Commencer

Présentation des jeux de données d’estimation de pose

Formats de jeux de données pris en charge#

Format Ultralytics YOLO#

Le format des étiquettes du jeu de données utilisé pour entraîner les modèles YOLO de pose est le suivant :

  1. Un fichier texte par image : chaque image du jeu de données possède un fichier texte correspondant, portant le même nom que le fichier image et l’extension « .txt ».
  2. Une ligne par objet : chaque ligne du fichier texte correspond à une instance d’objet dans l’image.
  3. Informations sur l’objet par ligne : chaque ligne contient les informations suivantes sur l’instance d’objet :
    • Indice de classe de l’objet : un entier représentant la classe de l’objet (par exemple, 0 pour une personne, 1 pour une voiture, etc.).
    • Coordonnées du centre de l’objet : coordonnées x et y du centre de l’objet, normalisées entre 0 et 1.
    • Largeur et hauteur de l’objet : largeur et hauteur de l’objet, normalisées entre 0 et 1.
    • Coordonnées des points clés de l’objet : les points clés de l’objet, normalisés entre 0 et 1.

Voici un exemple de format d’annotation pour une tâche d’estimation de pose :

Format avec points clés 2D

<class-index> <x> <y> <width> <height> <px1> <py1> <px2> <py2> ... <pxn> <pyn>

Format avec visibilité des points clés (visibilité de chaque point incluse)

<class-index> <x> <y> <width> <height> <px1> <py1> <p1-visibility> <px2> <py2> <p2-visibility> ... <pxn> <pyn> <pn-visibility>

Dans ce format, <class-index> correspond à l’indice de classe de l’objet, <x> <y> <width> <height> aux coordonnées normalisées de la boîte englobante et <px1> <py1> <px2> <py2> ... <pxn> <pyn> aux coordonnées normalisées des points clés. Le canal de visibilité est facultatif, mais utile pour les jeux de données qui annotent les occultations.

Format YAML du jeu de données#

Le framework Ultralytics utilise le format de fichier YAML pour définir le jeu de données et la configuration du modèle lors de l’entraînement de modèles d’estimation de pose. Voici un exemple du format YAML utilisé pour définir un jeu de données de pose :

ultralytics/cfg/datasets/coco8-pose.yaml
# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license

# COCO8-pose dataset (first 8 images from COCO train2017) by Ultralytics
# Documentation: https://docs.ultralytics.com/datasets/pose/coco8-pose
# Example usage: yolo train data=coco8-pose.yaml
# parent
# ├── ultralytics
# └── datasets
#     └── coco8-pose ← downloads here (1 MB)

# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
path: coco8-pose # dataset root dir
train: images/train # train images (relative to 'path') 4 images
val: images/val # val images (relative to 'path') 4 images
test: # test images (optional)

# Keypoints
kpt_shape: [17, 3] # number of keypoints, number of dims (2 for x,y or 3 for x,y,visible)
flip_idx: [0, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 12, 11, 14, 13, 16, 15]

# Classes
names:
  0: person

# Keypoint names per class
kpt_names:
  0:
    - nose
    - left_eye
    - right_eye
    - left_ear
    - right_ear
    - left_shoulder
    - right_shoulder
    - left_elbow
    - right_elbow
    - left_wrist
    - right_wrist
    - left_hip
    - right_hip
    - left_knee
    - right_knee
    - left_ankle
    - right_ankle

# Download script/URL (optional)
download: https://github.com/ultralytics/assets/releases/download/v0.0.0/coco8-pose.zip

Les champs train, val et test désignent les images d’entraînement, de validation et de test. Chacun accepte un répertoire, un fichier image, une liste de répertoires ou de fichiers image, ou un fichier UTF-8 *.txt répertoriant un chemin d’image par ligne (les chemins commençant par ./ sont résolus par rapport au fichier *.txt). Un fichier *.txt est utile pour entraîner un modèle sur un sous-ensemble d’un répertoire, ignorer les images non annotées ou combiner des images provenant de plusieurs sources dans une même partition.

Chemins des images dans un fichier `*.txt`
path: datasets/coco8-pose # dataset root
train: train.txt # a directory, an image, a list e.g. [images/a, images/b], or a *.txt file
val: val.txt
kpt_shape: [17, 3] # required for pose datasets
names:
  0: person

names est un dictionnaire de noms de classes. L’ordre des noms doit correspondre à celui des indices de classe des objets dans les fichiers du jeu de données YOLO.

(Facultatif) flip_idx associe chaque point clé à son image miroir, afin que l’augmentation par retournement horizontal conserve la cohérence entre la gauche et la droite pour les squelettes symétriques, comme le corps ou le visage humains. Pour cinq repères faciaux indexés ainsi : [œil gauche, œil droit, nez, bouche gauche, bouche droite] = [0, 1, 2, 3, 4], flip_idx est [1, 0, 2, 4, 3] : les paires gauche-droite 0-1 et 3-4 sont inversées, et le nez conserve son propre indice.

(Facultatif) kpt_oks_sigmas définit des sigmas OKS personnalisés pour chaque point clé, utilisés pendant l’entraînement et la validation, par exemple [0.26, 0.25, 0.25, ...]. La longueur de la liste doit correspondre au nombre de points clés N provenant de kpt_shape, et chaque valeur doit être positive. Si cette option n’est pas spécifiée, les sigmas COCO à 17 points clés sont utilisés pour kpt_shape: [17, 3], et un 1/N uniforme dans les autres cas.

Utilisation#

Exemple
from ultralytics import YOLO

# Charger un modèle
model = YOLO("yolo26n-pose.pt")  # charger un modèle préentraîné (recommandé pour l’entraînement)

# Entraîner le modèle
results = model.train(data="coco8-pose.yaml", epochs=100, imgsz=640)

Jeux de données pris en charge#

Cette section présente les jeux de données compatibles avec le format Ultralytics YOLO et utilisables pour entraîner des modèles d’estimation de pose. La plupart de ces jeux de données sont également hébergés sur Ultralytics Platform, où tu peux parcourir les images et les annotations, consulter les statistiques des jeux de données et les cloner pour l’entraînement dans le cloud.

COCO-Pose#

  • Description : COCO-Pose est un jeu de données d’estimation de pose humaine à grande échelle qui couvre les images COCO 2017 contenant des personnes annotées avec des points clés.
  • Format des annotations : Identique au format Ultralytics YOLO décrit ci-dessus, avec des points clés pour les poses humaines.
  • Nombre de classes : 1 (personne).
  • Points clés : 17 types de points clés, notamment le nez, les yeux, les oreilles, les épaules, les coudes, les poignets, les hanches, les genoux et les chevilles, chacun doté d’une dimension de visibilité.
  • Utilisation : Adapté à l’entraînement de modèles d’estimation de pose humaine.
  • Remarques supplémentaires : Le jeu de données s’appuie sur le défi COCO Keypoints 2017 : 58 945 images annotées représentant 156 165 personnes.
  • En savoir plus sur COCO-Pose

COCO8-Pose#

  • Description : Ultralytics COCO8-Pose est un petit jeu de données d’estimation de pose composé des 8 premières images du jeu COCO train 2017, dont 4 pour l’entraînement et 4 pour la validation.
  • Format des annotations : Identique au format Ultralytics YOLO décrit ci-dessus, avec des points clés pour les poses humaines.
  • Nombre de classes : 1 (personne).
  • Points clés : 17 types de points clés, notamment le nez, les yeux, les oreilles, les épaules, les coudes, les poignets, les hanches, les genoux et les chevilles, chacun doté d’une dimension de visibilité.
  • Utilisation : Adapté aux tests et au débogage de modèles d’estimation de pose, ou à l’expérimentation de nouvelles approches de détection de points clés.
  • Remarques supplémentaires : COCO8-Pose est idéal pour les contrôles de cohérence et les vérifications CI.
  • En savoir plus sur COCO8-Pose

Dog-Pose#

  • Description : Le jeu de données Dog-Pose d’Ultralytics contient 6 773 images d’entraînement et 1 703 images de validation pour l’estimation des points clés canins.
  • Format des annotations : Suit le format Ultralytics YOLO, avec des annotations pour plusieurs points clés propres à l’anatomie canine.
  • Nombre de classes : 1 (chien).
  • Points clés : 24 points clés, chacun doté d’une dimension de visibilité, adaptés aux poses des chiens, notamment aux membres, aux articulations et à la position de la tête.
  • Utilisation : idéal pour entraîner des modèles à estimer les poses des chiens dans divers contextes, de la recherche aux applications réelles.
  • Remarques supplémentaires : Les images sources proviennent du Stanford Dogs Dataset.
  • En savoir plus sur Dog-Pose

Points clés de la main#

  • Description : Le jeu de données Hand Keypoints d’Ultralytics comprend 26 768 images, dont 18 776 sont destinées à l’entraînement et 7 992 à la validation.
  • Format des annotations : Identique au format Ultralytics YOLO décrit ci-dessus, mais avec 21 points clés pour une main humaine et une dimension de visibilité.
  • Nombre de classes : 1 (main).
  • Points clés : 21 points clés.
  • Utilisation : idéal pour l’estimation de la pose de la main humaine et la reconnaissance des gestes.
  • Remarques supplémentaires : Les annotations des points clés sont générées avec Google MediaPipe afin d’assurer la cohérence des annotations.
  • En savoir plus sur Hand Keypoints

Tiger-Pose#

  • Description : Le jeu de données Tiger-Pose d’Ultralytics comprend 263 images extraites d’une vidéo YouTube, dont 210 sont destinées à l’entraînement et 53 à la validation.
  • Format des annotations : Identique au format Ultralytics YOLO décrit ci-dessus, avec 12 points clés pour la pose animale et sans dimension de visibilité.
  • Nombre de classes : 1 (tigre).
  • Points clés : 12 points clés.
  • Utilisation : Idéal pour la pose animale ou toute autre pose qui n’est pas humaine.
  • Remarques supplémentaires : Publié sous la licence AGPL-3.0.
  • En savoir plus sur Tiger-Pose

Ajouter ton propre jeu de données#

Si tu disposes de ton propre jeu de données et souhaites l’utiliser pour entraîner des modèles d’estimation de pose au format Ultralytics YOLO, vérifie qu’il respecte le format spécifié ci-dessus sous « Format Ultralytics YOLO ». Convertis tes annotations au format requis et indique les chemins, le nombre de classes et les noms des classes dans le fichier de configuration YAML.

Pour éviter complètement l’étape de conversion, Ultralytics Platform te permet de téléverser des images brutes, d’annoter les points clés dans le navigateur et d’entraîner directement un modèle sur le jeu de données obtenu.

Outil de conversion#

Ultralytics fournit un outil de conversion des annotations du format populaire du jeu de données COCO vers le format YOLO :

Exemple
from ultralytics.data.converter import convert_coco

convert_coco(labels_dir="path/to/coco/annotations/", use_keypoints=True)

Cet outil de conversion permet de convertir le jeu de données COCO ou tout autre jeu de données au format COCO vers le format Ultralytics YOLO. Le paramètre use_keypoints indique s’il faut inclure les points clés (pour l’estimation de pose) dans les annotations converties.

FAQ#

  • Le format Ultralytics YOLO pour les jeux de données d’estimation de pose consiste à associer à chaque image un fichier texte correspondant. Chaque ligne du fichier texte stocke des informations sur une instance d’objet :

    • Indice de classe de l’objet
    • Coordonnées du centre de l’objet (x et y normalisés)
    • Largeur et hauteur de l’objet (normalisées)
    • Coordonnées des points clés de l’objet (pxn et pyn normalisés)

    Pour les poses 2D, les points clés comprennent des coordonnées x et y normalisées. Avec une dimension de visibilité, chaque point clé possède également un indicateur de visibilité. Pour plus de détails, consulte le format Ultralytics YOLO.

  • coco-pose.yaml est fourni avec le package et télécharge les images et les annotations lors de sa première utilisation ; aucune préparation manuelle n’est donc nécessaire :

    from ultralytics import YOLO
    
    model = YOLO("yolo26n-pose.pt")  # charger un modèle préentraîné
    results = model.train(data="coco-pose.yaml", epochs=100, imgsz=640)

    Pour en savoir plus sur le jeu de données, consulte COCO-Pose, et pour obtenir la liste complète des arguments, consulte la page Entraînement.

  • Pour ajouter ton jeu de données :

    1. Convertis tes annotations au format Ultralytics YOLO.

    2. Crée un fichier de configuration YAML qui spécifie les chemins du jeu de données, le nombre de classes et les noms des classes.

    3. Utilise le fichier de configuration pour entraîner ton modèle :

      from ultralytics import YOLO
      
      model = YOLO("yolo26n-pose.pt")
      results = model.train(data="your-dataset.yaml", epochs=100, imgsz=640)

      Pour suivre toutes les étapes, consulte la section Ajouter ton propre jeu de données.

  • Dans Ultralytics YOLO, le fichier YAML du jeu de données définit le jeu de données et la configuration du modèle pour l’entraînement. Il spécifie les chemins vers les images d’entraînement, de validation et de test, la forme des points clés, les noms des classes et d’autres options de configuration. Ce format structuré simplifie la gestion des jeux de données et l’entraînement des modèles. Voici un exemple de format YAML :

    # Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
    
    # COCO8-pose dataset (first 8 images from COCO train2017) by Ultralytics
    # Documentation: https://docs.ultralytics.com/datasets/pose/coco8-pose
    # Example usage: yolo train data=coco8-pose.yaml
    # parent
    # ├── ultralytics
    # └── datasets
    #     └── coco8-pose ← downloads here (1 MB)
    
    # Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
    path: coco8-pose # dataset root dir
    train: images/train # train images (relative to 'path') 4 images
    val: images/val # val images (relative to 'path') 4 images
    test: # test images (optional)
    
    # Keypoints
    kpt_shape: [17, 3] # number of keypoints, number of dims (2 for x,y or 3 for x,y,visible)
    flip_idx: [0, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 12, 11, 14, 13, 16, 15]
    
    # Classes
    names:
      0: person
    
    # Keypoint names per class
    kpt_names:
      0:
        - nose
        - left_eye
        - right_eye
        - left_ear
        - right_ear
        - left_shoulder
        - right_shoulder
        - left_elbow
        - right_elbow
        - left_wrist
        - right_wrist
        - left_hip
        - right_hip
        - left_knee
        - right_knee
        - left_ankle
        - right_ankle
    
    # Download script/URL (optional)
    download: https://github.com/ultralytics/assets/releases/download/v0.0.0/coco8-pose.zip

    Pour en savoir plus sur la création de fichiers de configuration YAML, consulte le format YAML des jeux de données.

  • Ultralytics fournit un outil de conversion qui convertit les annotations d’un jeu de données COCO au format YOLO, y compris les informations sur les points clés :

    from ultralytics.data.converter import convert_coco
    
    convert_coco(labels_dir="path/to/coco/annotations/", use_keypoints=True)

    Cet outil te permet d’intégrer des jeux de données au format COCO dans des projets YOLO.

Commentaires69

Pose une question, partage tes résultats ou aide quelqu’un d’autre.

  1. OfekDayanGitHub

    I trained pose models (n - x) on a custom data - pitch on a football playground (skeleton with 32 points). Most of the key-points are detected on the correct location, but they have wrong ids.

    What could be the reason?

  2. UltralyticsAssistantGitHubUltralytics

    👋 Hello, thank you for your interest in Ultralytics 🚀! We recommend checking out the Docs for detailed information about pose estimation datasets. Specifically, you can explore the Pose Datasets section, where you'll find guidance on supported formats like COCO-Pose, COCO8-Pose, Tiger-Pose, and instructions for adding custom datasets.

    If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug the issue.

    For custom training ❓ Questions, please include as much detail as possible, such as dataset examples, training logs, or specific challenges. Ensure you're following our Tips for Best Training Results for optimal performance.

    Join the Ultralytics community where it suits you best:

    • For real-time discussions, visit Discord 🎧.
    • Explore in-depth topics on Discourse.
    • Share insights and connect on our Subreddit.

    Upgrade

    To ensure you are using the latest features and fixes, upgrade to the latest ultralytics package with:

    pip install -U ultralytics

    Verify your environment includes all requirements and is running Python>=3.8 and PyTorch>=1.8.

    Environments

    Run YOLO in any of these up-to-date verified environments with preinstalled dependencies:

    Status

    Ultralytics CI

    If this badge is green, all Ultralytics CI tests are passing. CI tests validate the functionality of YOLO Modes and Tasks across macOS, Windows, and Ubuntu.

    This is an automated response, but an Ultralytics engineer will assist you shortly. Let us know if you need further guidance! 😊

    1. Glenn JocherUltralytics

      @Alebep Currently in the Ultralytics YOLO framework, all classes within a single dataset must use the same number of keypoints as defined by the kpt_shape parameter in your YAML file. This is a framework-level constraint where the model architecture expects a consistent keypoint structure across all detected objects.

      If you need to work with objects requiring different numbers of keypoints (like your dog with 5 and turtle with 3 example), you would need to:

      1. Train separate models for each keypoint configuration using different datasets
      2. Or use a unified keypoint count approach where you define all possible keypoints and mark unused ones as invisible for certain classes

      This is because the pose estimation head in YOLO is structured to predict a fixed number of keypoints for all objects. Each dataset's configuration defines this through the kpt_shape parameter, which applies uniformly to all classes in that dataset.

      For examples of different keypoint configurations, you can look at our existing pose datasets like Dog-Pose (24 keypoints) and Tiger-Pose (12 keypoints), each optimized for their specific subject.

  3. AlebepGitHub

    Can the number of keypoints vary by class? For example, could the 'dog' class have 5 keypoints while the 'turtle' class has only 3? Or do all classes need to have the same number of keypoints?

    1. Glenn JocherUltralytics

      Hi there! 👋 In Ultralytics YOLO pose estimation, all classes in a single dataset must share the same number of keypoints as defined by the kpt_shape parameter in the dataset YAML file. This configuration applies uniformly across all classes within that dataset. For mixed keypoint requirements across object types, you'll need to train separate models on different datasets. Check out our Pose Estimation Datasets Guide for details on configuring keypoint dimensions.

      1
  4. foolishcqxGitHub

    首先,非常感谢你们的工作!最近我在进行自定义的关键点检测时遇到了一个问题,我们的模型在训练集和验证集上的效果在训练过程中指标都表现的很好,但是在我测试效果很糟糕。是否有什么技巧可以改善这个问题? 数据标签格式: 0 0.528905 0.425485 0.501863 0.742139 0.385417 0.384615 0.406250 0.377622 0.429688 0.374126 0.450521 0.384615 0.427083 0.391608 0.403646 0.395105 0.565104 0.388112 0.585938 0.377622 0.609375 0.381119 0.630208 0.391608 0.609375 0.398601 0.585938 0.395105 0.450521 0.636364 0.486979 0.632867 0.510417 0.636364 0.531250 0.632867 0.565104 0.632867 0.531250 0.632867 0.510417 0.636364 0.486979 0.632867 配置文件的格式:

    path: /data/zhongjiafeng/ultralytics-main/posedataset # dataset root dir train: /data/zhongjiafeng/ultralytics-main/posedataset/train/images # train images (relative to 'path') 4 images val: /data/zhongjiafeng/ultralytics-main/posedataset/val/images # val images (relative to 'path') 4 images test: # test images (optional)

    Keypoints

    kpt_shape: [20, 2] # number of keypoints, number of dims (2 for x,y or 3 for x,y,visible)

    Classes

    names: 0: face

    我们是在加载yolo11s-pose.pt模型的基础上进行训练的

    1. Glenn JocherUltralytics

      @foolishcqx 感谢您的详细描述!从您的配置来看,关键点维度设置为2但缺少必要的flip_idx对称翻转索引,这会影响数据增强效果。建议参考COCO-Pose的YAML配置,为面部对称关键点(如双眼/嘴角)添加正确的flip_idx,并验证关键点坐标是否已归一化到0-1范围。具体可查看Pose数据集格式文档中的对称点配置说明。

  5. matteobonaGitHub

    Hi, I was trying to train a custom pose model on a dataset with 4 classes. The first class has 31 keypoints, while the others do not have keypoints. I created my YAML file as follows:

    train: "images\train" # train images val: "images\valid" #val images kpt_shape: [31, 3] nc: 4 names: ['pitch', 'player', 'referee', 'ball']

    And this is an example of a .txt file of annotations: 0 0.5002604166666667 0.5421296296296296 0.9994791666666667 0.8953703703703704 0.367078125 0.12500925925925926 2 0.2983958333333333 0.1687037037037037 2 0.22916145833333335 0.2129722222222222 2 0.0834375 0.3101851851851852 2 0 0 0 0 0 0 0.283984375 0.22016666666666668 2 0.14989583333333334 0.32462962962962966 2 0.463375 0.18461111111111111 2 0.39940625 0.2422314814814815 2 0.30180208333333336 0.3329259259259259 2 0.1875 0.4439814814814815 2 0.89265625 0.18564814814814815 2 0.8919270833333334 0.2985185185185185 2 0.8911354166666666 0.4561018518518518 2 0.8863802083333333 0.9238518518518518 2 0.7358750000000001 0.3487314814814815 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.8905729166666667 0.36574074074074076 1 2 0.9135416666666667 0.9009259259259259 0.025 0.1037037037037037 2 0.48151041666666666 0.38842592592592595 0.009895833333333333 0.05648148148148148 1 0.25416666666666665 0.42083333333333334 0.014583333333333334 0.05462962962962963 1 0.3807291666666667 0.24537037037037038 0.010416666666666666 0.04814814814814815 1 0.07291666666666667 0.31851851851851853 0.010416666666666666 0.04814814814814815 1 0.63046875 0.1712962962962963 0.008854166666666666 0.037037037037037035 1 0.2185416666666667 0.37972222222222224 0.021458333333333336 0.05388888888888889 3 0.24916666666666665 0.3971296296296296 0.005625000000000001 0.009074074074074075 2 0.49213541666666666 0.1286574074074074 0.009270833333333334 0.03138888888888889 1 0.46328125 0.5268518518518519 0.0234375 0.06296296296296296 1 0.72734375 0.5412037037037037 0.018229166666666668 0.075

    In the first row I have the 0 label for the pitch + x_center y_center w h + 31 x 3 keypoints. In the following lines I have an object for each row with class x_center y_center w h.

    Starting the train I obtain this warning for each image train: WARNING ⚠️ "images\train\imm1: ignoring corrupt image/label: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (12,) + inhomogeneous part.

    I noticed that the number 12 in "the detected shape" corresponds to the number of rows in the .txt file, but I don't understand where I made a mistake. Thanks for your help!

    1. Glenn JocherUltralytics

      @matteobona thanks for sharing your details. The issue arises because pose estimation in YOLO requires consistent keypoint dimensions across all instances. Since only class 0 has keypoints while others don't, the label rows become inconsistent in length. For pose tasks, all instances must follow the same format defined in kpt_shape. You'll need to either:

      1. Add placeholder keypoints (0,0,0) for non-pose classes, or
      2. Separate pose and detection tasks into different datasets

      For mixed tasks, we recommend training separate detection and pose models. The pose estimation documentation explains the required label format where every instance must have keypoints matching the kpt_shape dimensions.

  6. AdiGaikwadNewpageGitHub

    Hello ultralytics team, I want to use this pose estimation for human action detection like push, walking, falling, fighting, sitting, etc. Can you help me to achieve this? How do I need to prepare dataset for this?

    1. Glenn JocherUltralytics

      @AdiGaikwadNewpage to achieve human action detection using pose estimation, prepare your dataset in the Ultralytics YOLO format, ensuring each image has a corresponding text file with object and keypoint details. For more guidance on dataset preparation, refer to the Pose Estimation Datasets Overview.

    2. Glenn JocherUltralytics

      To detect actions like walking or sitting, you'll need to train a model to recognize pose sequences associated with each action. After training, use the model to predict keypoints and analyze their movement patterns over time to classify actions.

  7. LiJiang97GitHub

    My training set label file is

    0 0.47899 0.45495 0.20011 0.29458 0.38743 0.31863 2 0.54499 0.31326 2 0.57668 0.51331 2 0.38480 0.53105 2 0.38988 0.38655 2 0.54744 0.37768 2 0.56915 0.57890 2 0.39076 0.59874 2

    and I set kpt_shape=[8,3] in the yaml file to detect 8 key points.

    Why does it prompt

    "train: WARNING ⚠️ C:\applications_data\pycharm\yolov8\datasets\object_260_dataset\images\train\image_37.jpg: ignoring corrupt image/label: labels require 14 columns each"

    and

    " File "C:\applications\miniconda3\envs\yolov8\Lib\site-packages\ultralytics\data\dataset.py", line 161, in get_labels len_cls, len_boxes, len_segments = (sum(x) for x in zip(*lengths)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ ValueError: not enough values to unpack (expected 3, got 0)"?

    thank you

    1. Glenn JocherUltralytics(modifié)

      @LiJiang97 for kpt_shape: [8, 3] each label row needs 5 + 8×3 = 29 values (class, box x y w h, then x y visibility for each of the 8 keypoints), and your row has exactly that. The warning says "labels require 14 columns", which means the dataset YAML being loaded does not have kpt_shape: [8, 3] (14 columns corresponds to kpt_shape: [3, 3]). Check that the data argument points to the YAML you edited and that the setting is kpt_shape: [8, 3]. For more details, please refer to the pose dataset format documentation.

  8. benicio22GitHub

    Hi, I would like to know if the mAP@0.50 for the key point detection using Ultralytics YOLO v8 is the key point the similarity used by the coco dataset. Also, what does this 0.50 mean? It is not the IoU right?

    1. benicio22GitHub

      Do you guys have a function that calculates the metrics per key point, not overall? How can we do that?

    2. Glenn JocherUltralytics

      @benicio22 currently, Ultralytics YOLOv8 does not provide a built-in function to calculate metrics per key point. However, you can achieve this by extracting the keypoint predictions and ground truth from the results and then computing the metrics manually. For more details on the dataset format and keypoint handling, you can refer to our documentation at https://docs.ultralytics.com/datasets/pose/.

  9. PaulCashman73GitHub

    I've used YoloV8 successfully for instance segmentation of community cats. Now I'm trying to use it for pose estimation and getting nowhere. I have a dataset of 1130 images (output from the instance segmentation) with 20 labeled keypoints. I used the following code to train a pre-trained model:

    model = YOLO('yolov8m-pose.pt') # Use a pre-trained model project = '/content/data/results' name = '100-epochs-'

    results = model.train(data='/content/data/data.yaml', epochs=100, project=project, name=name, patience=0, batch=4, imgsz=640)

    My YAML file is pretty vanilla:

    train: /content/train/images val: /content/valid/images test: /content/test/images

    kpt_shape: [20, 3] flip_idx: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19]

    nc: 1 names: ['cat']

    roboflow: workspace: feralcatposeestimation project: feral-cat-pose-detection-6 version: 1 license: CC BY 4.0 url: https://universe.roboflow.com/feralcatposeestimation/feral-cat-pose-detection-6/dataset/1

    My results are dismal. Literally, there are no detections found on any of my test images. The precision and recall for the pose is nearly zero. I feel I must be doing something really wrong to achieve these results.

    I looked at one of the label files. Reformatted and with extra blanks to show the keypoints it is as follows:

    0 # One class 0.33046875. 0.715625 # Are these the x-center and y-center, or lower-left x and y? 0.33203125 0.365625 # Width, height 0.7729089743589743 0.2127192307692308 2 # I assume 2 is visible; it's not documented 0.36486538461538454 0.16201346153846155 2 0.6729679487179487 0.3529711538461539 2 0.3790602564102564 0.32252403846153843 2 0.512726923076923 0.4760394230769231 2 0.4950102564102563 0.5223086538461539 2 0.3149961538461538 0.7037836538461539 0 0.4352820512820513 0.7231221153846155 0 0.19383205128205125 0.7414942307692307 0 0.192075641025641 0.8033788461538462 0 0.5024294871794871 0.954883653846154 2 0.2842653846153846 0.7559990384615386 0 0.279875641025641 0.8072471153846154 0 0.3159115384615384 0.9471951923076924 2 0.36943205128205125 0.7559990384615386 0 0.367675641025641 0.8120817307692307 0 0.367675641025641 0.8662307692307692 0 0.46513461538461537 0.7618 0 0.46601282051282045 0.8140153846153847 0 0.46513461538461537 0.8817019230769232 0

    There are 8 visible keypoints (if 2 means visible), which corresponds to the image. But I don't know how to interpret the x,y coordinates in the second line. If they are x-center and y-center of the bounding box, then some of the keypoints are outside the box. If they are the lower left of the bounding box, then other keypoints are outside the bounding box.

    I;m not sure where to begin to unravel what is happening. Any advice would be appreciated.

    1. Glenn JocherUltralytics

      @PaulCashman73 it looks like you're on the right track with your dataset and YAML configuration. The x and y coordinates in the second line of your label file should indeed be the center of the bounding box, normalized between 0 and 1. If keypoints appear outside the bounding box, it might indicate an issue with the normalization or annotation process. Ensure that all keypoints are correctly normalized and fall within the bounding box coordinates. Additionally, verify that your dataset follows the Ultralytics YOLO format precisely. For further details, you can refer to Ultralytics Pose Estimation Datasets. If the issue persists, try updating to the latest version of the package to see if it resolves any underlying bugs.

    2. PaulCashman73GitHub

      Thanks! This dataset was generated by Roboflow, so I will raise an issue in their support forums to suggest that perhaps the data is not being generated properly.

    3. Glenn JocherUltralytics

      Thanks for the update. It's a good idea to check with Roboflow regarding the dataset generation. Meanwhile, ensure that the x and y coordinates in your label files are indeed the center of the bounding box, normalized between 0 and 1. If the keypoints are outside the bounding box, it might indicate an issue with the normalization process. Also, make sure your dataset follows the Ultralytics YOLO format precisely. If the problem persists, try updating to the latest version of the Ultralytics package.

  10. 1-zymGitHub

    Hello, when I trained yolov8-pose, some indicators were not displayed, only various loss were displayed, but these indicators were included in the result.csv file, such as map50, MAP50-90

    1. 1-zymGitHub

      0 0.605961 0.471792 0.0645781 0.0777917 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.527508 0.501177 0.0356094 0.300104 0.534375 0.377083 2 0.5375 0.372917 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.5375 0.464583 2 0.51875 0.43125 2 0.545312 0.514583 1 0.5375 0.5125 2 0.526563 0.579167 2 0 0 0 0.542188 0.639583 2 0 0 0 0 0.415367 0.431646 0.00692187 0.0497917 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 This is the coco key json label that I converted to yolo format

    2. 1-zymGitHub

      This part of the tutorial is a little unclear for me as a novice, can you provide more detailed yolov8-pose tutorial?

    3. Glenn JocherUltralytics(modifié)

      @1-zym hi there! Thanks for reaching out. I understand that the tutorial might be a bit overwhelming at first. Let's break it down step-by-step to make it clearer for you.

      Training YOLO26-Pose

      To train a YOLO26-Pose model, you'll need to follow these steps:

      1. Prepare Your Dataset: Ensure your dataset is in the Ultralytics YOLO format. Each image should have a corresponding .txt file with annotations. The format is:

        • <class-index> <x> <y> <width> <height> <px1> <py1> <px2> <py2> ... <pxn> <pyn> for 2D keypoints.
        • <class-index> <x> <y> <width> <height> <px1> <py1> <p1-visibility> <px2> <py2> <p2-visibility> ... <pxn> <pyn> <p2-visibility> for 3D keypoints.
      2. Create a YAML Configuration File: This file defines the dataset paths and other configurations. Here’s an example for a dataset with 17 keypoints:

        path: ../datasets/coco8-pose
        train: images/train
        val: images/val
        kpt_shape: [17, 3]  # 17 keypoints, 3 dimensions (x, y, visibility)
        flip_idx: [0, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 12, 11, 14, 13, 16, 15]
        names:
          0: person
      3. Train the Model: Use the following Python code to train your model:

        from ultralytics import YOLO
        
        # Load a pretrained model
        model = YOLO("yolo26n-pose.pt")
        
        # Train the model
        results = model.train(data="coco8-pose.yaml", epochs=100, imgsz=640)

        Or, you can use the CLI:

        yolo pose train data=coco8-pose.yaml model=yolo26n-pose.pt epochs=100 imgsz=640
      Monitoring Training Metrics

      During training, you might notice that only loss values are displayed in the console. However, metrics like mAP50 and mAP50-95 are recorded in the results.csv file. This is expected behavior. You can visualize these metrics using tools like TensorBoard or by plotting the CSV data.

      Additional Resources

      For more detailed information, you can refer to the Pose Estimation Datasets Overview. This page provides comprehensive details on dataset formats, YAML configuration, and supported datasets.

      If you encounter any issues or need further assistance, please provide a reproducible example of your problem. This helps us diagnose and resolve the issue more effectively. You can find guidance on creating a minimum reproducible example here.

      Feel free to ask if you have more questions! 😊

    4. 1-zymGitHub

      Thank you for your answer, which is 3D keypoint format? COCO isn't just 2D? In addition, I would like to ask, how can I test the coco test set and generate json files, because the coco official website only accepts json format

    5. 1-zymGitHub

      I solved the problem by reusing the coco-pose.yaml file to download the comment file, probably because some missing key boxes were written after I processed them myself.I want to know how to test with coco, I didn't see it in the tutorial

    6. Glenn JocherUltralytics(modifié)

      Hi @1-zym, glad to hear you resolved the initial issue! For testing with the COCO test set and generating JSON files, you can follow these steps:

      Testing with COCO Test Set
      1. Prepare the COCO Test Set: Ensure you have the COCO test images and the corresponding annotations in the correct format.

      2. Modify the YAML Configuration: Update your dataset YAML file to include the path to the test images. Here’s an example:

        path: ../datasets/coco
        train: images/train
        val: images/val
        test: images/test  # Add this line for test images
        kpt_shape: [17, 3]  # number of keypoints, number of dims (2 for x,y or 3 for x,y,visible)
        flip_idx: [0, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 12, 11, 14, 13, 16, 15]
        names:
          0: person
      3. Run the Validation: Use the following command to run the validation on the test set:

        yolo pose val model=path/to/best.pt data=coco-pose.yaml split=test save_json=True
      Generating JSON Files for COCO Evaluation

      To generate the COCO-format JSON predictions, set save_json=True when validating on the test split:

      from ultralytics import YOLO
      
      model = YOLO("path/to/best.pt")
      metrics = model.val(data="coco-pose.yaml", split="test", save_json=True)

      This writes a predictions.json file with image IDs, category IDs, boxes, scores and keypoints in COCO format to the validation run directory (for example runs/pose/val), ready for the evaluation server. Note that model.val() returns a metrics object, not a list of results, so it cannot be iterated.

      Additional Resources

      For more detailed information, you can refer to the Pose Estimation Datasets Overview. This page provides comprehensive details on dataset formats, YAML configuration, and supported datasets.

      If you encounter any issues or need further assistance, please provide a reproducible example of your problem. This helps us diagnose and resolve the issue more effectively. You can find guidance on creating a minimum reproducible example here.

      Feel free to ask if you have more questions! 😊

    7. 1-zymGitHub

      I have the following error according to the way you provided. In addition, can you provide the detailed configuration of pose training? For example: epoch, lr, whether pretrain, etc., the yolom-pose results of training 300 epoches by myself have a big gap with the results provided Ultralytics YOLOv8.2.54 🚀 Python-3.8.19 torch-1.12.1+cu116 CUDA:0 (NVIDIA A40, 45449MiB) YOLOv8m-pose summary (fused): 237 layers, 26447596 parameters, 0 gradients, 81.0 GFLOPs val: Scanning /opt/storage/zhaoyiming/datasets/coco-pose/labels/test2017.cache.. WARNING ⚠️ No labels found in /opt/storage/zhaoyiming/datasets/coco-pose/labels/test2017.cache, training may not work correctly. See https://docs.ultralytics.com/datasets for dataset formatting guidance. Class Images Instances Box(P R mAP50 m all 40670 0 0 0 0 0 0 0 0 0 WARNING ⚠️ no labels found in pose set, can not compute metrics without labels Speed: 0.1ms preprocess, 4.1ms inference, 0.0ms loss, 0.5ms postprocess per image Results saved to /opt/storage/zhaoyiming/ultralytics/runs/pose/val14 Traceback (most recent call last): File "/opt/storage/zhaoyiming/ultralytics/ultralytics/utils/test.py", line 13, in for result in results: TypeError: 'PoseMetrics' object is not iterable

      Process finished with exit code 1

    8. Glenn JocherUltralytics(modifié)

      Hi @1-zym, thanks for reaching out! It looks like you're encountering an issue with missing labels during validation and a TypeError in your script. Here are a few steps to help resolve these issues:

      Missing Labels Warning

      The warning indicates that no labels were found in your test set. Ensure that your dataset is correctly formatted and that the labels are in the expected location. You can refer to the dataset formatting guidance for more details.

      Configuration for Pose Training

      Here's a detailed example configuration for training a YOLO26-Pose model:

      from ultralytics import YOLO
      
      # Load a pretrained model
      model = YOLO("yolo26m-pose.pt")
      
      # Train the model
      results = model.train(
          data="coco-pose.yaml",  # Path to your dataset YAML file
          epochs=300,             # Number of epochs
          imgsz=640,              # Image size
          lr0=0.01,               # Initial learning rate
          pretrained=True         # Use pretrained weights
      )
      Generating JSON Files for COCO Evaluation

      To generate the COCO-format JSON predictions, set save_json=True when validating on the test split:

      from ultralytics import YOLO
      
      model = YOLO("path/to/best.pt")
      metrics = model.val(data="coco-pose.yaml", split="test", save_json=True)

      This writes a predictions.json file with image IDs, category IDs, boxes, scores and keypoints in COCO format to the validation run directory (for example runs/pose/val), ready for the evaluation server. Note that model.val() returns a metrics object, not a list of results, so it cannot be iterated.

      TypeError: 'PoseMetrics' object is not iterable

      model.val() returns a metrics object, not a list of results, so it cannot be iterated; use save_json=True as shown above. Ensure you are using the latest version of the Ultralytics package and that your script aligns with the expected API.

      If the issue persists, please provide a minimum reproducible example to help us diagnose the problem more effectively. You can find guidance on creating a reproducible example here.

      Feel free to ask if you have more questions! 😊

    9. 1-zymGitHub

      path: ../datasets/coco-pose train: train2017.txt val: val2017.txt test: test-dev2017.txt, 通过coco-pose.yaml,我并没有在我的文件夹成功下载 test-dev2017.txt,是官方没有提供?需要我自己生成? 可以提供一下yolov8-pose的详细训练配置?我自己训练300epoch的yolov8m-pose与你们提供的yolov8m-pose的精度,在coco-pose数据集上map50-90差了将近10个点

    10. 1-zymGitHub

      0 0.620039 0.5939 0.172415 0.14608 0.658793 0.552 2 0.664042 0.55 2 0.656168 0.548 2 0.67979 0.554 2 0 0 0 0.67979 0.57 2 0.648294 0.568 2 0.674541 0.588 2 0.606299 0.582 2 0.627297 0.562 2 0.603675 0.564 2 0.627297 0.616 2 0 0 0 0.595801 0.604 2 0 0 0 0.569554 0.644 2 0 0 0 0 0.385525 0.58557 0.14937 0.12586 0.414698 0.54 1 0.419948 0.538 2 0.412073 0.538 1 0.43832 0.54 2 0 0 0 0.435696 0.558 2 0.417323 0.554 2 0.396325 0.574 2 0 0 0 0 0 0 0 0 0 0.414698 0.608 2 0.39895 0.604 2 0.372703 0.592 2 0 0 0 0.346457 0.638 2 0 0 0 0 0.328898 0.70199 0.0313386 0.06714 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.760577 0.69423 0.0285564 0.05486 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.510709 0.69215 0.0187927 0.04682 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.929554 0.67602 0.0388451 0.01844 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 上面的是我自己通过脚本将coco 关键点标注文件转换为yolo 格式 0 0.620039 0.593900 0.172415 0.146080 0.658793 0.552000 2.000000 0.664042 0.550000 2.000000 0.656168 0.548000 2.000000 0.679790 0.554000 2.000000 0.000000 0.000000 0.000000 0.679790 0.570000 2.000000 0.648294 0.568000 2.000000 0.674541 0.588000 2.000000 0.606299 0.582000 2.000000 0.627297 0.562000 2.000000 0.603675 0.564000 2.000000 0.627297 0.616000 2.000000 0.000000 0.000000 0.000000 0.595801 0.604000 2.000000 0.000000 0.000000 0.000000 0.569554 0.644000 2.000000 0.000000 0.000000 0.000000 0 0.385525 0.585570 0.149370 0.125860 0.414698 0.540000 1.000000 0.419948 0.538000 2.000000 0.412073 0.538000 1.000000 0.438320 0.540000 2.000000 0.000000 0.000000 0.000000 0.435696 0.558000 2.000000 0.417323 0.554000 2.000000 0.396325 0.574000 2.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000 0.414698 0.608000 2.000000 0.398950 0.604000 2.000000 0.372703 0.592000 2.000000 0.000000 0.000000 0.000000 0.346457 0.638000 2.000000 0.000000 0.000000 0.000000 这是我通过coco-pose.yaml下载的,用上面的生成的会对训练结果造成影响?

    11. Glenn JocherUltralytics(modifié)

      Hi @1-zym, thank you for your detailed message. It looks like you're encountering a few issues related to dataset preparation and training configuration. Let's address them step-by-step.

      Missing test-dev2017.txt

      The test-dev2017.txt file is not provided by default. You need to generate this file yourself by listing the paths to your test images. Here’s a quick way to create it:

      find /path/to/test/images -type f -name "*.jpg" > test-dev2017.txt
      Detailed Training Configuration for YOLO26-Pose

      To achieve optimal results, it's crucial to fine-tune your training configuration. Here’s an example configuration for training YOLO26m-Pose:

      from ultralytics import YOLO
      
      # Load a pretrained model
      model = YOLO("yolo26m-pose.pt")
      
      # Train the model
      results = model.train(
          data="coco-pose.yaml",  # Path to your dataset YAML file
          epochs=300,             # Number of epochs
          imgsz=640,              # Image size
          lr0=0.01,               # Initial learning rate
          pretrained=True,        # Use pretrained weights
          batch=16,               # Batch size
          optimizer='Adam',       # Optimizer
          patience=50,            # Early stopping patience
      )
      Generating JSON Files for COCO Evaluation

      To generate the COCO-format JSON predictions, set save_json=True when validating on the test split:

      from ultralytics import YOLO
      
      model = YOLO("path/to/best.pt")
      metrics = model.val(data="coco-pose.yaml", split="test", save_json=True)

      This writes a predictions.json file with image IDs, category IDs, boxes, scores and keypoints in COCO format to the validation run directory (for example runs/pose/val), ready for the evaluation server. Note that model.val() returns a metrics object, not a list of results, so it cannot be iterated.

      Addressing the TypeError

      model.val() returns a metrics object, not a list of results, so it cannot be iterated; use save_json=True as shown above. Ensure you are using the latest version of the Ultralytics package. If the issue persists, please provide a minimum reproducible example to help us diagnose the problem more effectively. You can find guidance on creating a reproducible example here.

      Feel free to ask if you have more questions! 😊

    12. 1-zymGitHub

      现在我遇到一个问题,我用yolov8m-pose使用相同的训练参数、环境,只是GPU不同,训练300个epoch,不使用预训练,其他参数都是默认,我的精度比他的低了10个点,此外我在自己的卡上使用100个epoch训练yolov8n-pose比使用300个epoch 精度也高了7,8个点,你能为我解答这个问题?

    13. 1-zymGitHub

      task: pose mode: train model: ultralytics/cfg/models/v8/yolov8-pose.yaml data: ultralytics/cfg/datasets/coco-pose.yaml epochs: 300 time: null patience: 100 batch: 64 imgsz: 800 save: true save_period: -1 cache: true device: '1' workers: 8 project: null name: train35 exist_ok: false pretrained: false optimizer: SGD verbose: true seed: 0 deterministic: true single_cls: false rect: false cos_lr: false close_mosaic: 10 resume: false amp: true fraction: 1.0 profile: false freeze: null multi_scale: false overlap_mask: true mask_ratio: 4 dropout: 0.0 val: true split: val save_json: false save_hybrid: false conf: null iou: 0.7 max_det: 300 half: false dnn: false plots: true source: null vid_stride: 1 stream_buffer: false visualize: false augment: false agnostic_nms: false classes: null retina_masks: false embed: null show: false save_frames: false save_txt: false save_conf: false save_crop: false show_labels: true show_conf: true show_boxes: true line_width: null format: torchscript keras: false optimize: false int8: false dynamic: false simplify: false opset: null workspace: 4 nms: false lr0: 0.01 lrf: 0.01 momentum: 0.937 weight_decay: 0.0005 warmup_epochs: 3.0 warmup_momentum: 0.8 warmup_bias_lr: 0.1 box: 7.5 cls: 0.5 dfl: 1.5 pose: 12.0 kobj: 1.0 label_smoothing: 0.0 nbs: 64 hsv_h: 0.015 hsv_s: 0.7 hsv_v: 0.4 degrees: 0.0 translate: 0.1 scale: 0.5 shear: 0.0 perspective: 0.0 flipud: 0.0 fliplr: 0.5 bgr: 0.0 mosaic: 1.0 mixup: 0.0 copy_paste: 0.0 auto_augment: randaugment erasing: 0.4 crop_fraction: 1.0 cfg: null tracker: botsort.yaml save_dir: /opt/storage/zhaoyiming/ultralytics/runs/pose/train35 这是参数文件

    14. Glenn JocherUltralytics

      It sounds like you're experiencing significant variations in performance despite using similar training parameters. Differences in GPU architecture can sometimes lead to variations in training results due to factors like hardware-specific optimizations and floating-point precision. Additionally, training stability and convergence can be influenced by random initialization and data augmentation. To address this, ensure you're using the latest version of the Ultralytics package and try setting a fixed random seed for reproducibility. If the issue persists, consider experimenting with different learning rates, batch sizes, and optimizers to find the optimal configuration for your specific setup.

    15. 1-zymGitHub

      为什么yolov8n-pose训练100个epoch会比300个epoch要好呢?lr0=0.01 lrf=0.01 ,其他的超参数都一样,都是默认,我不是很明白它的学习率是怎么衰减的。

    16. Glenn JocherUltralytics

      The difference in performance between training for 100 epochs and 300 epochs could be due to overfitting. With a learning rate of 0.01 and default parameters, the model might be learning too quickly and overfitting to the training data. To address this, consider using a learning rate scheduler or early stopping. Additionally, ensure that your dataset is well-augmented and that the training process is consistent across different GPUs. For more details on learning rate schedules and training strategies, you can refer to the Ultralytics documentation.