computer vision presents a low-cost method for easy and timely electricity asset inventory, especially in developing countries. This study also provides the guide to various considerations when adopting this technology in terms of the choice of deep learning architecture, adequate training samples over multiple fault characteristics, effects of data augmentation, and balancing of intra-class heterogeneity.

Sentiment Analysis Pedulilindungi Tweet Using Support Vector Machine Method

Pedulilindungi application has many benefits but many controversies arise in the community. Various opinions in the form of tweets were expressed by the public, both positive and negative opinions. In this study, the objective is to make a classification model to classify tweets into two types of sentiment, namely positive and negative. The model is made in several stages, namely data retrieval, data filtering, data labeling, data preprocessing, splitting data train and data test, feature selection using Information Gain and Genetic Algorithm, and then classification using the SVM method. The model using two-stage feature selection and SVM method, obtained an accuracy value of 64.08% with 841 features and processing time of 0.033 seconds with 9.6% CPU usage. The model with two-stage feature selection is more efficient and effective than the one-stage feature selection model whose accuracy value is only 60.56% with 1800 features and a processing time of 0.044 seconds with 15.4% CPU usage.

WGNAM: whole-genome nested association mapping

Key messageA powerful QTL analysis method for nested association mapping populations is presented. Based on a one-stage multi-locus model, it provides accurate predictions of founder specific QTL effects.Nested association mapping (NAM) populations have been created to enable the identification of quantitative trait loci (QTL) in different genetic backgrounds. A whole-genome nested association mapping (WGNAM) method is presented to perform QTL analysis in NAM populations. The WGNAM method is an adaptation of the multi-parent whole genome average interval mapping approach where the crossing design is incorporated through the probability of inheriting founder alleles for every marker across the genome. Based on a linear mixed model, this method provides a one-stage analysis of raw phenotypic data, molecular markers, and crossing design. It simultaneously scans the whole-genome through an iterative process leading to a model with all the identified QTL while keeping the false positive rate low. The WGNAM approach was assessed through a simulation study, confirming to be a powerful and accurate method for QTL analysis for a NAM population. This novel method can also accommodate a multi-reference NAM (MR-NAM) population where donor parents are crossed with multiple reference parents to increase genetic diversity. Therefore, a demonstration is presented using a MR-NAM population for wheat (Triticum aestivum L.) to perform a QTL analysis for plant height. The strength and size of the putative QTL were summarized enhancing the understanding of the QTL effects depending on the parental origin. Compared to other methods, the proposed methodology based on a one-stage analysis provides greater power to detect QTL and increased accuracy in the estimation of their effects. The WGNAM method establishes the basis for accurate QTL mapping studies for NAM and MR-NAM populations.

Surround-Net: A Multi-Branch Arbitrary-Oriented Detector for Remote Sensing

With the development of oriented object detection technology, especially in the area of remote sensing, significant progress has been made, and multiple excellent detection architectures have emerged. Oriented detection architectures can be broadly divided into five-parameter systems and eight-parameter systems that encounter the periodicity problem of angle regression and the discontinuous problem of vertex regression during training, respectively. Therefore, we propose a new multi-branch anchor-free one-stage model that can effectively alleviate the corner case when representing rotating objects, called Surround-Net. The creative contribution submitted in this paper mainly includes three aspects. Firstly, a multi-branch strategy is adopted to make the detector choose the best regression path adaptively for the discontinuity problem. Secondly, to address the inconsistency between classification and quality estimation (location), a modified high-dimensional Focal Loss and a new Surround IoU Loss are proposed to enhance the unity ability of the features. Thirdly, in the refined process after backbone feature extraction, a center vertex attention mechanism is adopted to deal with the environmental noise introduced in the remote sensing images. This type of auxiliary module is able to focus the model’s attention on the boundary of the bounding box. Finally, extensive experiments were carried out on the DOTA dataset, and the results demonstrate that Surround-Net can solve regression boundary problems and can achieve a more competitive performance (e.g., 75.875 mAP) than other anchor-free one-stage detectors with higher speeds.

Semantic segmentation and photogrammetry of crowdsourced images to monitor historic facades

Crowdsourced images hold information could potentially be used to remotely monitor heritage sites, and reduce human and capital resources devoted to on-site inspections. This article proposes a combination of semantic image segmentation and photogrammetry to monitor changes in built heritage sites. In particular, this article focuses on segmenting potentially damaging plants from the surrounding stone masonry and other image elements. The method compares different backend models and two model architectures: (i) a one-stage model that segments seven classes within the image, and (ii) a two-stage model that uses the results from the first stage to refine a binary segmentation for the plant class. The final selected model can achieve an overall IoU of 66.9% for seven classes (54.6% for one-stage plant, 56.2% for two-stage plant). Further, the segmentation output is combined with photogrammetry to build a 3D segmented model to measure the area of biological growth. Lastly, the main findings from this paper are: (i) With the help of transfer learning and proper choice of model architecture, image segmentation can be easily applied to analyze crowdsourcing data. (ii) Photogrammetry can be combined with image segmentation to alleviate image distortions for monitoring purpose. (iii) Beyond the measurement of plant area, this method has the potential to be easily transferred into other tasks, such as monitoring cracks and erosion, or as a masking tool in the photogrammetry workflow.

Research on Object Detection of PCB Assembly Scene Based on Effective Receptive Field Anchor Allocation.

Vision-based object detection of PCB (printed circuit board) assembly scenes is essential in accelerating the intelligent production of electronic products. In particular, it is necessary to improve the detection accuracy as much as possible to ensure the quality of assembly products. However, the lack of object detection datasets in PCB assembly scenes is the key to restricting intellectual PCB assembly research development. As an excellent representative of the one-stage object detection model, YOLOv3 (you only look once version 3) mainly relies on placing predefined anchors on the three feature pyramid layers and realizes recognition and positioning using regression. However, the number of anchors distributed in each grid cell of different scale feature layers is usually the same. The ERF (effective receptive field) corresponding to the grid cell at different locations varies. The contradiction between the uniform distribution of fixed-size anchors and the ERF size range in different feature layers will reduce the effectiveness of object detection. Few people use ERF as a standard for assigning anchors to improve detection accuracy. To address this issue, firstly, we constructed a PCB assembly scene object detection dataset, which includes 21 classes of detection objects in three scenes before assembly, during assembly, and after assembly. Secondly, we performed a refined ERF analysis on each grid of the three output layers of YOLOv3, determined the ERF range of each layer, and proposed an anchor allocation rule based on the ERF. Finally, for the small and difficult-to-detect TH (through-holes), we increased the context information and designed improved-ASPP (Atrous spatial pyramid pooling) and channel attention joint module. Through a series of experiments on the object detection dataset of the PCB assembly scene, we found that under the framework of YOLOv3, anchor allocation based on ERF can increase mAP (mean average precision) from 79.32% to 89.86%. At the same time, our proposed method is superior to Faster R-CNN (region convolution neural network), SSD (single shot multibox detector), and YOLOv4 (you only look once version 4) in the balance of high detection accuracy and low computational complexity.

Bayesian two-part modeling of phytoplankton biomass and occurrence

Phytoplankton biomass data often involve zero outcomes preventing a description by continuous distributions with positive support such as the lognormal distribution commonly used to describe ecological data. Two usual solutions: ignoring the zeroes and adding a small positive number to all outcomes, induce bias and reduce predictive power. To address these shortcomings, we design a Bayesian two-part model with a binary component for presence or absence and a continuous component involving a lognormal model for non-zero biomass. We specify two equations relating species-specific occurrence probabilities and expected log-biomasses when present to potential covariates, with spike-and-slab priors imposed on linear effects to selectively discard the irrelevant predictors. We analyze the biomass data of 74 phytoplankton (57 diatoms and 17 dinoflagellates) recorded weekly at Station L4 (Western English Channel, UK) between April 2003 and December 2009, along with measurements of abiotic covariates. Our results disclose different combinations of environmental predictors for the occurrence and the biomass of individual species. Overall, the occurrence of dinoflagellates is associated with higher temperature and irradiance levels compared to diatoms, with virtually no dependence on nutrient concentrations. Irradiance emerges as the key predictor of biomass when species are present. Optimum temperatures for biomass accumulation and temperature sensitivities vary widely among and within functional types. Compared to one-stage models based on usual zero-handling approaches, our two-part model stands out with higher prediction accuracy. The two-part modeling approach provides a valuable framework for decoupling the predictors of species occurrence and abundance from observational data.

Multi-Model Medical Image Segmentation Using Multi-Stage Generative Adversarial Networks

Image segmentation is a challenging problem in medical applications. Medical imaging has become an integral part of machine learning research, as it enables inspecting interior human body with no surgical intervention. Much research has been conducted to study brain segmentation. However, prior studies usually employ one-stage models to segment brain tissues, which could lead to a significant information loss. In this paper, we propose a multi-stage Generative Adversarial Network (<inline-formula> <tex-math notation="LaTeX">$GAN$ </tex-math></inline-formula>) model to resolve existing issues of one-stage models. To do this, we apply a <i>coarse-to-fine</i> method to improve brain segmentation using a multi-stage <inline-formula> <tex-math notation="LaTeX">$GAN$ </tex-math></inline-formula>. In the first stage, our model generates a <i>coarse</i> outline for both the background and brain tissues. Then, in the second stage, the model generates a <i>refine</i> outline for the white matter (<inline-formula> <tex-math notation="LaTeX">$WM$ </tex-math></inline-formula>), gray matter (<inline-formula> <tex-math notation="LaTeX">$GM$ </tex-math></inline-formula>), and cerebrospinal fluid (<inline-formula> <tex-math notation="LaTeX">$CSF$ </tex-math></inline-formula>). We perform a fusion of the <i>coarse</i> and <i>refine</i> outlines to achieve high results. Despite using very limited data, we obtain an improved Dice Coefficient (DC) accuracy of up to 5&#x0025; compared to one-stage models. We conclude that our model is more efficient and accurate in practice for brain segmentation of both infants and adults. In addition, we observe that our multi-stage model is 2.69&#x2013;13.93 minutes faster than prior models. Moreover, our multi-stage model achieves higher performance with only a few-shot learning, in which only limited labeled data is available. Therefore, for medical images, our solution is applicable to a wide range of image segmentation applications for which convolution neural networks and one-stage methods have failed. This helps to advance the process of analyzing brain images, thus providing many advantages to the healthcare system, especially in critical health situations where urgent intervention is needed.

ForestDet: Large-Vocabulary Long-Tailed Object Detection and Instance Segmentation

Object detection and instance segmentation with a large number of object categories and long-tailed data distribution are challenging for most existing deep learning models. As the number of classes increases, the outputs of a classifier become sensitive to likely noisy logits, which can easily result in an incorrect recognition. To alleviate the large-vocabulary problem, we cluster fine-grained classes into coarser parent classes and then build a classification tree to classify an object into a fine-grained class via its parent class. Because the number of parent class is much fewer, their logits are more stable to suppress the wrong/noisy logits existed in the fine-grained class nodes. Due to a variety of ways for clustering fine-grained classes into parent classes, we can further construct multiple trees to build a classification forest where each single tree contributes its vote to the fine-grained classification. Moreover, a simple yet effective resampling method, termed as NMS Resampling, is proposed aiming at solving the long tail (data imbalance) problem. Our method, coined as ForestDet, serves as a plug-and-play module, which can be readily employed in both one-stage and two-stage object recognition models for recognizing more than 1000 categories. Extensive experiments are conducted on the large vocabulary dataset LVIS. Compared to the Mask R-CNN baseline, our two-stage counterpart Forest R-CNN significantly boosts the performance by 11.5&#x0025; and 3.9&#x0025; AP improvements on the rare categories and overall categories, respectively. Compared to the RetinaNet baseline, our one-stage counterpart Forest RetinaNet improves 2.1&#x0025; AP on overall categories. Moreover, we achieve state-of-the-art results on the LVIS dataset. <i>Code and models are available at <uri>https://github.com/JialianW/Forest_RCNN</uri>.</i>

SODAR: Segmenting Objects by Dynamically Aggregating Neighboring Mask Representations.

Recent state-of-the-art one-stage instance segmentation model SOLO divides the input image into a grid and directly predicts per grid cell object masks with fully-convolutional networks, yielding comparably good performance as traditional two-stage Mask R-CNN yet enjoying much simpler architecture and higher efficiency. We observe SOLO generates similar masks for an object at nearby grid cells, and these neighboring predictions can complement each other as some may better segment certain object part, most of which are however directly discarded by non-maximum-suppression. Motivated by the observed gap, we develop a novel learning-based aggregation method that improves upon SOLO by leveraging the rich neighboring information while maintaining the architectural efficiency. The resulting model is named SODAR. Unlike the original per grid cell object masks, SODAR is implicitly supervised to learn mask representations that encode geometric structure of nearby objects and complement adjacent representations with context. The aggregation method further includes two novel designs: 1) a mask interpolation mechanism that enables the model to generate much fewer mask representations by sharing neighboring representations among nearby grid cells, and thus saves computation and memory; 2) a deformable neighbour sampling mechanism that allows the model to adaptively adjust neighbor sampling locations thus gathering mask representations with more relevant context and achieving higher performance. SODAR significantly improves the instance segmentation performance, e.g., it outperforms a SOLO model with ResNet-101 backbone by 2.2 AP on COCO test set, with only about 3% additional computation. We further show consistent performance gain with the SOLOv2 model.