Find research datasets worth reusing

Fig. 14 in Once upon a time in America: recognition of the species of Libitioides from USA, with comments on other American Cosmetidae (Opiliones, Laniatores)

Fig. 14. Libitioides albolineata (Sørensen, 1884), ♂ (USNMENT 01538062) from Palmyra, Virginia. a. Dorsal view of body. b. Left femur and trochanter IV, dorsal view. c. Left metatarsus and tarsomeres of leg I, lateral view. d. Left chelicera dorsal view. Scale bars = 1 mm.

Fig. 15 in Once upon a time in America: recognition of the species of Libitioides from USA, with comments on other American Cosmetidae (Opiliones, Laniatores)

Fig. 15. Libitioides albolineata (Sørensen, 1884) (USNMENT 0153862) from Palmyra, Virginia. a. Male, habitus, dorsal view. b. Same, sinistrolateral view. c. Same, frontal view. d. Same, ventral view. e. Same, panoramic, dorsal view. f. Female, habitus, dorsal view (USNMENT 0153862). Scale bars: a–c = 1 mm; d–f = 2 mm.

Fig. 13 in Once upon a time in America: recognition of the species of Libitioides from USA, with comments on other American Cosmetidae (Opiliones, Laniatores)

Fig. 13. Southeastern USA, marked with abbreviations of American states, showing the distribution of the five species of Libitioides Roewer, 1912 as proposed in this paper.

Fig. 7 in Once upon a time in America: recognition of the species of Libitioides from USA, with comments on other American Cosmetidae (Opiliones, Laniatores)

Fig. 7. Southeastern USA, showing the contrasting occurrences of prevailing body color backgrounds in the four morphs B/C/D/E of Libitioides Roewer, 1912.

Fig. 6 in Once upon a time in America: recognition of the species of Libitioides from USA, with comments on other American Cosmetidae (Opiliones, Laniatores)

Fig. 6. Schemes of scutal yellow markings and other features diagnosing the morphs discussed herein. a. Morph A: Libitioides ornata (Say, 1821). b. Morph B: "clean" Libitioides sayi (Simon, 1879). c. Morph C: Libitioides sayi, striped Texan morph "depressa". d. Morph D: Libitioides albolineata (Sørensen, 1884). e. Morph E: "yellow Libitioides albolineata".

Fig. 11 in Once upon a time in America: recognition of the species of Libitioides from USA, with comments on other American Cosmetidae (Opiliones, Laniatores)

Fig. 11. Southeastern USA, showing the contrasting conformations of the ribs in the four morphs B/C/ D/E of Libitioides Roewer, 1912.

Fig. 10 in Once upon a time in America: recognition of the species of Libitioides from USA, with comments on other American Cosmetidae (Opiliones, Laniatores)

Fig. 10. Southeastern USA, showing the contrasting conformations of the omega stripe in the four morphs B/C/D/E of Libitioides Roewer, 1912.

Fig. 9 in Once upon a time in America: recognition of the species of Libitioides from USA, with comments on other American Cosmetidae (Opiliones, Laniatores)

Fig. 9. Southeastern USA, showing the contrasting conformations of the backbone in the four morphs B/C/D/E of Libitioides Roewer, 1912.

Fig. 22 in Once upon a time in America: recognition of the species of Libitioides from USA, with comments on other American Cosmetidae (Opiliones, Laniatores)

Fig. 22. Libitioides sayi (Simon, 1879), ♂ (MNRJ 58916) from Lake Kirby, Texas, distal part of penis. a. Panoramic dorsal view. b. Ventral view. c. Detail of glans, dextrolateral view. d. Apical view. Scale bars: a = 100 μm; b, d = 50 μm; c = 40 μm.

opencc-zeroJul 2023View details →

dryad40/100

Data from: Both learning and syntax recognition are used by great tits when answering to mobbing calls

Mobbing behavior, in addition to its complex cooperative aspects, is particularly suitable to study the mechanisms implicated in heterospecific communication. Indeed, various mechanisms ranging from pure learning to innate recognition have been proposed. One promising, yet understudied mechanism could be syntax recognition, especially given the latest works published on syntax comprehension in birds. In this experiment, we test whether great tits use both learning and syntax recognition when responding to heterospecifics. In the first part of the experiment, we demonstrate that great tits show different responses to the same heterospecific calls depending on their sympatric status. In a second part, we explore the impact of reorganizing the notes of the heterospecific mobbing calls to fit the syntax of great tits. Great tits showed an increased mobbing response toward the heterospecific calls when they shared their own call organization. Our results corroborate the recent finding that syntactic rules in bird calls may have a strong impact on their communication systems and enlighten how various mechanisms can be used by the same species to respond to heterospecific calls.

opencc-by-4.0Jul 2023View details →

GamER: Gameplay Physiological Signal Dataset for Emotion Recognition

A dataset of physiological signals (EEG, ECG, EDA, EOG and Respiration) collected from 48 subjects while they were playing simple games designed to elicit emotional responses. Data were collected using the Biosignals PLUX Researcher kit, which allows up to 10 hours of signal recordings at up to 3kHz sampling rate and 16-bit resolution per channel, while recording data from up to eight sensors simultaneously. The dataset consists of 3 files: <ol> <li>data.zip: Contains the physiological signals. They are organized in 1 folder per subject, which contains the recorded data in .h5 and .txt format as provided by the Researcher kit.</li> <li>emotion_annotation.csv: Contains the emotion annotation that was produced by a psychologist based on the subjects self-reports and by examining the video of the data collection process (the videos are not included for privacy reasons)</li> <li>video_offsets.csv: Additional information for synchronizing the video-based annotations with the signals.</li> </ol> For more information, refer to the accompanying publication. Disclaimer: All subjects have provided written consent for the publication of their anonymized physiological signals.

opencc-by-4.0Aug 2023View details →

A more accurate risk biomarkers recognition for Progressive Multifocal Leukoencephalopathy (PML) caused by Polyomavirus JC in patients with multiple sclerosis during treatment with disease-modifying therapies (DMTs): an ongoing clinical challenge.

The therapeutic scenario for the treatment of MS has recently been characterized by a veritable revolution, which has included the introduction, for the first time, of drug treatment guidelines and the entry, in the therapeutic landscape of the last decade, of numerous new drugs that, in varying but significantly relevant ways, have proven capable of modifying the course of the disease (Disease-Modifying Therapies - DMTs).  While the arsenal of available drugs has resulted in advances in efficacy and selectivity, it has also exposed them to the danger of side effects, potentially serious and in some cases even fatal. Among the most important side effects, complications of infectious origin, characterized by cases of viral infection/reactivation, as in the case of JCPyV, the etiologic agent of PML, are the most represented. This study, based on the follow-up of MS patients treated with different DMTs, contributes to implementing the data in the literature regarding a greater understanding of the risks related to the administration of these drugs and more appropriate monitoring of MS treatment. The risk of JCPyV reactivation with Fingolimod and Dimethyl fumarate is lower than the risk of viral reactivation associated with natalizumab. In light of the data obtained, it is possible to conclude that, in the case of natalizumab, testing for JC viruria would seem to be more useful in identifying those patients with a JCPyV-specific humoral response that is not yet detectable. In addition, our results, draw attention to the importance of analyzing the NCCR rearrangements of JCPyV. Indeed, the particular rearrangements found in plasma and PBMCs of patients with RRMS treated with natalizumab could represent an alert of neuroinvasiveness in order to detect early those patients with a higher risk of developing PML. In the case of dimethyl fumarate, it seems likely that monitoring of lymphopenia would identify a higher risk group of patients in whom alternative therapy should be sought. Prolonged lymphopenia, with absolute lymphocyte counts less than 750 lymphocytes/mL, might be the major risk factor for PML although, a greater risk might lie in the loss of CD8+ cells that are crucial for JCPyV control. For fingolimod, this strategy cannot be applied because the number of circulating lymphocytes decreases while the actual lymphocyte function appears largely normal therefore, monitoring viruria and viremia along with identification of the neurotrophic variant of the virus seems a more exploitable means for risk stratification. In conclusion, the results of this study can be considered directly transferable to the National Health System in that, both the monitoring of JCPyV reactivation by viruria and the sequence analysis of viral NCCR and host immune set-up could play the role of translatable biomarkers in clinical practice in order to assess the risk of PML onset. In addition to improving risk stratification of this disease, these biomarkers of viral reactivation and pathogenicity could facilitate timely diagnosis, optimizing the use of health care resources and contributing to the reduction of direct and indirect costs of MS disease. Prezioso Carla was supported by the Italian Ministry of Health (Starting Grant: SG-2018-12366194).

opencc-by-4.0Aug 2023View details →

TongueTap: Multimodal Tongue Gesture Recognition with Head-Worn Devices

Please cite the primary paper at <a href="http://doi.org/10.1145/3577190.3614120">https://doi.org/10.1145/3577190.3614120</a> when referencing this dataset. This dataset contains multimodal tongue gesture data as a supplement for "TongueTap: Multimodal Tongue Gesture Recognition with Head-Worn Devices" published in ICMI (International Conference on Multimodal Interfaces) 2023. The data is presented in three formats (XDF, pickle and NumPy) at various stages of pre-processing. Please review the READMEs in each file before working with them. Please also review the paper at <a href="http://doi.org/10.1145/3577190.3614120">https://doi.org/10.1145/3577190.3614120</a> for more information about the data and how it was collected. Abstract Mouth-based interfaces are a promising new approach enabling silent, hands-free and eyes-free interaction with wearable devices. However, interfaces sensing mouth movements are traditionally custom-designed and placed near or within the mouth. TongueTap synchronizes multimodal EEG, PPG, IMU, eye tracking and head tracking data from two commercial headsets to facilitate tongue gesture recognition using only off-the-shelf devices on the upper face. We classified eight closed-mouth tongue gestures with 94% accuracy, offering an invisible and inaudible method for discreet control of head-worn devices. Moreover, we found that the IMU alone differentiates eight gestures with 80% accuracy and a subset of four gestures with 92% accuracy. We built a dataset of 48,000 gesture trials across 16 participants, allowing TongueTap to perform user-independent classification. Our findings suggest tongue gestures can be a viable interaction technique for VR/AR headsets and earables without requiring novel hardware.

opencc-by-nc-4.0Aug 2023View details →

Logistic Activity Recognition Challenge (LARa Version 03) – A Motion Capture and Inertial Measurement Dataset

LARa Version 03 is a freely accessible logistics-dataset for human activity recognition. In the “Innovationlab Hybrid Services in Logistics” at TU Dortmund University, two picking and one packing scenarios with 16 subjects were recorded using an optical marker-based Motion Capturing system (OMoCap), Inertial Measurement Units (IMUs), and an RGB camera. Each subject was recorded for one hour (960 minutes in total). All the given data have been labelled and categorised into eight activity classes and 19 binary coarse-semantic descriptions, also called attributes. In total, the dataset contains 221 unique attribute representations. The dataset was created according to the guideline of the following paper: “A Tutorial on Dataset Creation for Sensor-based Human Activity Recognition”, PerCom, 2023 DOI: <a href="http://dx.doi.org/10.1109/PerComWorkshops56833.2023.10150401">10.1109/PerComWorkshops56833.2023.10150401</a> The LARa Version 03 contains a new Annotation tool for OMoCap and RGB Videos, namely, the Sequence Attribute Retrieval Annotator (SARA). SARA, developed and modified based on the LARa Version 02 annotation tool, includes desirable features and attempts to overcome limitations as found in the LARa annotation tool. Furthermore, few features were included based on the explorative study of previously developed annotation tools, see journal. In alignment with the LARa annotation tool, SARA focuses on OMoCap and video annotations. However, it is to be noted that SARA was not intended to be a video annotation tool with features such as subject tracking and multiple subject annotations. Here, the video is considered to be a supporting input to the OMoCap annotation. We would recommend other tools for pure video-based multiple-human activity annotation, including subject tracking, segmentation, and pose estimation. There are different ways of installing the annotation tool: Compiled binaries (executable files) for Windows and Mac can be directly downloaded from here. Python users can install the tool from https://pypi.org/project/annotation-tool/ (PyPi): “pip install annotation-tool”. For more information, please refer to the “Annotation Tool - Installation and User Manual”. Upgrade: <ul> <li>Annotation tool (SARA) added (for Windows and MacOS, including an installation and user manual)</li> <li>Neural Networks updated (can be used with the annotation tool)</li> <li>OMoCap data: <ul> <li>Annotation errors corrected</li> <li>Annotations reformatted, fitting the SARA annotation tool</li> <li>“additional annotated data” extended</li> <li>“Markers_Exports” added</li> </ul> </li> <li>IMU data (MbientLab and MotionMiners Sensors) <ul> <li>Annotation errors corrected</li> </ul> </li> <li>README file (protocol) updated and extended</li> </ul>   If you use this dataset for research, please cite the following paper: “LARa: Creating a Dataset for Human Activity Recognition in Logistics Using Semantic Attributes”, Sensors 2020, DOI: <a href="https://doi.org/10.3390/s20154083">10.3390/s20154083</a>. If you use the Mbientlab Networks, please cite the following paper: “From Human Pose to On-Body Devices for Human-Activity Recognition”, 25th International Conference on Pattern Recognition (ICPR), 2021, DOI: <a href="https://doi.org/10.1109/ICPR48806.2021.9412283">10.1109/ICPR48806.2021.9412283</a>. For any questions about the dataset, please contact Friedrich Niemann at friedrich.niemann@tu-dortmund.de.

PPAST-GR Dataset: Greek Parliament Proceedings post WWII Analysis and Recognition

The first post-WWII years in Greece were devastating. After a brutal Nazi occupation, the Greek Civil War (1946–1949) erupted. It wrecked the economy and the country’s infrastructure and altered politics and the social fabric for decades to come. For the tense and unstable first years of the conflict (1946–1947), a study of the issues discussed in the Greek parliament could facilitate our understanding of the society at the time. An obstacle is that parliament proceedings are publicly available in a machine-readable form from 1989; before that only scanned images of the original records exist. We show that text recognition followed by natural language processing can unlock this corpus for historical research. Using Transkribus, we trained a text recogniser (1.5% CER) that we applied to 3,156 images from 1946 and 1947. As low-quality recognition is inevitable, we trained a language model on the transcribed text and applied it to recognised text, discarding records with high average cross-entropy. Using information extraction techniques, we sampled speeches that were applauded and we introduce the first quantification of issues that were thus received. All our resources will be made public. Our model is publicly available through the Transkribus platform (http://www.transkribus.org/) under the name "Greek Parliamentary Proceedings 1946".

WiFi CSI-Based Long-Range Through-Wall Human Activity Recognition with the ESP32

WiFi CSI-Based Long-Range Through-Wall Human Activity Recognition with the ESP32 This repository contains the WiFi CSI human presence detection and activity recognition datasets proposed in [1]. Datasets <ul> <li>DP_LOS - Line-of-sight (LOS) presence detection dataset, comprised of 392 CSI amplitude spectrograms.</li> <li>DP_NLOS - Non-line-of-sight (NLOS) presence detection dataset, comprised of 384 CSI amplitude spectrograms.</li> <li>DA_LOS - LOS activity recognition dataset, comprised of 392 CSI amplitude spectrograms.</li> <li>DA_NLOS - NLOS activity recognition dataset, comprised of 384 CSI amplitude spectrograms.</li> </ul> Table 1: Characteristics of presence detection and activity recognition datasets.  <table> <tbody> <tr> <td>Dataset</td> <td>Scenario</td> <td>#Rooms</td> <td>#Persons</td> <td>#Classes</td> <td>Packet Sending Rate</td> <td>Interval </td> <td>#Spectrograms</td> </tr> <tr> <td>DP_LOS</td> <td>LOS</td> <td>1</td> <td>1</td> <td>6</td> <td>100Hz</td> <td>4s (400 packets)</td> <td>392</td> </tr> <tr> <td>DP_NLOS</td> <td>NLOS</td> <td>5</td> <td>1</td> <td>6</td> <td>100Hz</td> <td>4s (400 packets)</td> <td>384</td> </tr> <tr> <td>DA_LOS</td> <td>LOS</td> <td>1</td> <td>1</td> <td>3</td> <td>100Hz</td> <td>4s (400 packets)</td> <td>392</td> </tr> <tr> <td>DA_NLOS</td> <td>NLOS</td> <td>5</td> <td>1</td> <td>3</td> <td>100Hz</td> <td>4s (400 packets)</td> <td>384</td> </tr> </tbody> </table>   Data Format Each dataset employs an 8:1:1 training-validation-test split, defined in the provided label files trainLabels.csv, validationLabels.csv, and testLabels.csv. Label files use the sample format [i c], with i corresponding to the spectrogram index (i.png) and c corresponding to the class. For presence detection datasets (DP_LOS , DP_NLOS), c in {0 = "no presence", 1 = "presence in room 1", ..., 5 = "presence in room 5"}. For activity recognition datasets (DA_LOS , DA_NLOS), c in {0="no activity", 1="walking", and 2="walking + arm-waving"}. Furthermore, the mean and standard deviation of a given dataset are provided in meanStd.csv. Download and Use This data may be used for non-commercial research purposes only. If you publish material based on this data, we request that you include a reference to our paper [1]. [1] Strohmayer, Julian, and Martin Kampel. "WiFi CSI-Based Long-Range Through-Wall Human Activity Recognition with the ESP32" International Conference on Computer Vision Systems. Cham: Springer Nature Switzerland, 2023.  BibTeX citation: <pre>@inproceedings{strohmayer2023wifi, title={WiFi CSI-Based Long-Range Through-Wall Human Activity Recognition with the ESP32}, author={Strohmayer, Julian and Kampel, Martin}, booktitle={International Conference on Computer Vision Systems}, pages={41--50}, year={2023}, organization={Springer} }</pre>

Domain-adaptive Data Synthesis for Large-scale Supermarket Product Recognition

Domain-Adaptive Data Synthesis for Large-Scale Supermarket Product Recognition This repository contains the data synthesis pipeline and synthetic product recognition datasets proposed in [1]. Data Synthesis Pipeline: We provide the Blender 3.1 project files and Python source code of our data synthesis pipeline pipeline.zip, accompanied by the <a href="https://github.com/taesungp/contrastive-unpaired-translation">FastCUT</a> models used for synthetic-to-real domain translation models.zip. For the synthesis of new shelf images, a product assortment list and product images must be provided in the corresponding directories products/assortment/ and products/img/. The pipeline expects product images to follow the naming convention c.png, with c corresponding to a GTIN or generic class label (e.g., 9120050882171.png). The assortment list, assortment.csv, is expected to use the sample format [c, w, d, h], with c being the class label and w, d, and h being the packaging dimensions of the given product in mm (e.g., [4004218143128, 140, 70, 160]). The assortment list to use and the number of images to generate can be specified in generateImages.py (see comments). The rendering process is initiated by either executing load.py from within Blender or within a command-line terminal as a background process.  Datasets: <ul> <li>SG3k - Synthetic GroZi-3.2k (SG3k) dataset, consisting of 10,000 synthetic shelf images with 851,801 instances of 3,234 GroZi-3.2k products. Instance-level bounding boxes and generic class labels are provided for all product instances.</li> <li>SG3kt - Domain-translated version of SGI3k, utilizing GroZi-3.2k as the target domain. Instance-level bounding boxes and generic class labels are provided for all product instances.</li> <li>SGI3k - Synthetic GroZi-3.2k (SG3k) dataset, consisting of 10,000 synthetic shelf images with 838,696 instances of 1,063 GroZi-3.2k products. Instance-level bounding boxes and generic class labels are provided for all product instances.</li> <li>SGI3kt - Domain-translated version of SGI3k, utilizing GroZi-3.2k as the target domain. Instance-level bounding boxes and generic class labels are provided for all product instances.</li> <li>SPS8k - Synthetic Product Shelves 8k (SPS8k) dataset, comprised of 16,224 synthetic shelf images with 1,981,967 instances of 8,112 supermarket products. Instance-level bounding boxes and GTIN class labels are provided for all product instances.</li> <li>SPS8kt - Domain-translated version of SPS8k, utilizing SKU110k as the target domain. Instance-level bounding boxes and GTIN class labels for all product instances.</li> </ul> Table 1: Dataset characteristics.  <table> <tbody> <tr> <td>Dataset</td> <td>#images</td> <td>#products</td> <td>#instances</td> <td>  labels     </td> <td>translation</td> </tr> <tr> <td>SG3k</td> <td>10,000</td> <td>3,234</td> <td>851,801</td> <td>bounding box & generic class¹</td> <td>none</td> </tr> <tr> <td>SG3kt</td> <td>10,000</td> <td>3,234</td> <td>851,801</td> <td>bounding box & generic class¹</td> <td>GroZi-3.2k</td> </tr> <tr> <td>SGI3k</td> <td>10,000</td> <td>1,063</td> <td>838,696</td> <td>bounding box & generic class²</td> <td>none</td> </tr> <tr> <td>SGI3kt</td> <td>10,000</td> <td>1,063</td> <td>838,696</td> <td>bounding box & generic class²</td> <td>GroZi-3.2k</td> </tr> <tr> <td>SPS8k</td> <td>16,224</td> <td>8,112</td> <td>1,981,967</td> <td>bounding box & GTIN</td> <td>none</td> </tr> <tr> <td>SPS8kt</td> <td>16,224</td> <td>8,112</td> <td>1,981,967</td> <td>bounding box & GTIN</td> <td>SKU110k</td> </tr> </tbody> </table>   Sample Format A sample consists of an RGB image (i.png) and an accompanying label file (i.txt), which contains the labels for all product instances present in the image. Labels use the YOLO format [c, x, y, w, h]. ¹SG3k and SG3kt use generic pseudo-GTIN class labels, created by combining the GroZi-3.2k food product category number i (1-27) with the product image index j (j.jpg), following the convention i0000j (e.g., 13000097). ²SGI3k and SGI3kt use the generic GroZi-3.2k class labels from <a href="https://arxiv.org/abs/2003.06800">https://arxiv.org/abs/2003.06800</a>. Download and Use This data may be used for non-commercial research purposes only. If you publish material based on this data, we request that you include a reference to our paper [1]. [1] Strohmayer, Julian, and Martin Kampel. "Domain-Adaptive Data Synthesis for Large-Scale Supermarket Product Recognition." International Conference on Computer Analysis of Images and Patterns. Cham: Springer Nature Switzerland, 2023. BibTeX citation: <pre>@inproceedings{strohmayer2023domain, title={Domain-Adaptive Data Synthesis for Large-Scale Supermarket Product Recognition}, author={Strohmayer, Julian and Kampel, Martin}, booktitle={International Conference on Computer Analysis of Images and Patterns}, pages={239--250}, year={2023}, organization={Springer} }</pre>

Data from: Automatic patient-level recognition of four Plasmodium species on thin blood smear by a Real Time Detector Transformer (RT-DETR) object detection algorithm: a proof-of-concept and evaluation

Automatic patient-level recognition of four Plasmodium species on thin blood smear by a Real Time Dectector Transformer (RT-DETR) object detection algorithm: a proof-of-concept and evaluation Emilie Guemas, Baptiste Routier, Théo Ghelfenstein-Ferreira, Camille Cordier, Sophie Hartuis, Bénédicte Marion, Sébastien Bertout, Emmanuelle Varlet-Marie, Damien Costa, Grégoire Pasquier Abstract: Malaria remains a global health problem with 247 million cases and 619,000 deaths in 2021. Diagnostic of Plasmodium species is important for administering the appropriate treatment. The gold-standard diagnosis from accurate species identification remains the thin blood smear. Nevertheless, this method is time-consuming and requires highly skilled and trained microscopists. To overcome these issues, new diagnostic tools based on deep learning are emerging. This study aimed to evaluate the performances of a RT-DETR (Real-Time Detection Transformer)object detection algorithm to discriminate Plasmodium species on thin blood smears images. The algorithm was trained and validated on a dataset consisting in 24,720 images from 475 thin blood smears corresponding to 2,002,597 labels. Performances were calculated with a test dataset of 4,508 images from 170 smears corresponding to 358,825labels coming from six French university hospital. At the patient level, the RT-DETR algorithm exhibited an overall accuracy of 79.4% (135/170) with a recall of 74% (40/54) and 81.9% (95/116) for negative and positive smears, respectively. Among Plasmodium positive smears, the global sensitivity was 82.7% (91/110) with a sensitivity of 90% (38/42), 81.8% (18/22) and 76.1% (35/46) for P. falciparum, P. malariae and P. ovale/vivax, respectively. The YOLOv5 model achieved a World Health Organization (WHO) competence level 2 for species identification. Besides, the RT-DETR algorithm may be run in real-time on low-cost devices such as a smartphone and could be suitable for deployment in low-resource setting areas where microscopy experts are lacking. Data collection: The training and validation dataset included 24,720 pictures taken from 475 manually May Grunwald-Giemsa (MGG)-stained thin blood smears from the Montpellier University Hospital collection and for a smaller part from the Toulouse University Hospital collection. In Montpellier, the pictures were taken with a Flexcam C1 microscope camera (Leica) attached to a Leica DM 2000 microscope and Leica DF450C microscope camera adapted with a Leica DM2500 microscope at X1000 magnification. Labelling of pictures was performed manually, and then automatically with manual correction with a Computer Visual Annotation Tools (CVAT) free software. Nine categories of labels were used: white blood cells (n=3,338), red blood cells (n=1,887,781), platelets (n=48,520), Trypanosoma brucei (n=2,773), and red blood cells infected by P. falciparum (n=43,545), P. ovale (n=4,651), P. vivax (n=4,115), P. malariae (n=2,849) and Babesia divergens (n=5,142). The test dataset included 4,508 pictures taken from 170 thin blood smears from the same number of patients from the Parasitology laboratories of University Hospitals of Montpellier, Toulouse, Rouen, Lille, Nantes and Saint-Louis in Paris (Table 1). Among these 170 patients, 54 were not infected, including two patients with Howell-Jolly bodies, and 116 were infected with hematozoa. For each patient, between 20 and 30 photos were taken from one thin blood smear with at least one hematozoan parasite per picture for infected patients. Accurate species diagnostic was made by a senior parasitologist, and for recent smears, it was confirmed by specific PCR, either performed locally (Toulouse) or at the Malaria French National Reference Center (Montpellier, Saint Louis, Rouen, Lille, Nantes).