Accurate Counting Research Articles

A method for counting fish based on improved YOLOv8

In industrial aquaculture, accurately counting fish in real-time is crucial for optimizing feeding strategies, preventing disease, and managing water quality. Current methods utilizing sensors, acoustics, machine learning, and density map regression face challenges such as high costs, invasiveness, and computational complexity. To address these limitations, we propose the YOLOv8n-MEMAGD method for accurate real-time fish counting. This method enhances YOLOv8 by incorporating the GELU activation function, MPDIoU for localization loss, and the C2f-FAM and C2f-MSCA modules into the backbone network. Additionally, the neck network is redesigned with a gather-and-distribution mechanism. Experimental results under land-based industrial aquaculture conditions demonstrate that YOLOv8n-MEMAGD achieved a mean absolute error (MAE) of 2.28 and a root mean square error (RMSE) of 2.84, even in challenging conditions such as fish overlap, aggregation, and background confounding. Compared to YOLOv8n, the proposed method increased average precision (AP50) for fish detection by 6.2 %, and significantly reduced MAE and RMSE by 61.4 % and 65.2 %, respectively, compared to CSRNet. Additionally, the method achieved a frame rate of 61 frames per second (FPS) when the number of fish ranged from 79 to 91, representing a 390.4 % increase over CSRNet. By comparing heatmaps, the proposed method demonstrates more effective detection of fish edge contours than current advanced algorithms. In conclusion, the proposed method shows promise for application in aquacultural scenarios with higher turbidity and larger number of fish.

Automatic localization of image semantic patches for crop disease recognition

Crop disease recognition plays a crucial role in agricultural production. However, disease images are large in scale and have a lot of redundant information, which reduces the effectiveness of deep neural networks in extracting diseases. To address the above issues and considering that not all image regions are relevant to disease recognition, this study proposes an efficient crop disease recognition method with dynamic reduction of image redundancy. The method is a two-stage process. In the first stage, we employ the lightweight CA-AnchorNet, which incorporates coordinate attention, to swiftly generate a feature map of the affected crop areas. Subsequently, class activation maps (CAMs) are utilized to identify the disease feature regions, highlighting areas that exhibit class discriminability. These regions are then mapped to a higher resolution from the original lower-resolution image, and the target patch is extracted. In the second stage, these local semantic patches, characterized by reduced spatial redundancy, are fed into the lightweight PatchNet for accurate recognition. PatchNet incorporates the Inception-C module and the ACON-C activation function. These features enhance the model's ability to express multi-scale and non-linear characteristics of crop disease features. This method does not require manually annotated region boxes and achieves an identification accuracy of 99.86 % on a 12-class crop disease dataset with complex environments. The parameter count is only 0.98 M. This method has the characteristics of accurate localization and low parameter count, and can be used for effective and high-precision recognition of crop diseases in complex environments.

Sub-window inference: A novel approach for improved sheep counting in high-density aerial images

Traditional sheep counting methods are labour-intensive, time-consuming, and potentially disruptive to sheep behaviour. Unmanned aerial vehicles (UAVs) and machine learning (ML) techniques have emerged to address these challenges by automating the process. However, these solutions face difficulties due to low object-to-image-pixel ratios and high object densities in images captured for sheep counting, which can compromise detection and counting accuracy. In this study, we introduce and evaluate a novel approach, sub-window inference, designed to increase the object-to-image-pixel ratios, thereby enhancing the performance of existing object detection and segmentation models. Our method was compared to four other object counting techniques, demonstrating superior performance in terms of a reduced mean absolute error (MAE) of 3.21 sheep and a mean absolute percentage error (MAPE) of 1.27%. Furthermore, our findings indicate that incorporating random cropping data augmentation during model training significantly enhances both detection and counting accuracy. It is important to note that a limitation of sub-window inference is that it does not facilitate real-time count predictions. Overall, our proposed method of sub-window inference reduces the MAE in automated sheep counting techniques involving UAVs and ML, presenting a highly effective solution that benefits sheep farmers.

COTTON-YOLO: Enhancing Cotton Boll Detection and Counting in Complex Environmental Conditions Using an Advanced YOLO Model

This study aims to enhance the detection accuracy and efficiency of cotton bolls in complex natural environments. Addressing the limitations of traditional methods, we developed an automated detection system based on computer vision, designed to optimize performance under variable lighting and weather conditions. We introduced COTTON-YOLO, an improved model based on YOLOv8n, incorporating specific algorithmic optimizations and data augmentation techniques. Key innovations include the C2F-CBAM module to boost feature recognition capabilities, the Gold-YOLO neck structure for enhanced information flow and feature integration, and the WIoU loss function to improve bounding box precision. These advancements significantly enhance the model’s environmental adaptability and detection precision. Comparative experiments with the baseline YOLOv8 model demonstrated substantial performance improvements with COTTON-YOLO, particularly a 10.3% increase in the AP50 metric, validating its superiority in accuracy. Additionally, COTTON-YOLO showed efficient real-time processing capabilities and a low false detection rate in field tests. The model’s performance in static and dynamic counting scenarios was assessed, showing high accuracy in static cotton boll counting and effective tracking of cotton bolls in video sequences using the ByteTrack algorithm, maintaining low false detections and ID switch rates even in complex backgrounds.

The Impact of Tree Density on Automated Oil Palm Tree Counting Accuracy

The rapid development of oil palm plantations in Kualuh Leidong Subdistrict, Labuhanbatu Utara District, North Sumatra Province has led to an increased need for effective and efficient monitoring and supervision of oil palm trees. One method that supports such monitoring and supervision is the use of an automatic counting method using orthophoto data. This orthophoto data was used for automatic tree counting using a deep learning method with the Faster R-CNN algorithm. The study considered two planting patterns: regular planting patterns with spacing of 4 to 9 meters, and random planting patterns with varying spacing. Data processing involved an epoch value of 80 and a batch size value of 4. The accuracy of the automatic oil palm tree counting was evaluated based on the density of the spacing between trees with reference to the ground truth. The findings indicated that the deep learning Faster R-CNN algorithm achieved higher accuracy in automatic calculations for regular planting patterns.

Viability of high-frequency environmental DNA (eDNA) sampling as a fish enumeration tool

Despite growing interest in environmental DNA (eDNA) monitoring as a low-cost, non-invasive tool for monitoring fish presence and relative abundance, quantitative comparisons between fish abundance and eDNA are still rare, especially for anadromous species of conservation concern. To examine the relationship between eDNA data and management-relevant abundance data, we assessed eDNA abundance of river herring, an ecologically relevant anadromous sister-species pair, alongside sonar-generated fish counts at a well-characterized site in the Choptank River, Chesapeake Bay, MD. Over the 2021 spawning season, there was a strong correlation between daily eDNA concentration and daily counts of fish swimming upstream (Spearman’s rho = 0.86). A generalized additive model for location scale and shape (GAMLSS) was used to estimate fish count as a function of flow-corrected eDNA concentration and relevant covariates. Model- and sonar-generated run counts were 729,385 and 733,208 fish respectively, differing by just 0.5 %. A sub-sampling analysis showed that a model built on at least five weeks of data could utilize eDNA abundance to produce accurate estimates for the remainder of the season, while a reduction in eDNA sampling frequency substantially reduced run count accuracy. This work emphasizes the utility of quantitative eDNA monitoring for river herring populations and shows the necessity of high-frequency eDNA sampling to estimate run count.

Green pepper fruits counting based on improved DeepSort and optimized Yolov5s.

Green pepper yield estimation is crucial for establishing harvest and storage strategies. This paper proposes an automatic counting method for green pepper fruits based on object detection and multi-object tracking algorithm. Green pepper fruits have colors similar to leaves and are often occluded by each other, posing challenges for detection. Based on the YOLOv5s, the CS_YOLOv5s model is specifically designed for green pepper fruit detection. In the CS_YOLOv5s model, a Slim-Nick combined with GSConv structure is utilized in the Neck to reduce model parameters while enhancing detection speed. Additionally, the CBAM attention mechanism is integrated into the Neck to enhance the feature perception of green peppers at various locations and enhance the feature extraction capabilities of the model. According to the test results, the CS_YOLOv5s model of mAP, Precision and Recall, and Detection time of a single image are 98.96%, 95%, 97.3%, and 6.3 ms respectively. Compared to the YOLOv5s model, the Detection time of a single image is reduced by 34.4%, while Recall and mAP values are improved. Additionally, for green pepper fruit tracking, this paper combines appearance matching algorithms and track optimization algorithms from SportsTrack to optimize the DeepSort algorithm. Considering three different scenarios of tracking, the MOTA and MOTP are stable, but the ID switch is reduced by 29.41%. Based on the CS_YOLOv5s model, the counting performance before and after DeepSort optimization is compared. For green pepper counting in videos, the optimized DeepSort algorithm achieves ACP (Average Counting Precision), MAE (Mean Absolute Error), and RMSE (Root Mean Squared Error) values of 95.33%, 3.33, and 3.74, respectively. Compared to the original algorithm, ACP increases by 7.2%, while MAE and RMSE decrease by 6.67 and 6.94, respectively. Additionally, Based on the optimized DeepSort, the fruit counting results using YOLOv5s model and CS_YOLOv5s model were compared, and the results show that using the better object detector CS_YOLOv5s has better counting accuracy and robustness.

The use of image classification to estimate flamingo abundance from aerial, drone and satellite imagery

The Lesser Flamingo Phoeniconaias minor is a Near Threatened species known to be highly gregarious and that can concentrate in large numbers at core foraging and breeding sites, yet is also known to disperse widely in search of suitable foraging habitat. The density of flamingos at occupied sites, together with frequent dispersal between foraging sites, poses significant challenges to obtaining reliable counts and associated population estimates. Our study tested the use of unmanned aerial vehicles (drones) imagery and high-resolution satellite imagery for producing density estimates of Lesser Flamingos at the only known breeding site in South Africa (Kamfers Dam, Northern Cape Province). We implemented a supervised classification approach using multiple supervised classification algorithms, including xgbTree, randomForest, nnet and GBM. The four supervised classification models performed very well at classifying flamingos from both drone and satellite imagery, with classification accuracies ranging between 0.98 and 0.99. The inclusion of an additional predictor (other than RGB bands) yielded the highest variable importance score and greatly increased model accuracy and predictive performance. Estimates of adult flamingos from drone imagery acquired across nesting mudflats on 5 February (n = 1 320) and 14 February (n = 330) 2019 suggested a 75% decline in incubating adults. Furthermore, a total of 16 134 (15 973–16 295) flamingos were estimated from drone imagery on 1 March 2019, and 17 291 (17 118–17 290) flamingos from satellite imagery on 19 March 2019. The modelled size of the creche across dates suggested a loss of 234 birds (4.5%) during the period of assessment, with the highest count estimating 5 218 juveniles. Our study successfully implemented machine learning classifiers that were able to detect and derive highly accurate counts of Lesser Flamingos from both drone and satellite imagery. This use of machine learning could greatly improve the efficiency of monitoring and associated conservation efforts underway for the Lesser Flamingo.

Deterministic preparation of a dual-species two-ion crystal

The demand for efficient preparation methods for dual-species ion crystals is rapidly expanding across quantum technology and fundamental physics applications with trapped ions. We present a deterministic and efficient technique to produce such crystals, utilizing the segmented structure of a linear Paul trap. By precisely tailoring the trapping potentials, we can split, move, and discard parts of an ion chain. This process is automated in a sequence that converts a larger ion sample into the desired configuration. A critical component of our approach is the accurate identification of crystal constituents. This is achieved by matching the measured positions of fluorescing ions against theoretical expectations for larger crystals, thus facilitating the detection of nonfluorescing ions and enabling accurate ion counting. We demonstrate that our method reliably produces two-ion crystals within minutes. These results represent a significant advance in the production of two-species ion crystals with applications ranging from quantum logic spectroscopy and optical clocks to quantum computing and simulations with trapped ions. Published by the American Physical Society 2024

High-accuracy occupancy counting at crowded entrances for smart buildings

The Buildings Global Status Report (Buildings-GSR) highlights the significant impact of the buildings and construction sector on global climate change. This sector is responsible for approximately 34% of global energy consumption and 37% of carbon dioxide (CO2) emissions. Smart building is a promising solution to achieve Zero Net goals. High-accuracy occupancy counting plays a crucial role in enhancing building energy efficiency, improving environmental comfort, and reducing disease transmission. Occupancy information can serve as feedback in a closed-loop building control strategy, enabling occupant-centric control (OCC). Even though existing studies have made developments, they struggle to obtain high-accuracy occupancy counting because of many challenges, e.g., cumulative errors and occlusion. To address these challenges, a novel deep learning-based occupancy counting method by edge cameras at crowded entrances is introduced. We annotate the occupancy tracking dataset and train tracking neural networks to count occupants, achieving 80.4% and 69.9% MOTA in two different cameras. We fuse multi-camera information to achieve high performance. Our methods achieve an occupancy counting accuracy of approximately 98.1% with two real-scene surveillance cameras, surpassing the existing occupancy counting methods by 7.7%. Besides, we design an occupant-centric control method to save energy in buildings. Our experiments demonstrate that our occupancy counting method achieves better performance on a real-world dataset, compared to other methods. Our experiments verify that our occupant-centric control method has the potential for building energy saving about 17%-25% in a room. Our study deepens the insights into occupancy measurement, paving the way toward energy-efficient buildings and comfortable indoor environments. We make our code available at https://github.com/kailaisun/Head-tracking-for-occupancy-counting.

One-Shot Any-Scene Crowd Counting With Local-to-Global Guidance.

Due to different installation angles, heights, and positions of the camera installation in real-world scenes, it is difficult for crowd counting models to work in unseen surveillance scenes. In this paper, we are interested in accurate crowd counting based on the data collected by any surveillance camera, that is to count the crowd from any scene given only one annotated image from that scene. To this end, we firstly pose crowd counting as a one-shot learning task. Through the metric-learning, we propose a simple yet effective method that firstly estimates crowd characteristics and then transfers them to guide the model to count the crowd. Specifically, to fully capture these crowd characteristics of the target scene, we devise the Multi-Prototype Learner to learn the prototypes of foreground and density from the limited support image using the Expectation-Maximization algorithm. To learn the adaptation capability for any unseen scene, estimated multi prototypes are proposed to guide the crowd counting of query images in a local-to-global way. CNN is utilized to activate the local features. And transformer is introduced to correlate global features. Extensive experiments on three surveillance datasets suggest that our method outperforms the SOTA methods in the few-shot crowd counting.

A High-Throughput Method for Accurate Extraction of Intact Rice Panicle Traits.

Rice panicle traits serve as critical indicators of both yield potential and germplasm resource quality. However, traditional manual measurements of these traits, which typically involve threshing, are not only laborious and time-consuming but also prone to introducing measurement errors. This study introduces a high-throughput and nondestructive method, termed extraction of panicle traits (EOPT), along with the software Panicle Analyzer, which is designed to assess unshaped intact rice panicle traits, including the panicle grain number, grain length, grain width, and panicle length. To address the challenge of grain occlusion within an intact panicle, we define a panicle morphology index to quantify the occlusion levels among the rice grains within the panicle. By calibrating the grain number obtained directly from rice panicle images based on the panicle morphology index, we substantially improve the grain number detection accuracy. For measuring grain length and width, the EOPT selects rice grains using an intersection over union threshold of 0.8 and a confidence threshold of 0.7 during the grain detection process. The mean values of these grains were calculated to represent all the panicle grain lengths and widths. In addition, EOPT extracted the main path of the skeleton of the rice panicle using the Astar algorithm to determine panicle lengths. Validation on a dataset of 1,554 panicle images demonstrated the effectiveness of the proposed method, achieving 93.57% accuracy in panicle grain counting with a mean absolute percentage error of 6.62%. High accuracy rates were also recorded for grain length (96.83%) and panicle length (97.13%). Moreover, the utility of EOPT was confirmed across different years and scenes, both indoors and outdoors. A genome-wide association study was conducted, leveraging the phenotypic traits obtained via EOPT and genotypic data. This study identified single-nucleotide polymorphisms associated with grain length, width, number per panicle, and panicle length, further emphasizing the utility and potential of this method in advancing rice breeding.

CTHNet: a network for wheat ear counting with local-global features fusion based on hybrid architecture.

Accurate wheat ear counting is one of the key indicators for wheat phenotyping. Convolutional neural network (CNN) algorithms for counting wheat have evolved into sophisticated tools, however because of the limitations of sensory fields, CNN is unable to simulate global context information, which has an impact on counting performance. In this study, we present a hybrid attention network (CTHNet) for wheat ear counting from RGB images that combines local features and global context information. On the one hand, to extract multi-scale local features, a convolutional neural network is built using the Cross Stage Partial framework. On the other hand, to acquire better global context information, tokenized image patches from convolutional neural network feature maps are encoded as input sequences using Pyramid Pooling Transformer. Then, the feature fusion module merges the local features with the global context information to significantly enhance the feature representation. The Global Wheat Head Detection Dataset and Wheat Ear Detection Dataset are used to assess the proposed model. There were 3.40 and 5.21 average absolute errors, respectively. The performance of the proposed model was significantly better than previous studies.

CrowdUNet: Segmentation assisted U-shaped crowd counting network

With the end of the COVID-19 pandemic, the number of pedestrians in various public places has increased dramatically. Estimating the size and density distribution of crowds accurately from images is essential for public safety. At present, there are still many factors that limit the accuracy of dense crowd counting, such as perspective distortion, background clutter and heavy occlusion. To be capable of accurately estimating the crowd size in RGB images, we propose a crowd counting network called CrowdUNet, which is assisted by a segmentation task. It applies the segmentation results to the crowd counts, making the network more focused on the prediction of foreground regions. We combine Swin Transformer Block and CNN to build a Swin Transformer Convolution(STC) module to extract deep semantic features. We analyze the characteristics of crowd images and propose a novel decoder structure called Coordinate Decoder(CD), which better aggregates low and high level features and improve the robustness of the network. In order to obtain accurate regression results, we also propose a regression head with multi-scale receptive fields, which is called Spatial Pyramid Convolution (SPC). Extensive experiments on four challenging crowd counting datasets namely ShanghaiTech A, ShanghaiTech B, UCF p=CC 50, and UCF-QNRF have validated the proposed method.

Age estimation in tortoises: an evaluation of methods for the spur-thighed tortoise (Testudo graeca)

Abstract Age and growth-related parameters provide a better understanding of life-history adaptations and trade-offs and can be studied in the wild through capture-mark-recapture (CMR) or sclerochronology (i.e., skeletochronology and counting growth marks on hard structures). In chelonians, skeletochronology is limited to carcasses, CMR is limited by time and funding constraints, but counting shell growth annuli (SGA count) is a cheaper and faster alternative. A long-term CMR study in Testudo graeca from south-eastern Romania allowed us to compare the age estimated by the two sclerochronology methods using carcasses, and calibrate the results with data obtained from recaptured tortoises. We found that the accuracy of SGA counts significantly decreased with time (i.e., years between captures and older individuals). Skeletochronology showed the best results with long bones (i.e., ulna, radius, femur), but overall underestimated the age when compared to SGA counts; we obtained comparable results from these two methods in tortoises up to 17 years old. The oldest age estimated by skeletochronology alone was 28 years, 30 years by SGA counts, and 40 years by SGA counts when calibrated by CMR. The growth models using age estimations from both skeletochronology and SGA counts were similar, highlighting their usefulness in monitoring growth trajectories. Our study showed that while age estimation in tortoises is challenging, a combination of methods can improve its accuracy.

Optimizing Crowd Counting in Dense Environments Through Curriculum Learning Training Strategy

Counting individuals in highly crowded environments, characterized by thousands of people, has garnered significant attention in recent years, due to the high number of vertical markets wherein such algorithms can prove beneficial, ranging from smart city and transportation to retail sectors, among others. Within this context, in this paper we introduce a novel training methodology tailored for estimating the number of people, ensuring precise counting accuracy in both moderately and highly crowded scenarios. The proposed approach exploits a formulation of the problem based on point detection, where each point represents an individual’s head. Our innovative contributions center around the designing of a novel training strategy employing Curriculum Learning (CL), which aims to replicate the gradual learning process observed in human cognition, training on simpler tasks at the beginning and tackling more complex tasks as the training evolves. In order to evaluate the complexity of each sample image, we propose a novel indicator taking into account both the number of people and their distribution within the image. The experimentation phase encompassed 18 publicly available datasets; the obtained results validate the effectiveness of the proposed approach, surpassing the baseline state-of-the-art point detection by 71% and 70% in terms of Mean Absolute Error (MAE) and Mean Squared Error (MSE), respectively

Large-scale performance assessment of the BG-Counter 2 used with two different mosquito traps

BackgroundMosquitoes are important vectors of pathogens. They are usually collected with CO2-baited traps and subsequently identified by morphology. This procedure is very time-consuming. Automatic counting traps could facilitate timely evaluation of the local risk for mosquito-borne pathogen transmission or decision-making on vector control measures, but the counting accuracy of such devices has rarely been validated in the field.MethodsThe Biogents (BG)-Counter 2 automatically counts mosquitoes by discriminating the size of captured objects directly in the field and transmits the data to a cloud server. To assess the accuracy of this counting device, 27 traps were placed at 19 sampling sites across Germany and used in daily, weekly or bimonthly intervals from April until October 2021. The BG-Counter 2 was attached to a CO2-trap (BG-Pro trap = CO2-Pro) and the same trap was converted to also attract gravid mosquitoes (upside-down BG-Pro trap with a water container beneath = CO2-Pro-gravid). All captured mosquitoes were identified by morphology. The number of females (unfed and gravid), mosquito diversity and the number of identified specimens in relation to the counting data of the BG-Counter were compared between the two trapping devices to evaluate sampling success and counting accuracy.ResultsIn total 26,714 mosquitoes were collected during 854 trap days. The CO2-Pro-gravid trap captured significantly more mosquitoes per trap day for all specimens, gravid females and non-gravid females, while there was no difference in the mosquito diversity. The linear model with the captured mosquitoes as a response and the counted specimens as a predictor explained only a small degree of the variation within the data (R2 = 0.16), but per individual trap the value could reach up to 0.62 (mean R2 = 0.23). The counting accuracy for the daily samples had a significant positive correlation with sample size, resulting in higher accuracy for the CO2-Pro-gravid trap and higher accuracy for sites and sampling months with high mosquito abundance.ConclusionsWhile the accuracy of the BG-Counter 2 is quite low, the device is able to depict mosquito phenology and provide information about local population dynamics.Graphical

CrowdAlign: Shared-weight dual-level alignment fusion for RGB-T crowd counting

The combination of visible and thermal images has been proven to be effective in improving accuracy for crowd counting in illumination-unconstrained scenes. However, the challenging problem of misalignment in RGB-T image pairs has not been extensively explored in this context. This study aims to address the issue of misalignment between RGB and thermal image pairs to enhance the counting accuracy of cross-modal models. Specifically, we propose CrowdAlign, a cross-modal feature alignment fusion network that utilizes a shared-weight strategy for efficient feature extraction. Additionally, CrowdAlign addresses alignment adjustments through two stages: pre-fusion and post-fusion alignment. For pre-fusion feature extraction, we design a dual-level spatial-semantic parallel alignment module, while for post-fusion feature extraction, a low-frequency feature attention filtering module is developed. This two-stage alignment approach enables cross-modal feature alignment without requiring additional supervision. Experiments on the public benchmarks demonstrate that our method is effective under RGB-T misalignment or dark conditions. We hope CrowdAlign will inspire researchers to focus on and explore the issue of misalignment between RGB image and thermal image for cross-modal crowd counting.

Intelligently Counting Agricultural Pests by Integrating SAM with FamNet

The utilization of the large pretrained model (LPM) based on Transformer has emerged as a prominent research area in various fields, owing to its robust computational capabilities. However, there remains a need to explore how LPM can be effectively employed in the agricultural domain. This research aims to enhance agricultural pest detection with limited samples by leveraging the strong generalization performance of the LPM. Through extensive research, this study has revealed that in tasks involving the counting of a small number of samples, complex agricultural scenes with varying lighting and environmental conditions can significantly impede the accuracy of pest counting. Consequently, accurately counting pests in diverse lighting and environmental conditions with limited samples remains a challenging task. To address this issue, the present research suggests a unique approach that integrates the outstanding performance of the segment anything model in class-agnostic segmentation with the counting network. Moreover, by intelligently utilizing a straightforward TopK matching algorithm to propagate accurate labels, and drawing inspiration from the GPT model while incorporating the forgetting mechanism, a more robust model can be achieved. This approach transforms the problem of matching instances in different scenarios into a problem of matching similar instances within a single image. Experimental results demonstrate that our method enhances the accuracy of the FamNet baseline model by 69.17% on this dataset. Exploring the synergy between large models and agricultural scenes warrants further discussion and consideration.

Deep learning-based localization algorithms on fluorescence human brain 3D reconstruction: a comparative study using stereology as a reference

3D reconstruction of human brain volumes at high resolution is now possible thanks to advancements in tissue clearing methods and fluorescence microscopy techniques. Analyzing the massive data produced with these approaches requires automatic methods able to perform fast and accurate cell counting and localization. Recent advances in deep learning have enabled the development of various tools for cell segmentation. However, accurate quantification of neurons in the human brain presents specific challenges, such as high pixel intensity variability, autofluorescence, non-specific fluorescence and very large size of data. In this paper, we provide a thorough empirical evaluation of three techniques based on deep learning (StarDist, CellPose and BCFind-v2, an updated version of BCFind) using a recently introduced three-dimensional stereological design as a reference for large-scale insights. As a representative problem in human brain analysis, we focus on a 4-cm3\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$4~\ ext {-cm}^3$$\\end{document} portion of the Broca’s area. We aim at helping users in selecting appropriate techniques depending on their research objectives. To this end, we compare methods along various dimensions of analysis, including correctness of the predicted density and localization, computational efficiency, and human annotation effort. Our results suggest that deep learning approaches are very effective, have a high throughput providing each cell 3D location, and obtain results comparable to the estimates of the adopted stereological design.