Prepare Datasets

In this document, we will give a guide on the process of preparing datasets for the MMPose. Various aspects of dataset preparation will be discussed, including using built-in datasets, creating custom datasets, combining datasets for training, browsing and downloading the datasets.

Use built-in datasets

Step 1: Prepare Data

MMPose supports multiple tasks and corresponding datasets. You can find them in dataset zoo. To properly prepare your data, please follow the guidelines associated with your chosen dataset.

Step 2: Configure Dataset Settings in the Config File

Before training or evaluating models, you must configure the dataset settings. Take for example, which can be used to train or evaluate the HRNet pose estimator on COCO dataset. We will go through the dataset configuration.

  • Basic Dataset Arguments

    # base dataset settings
    dataset_type = 'CocoDataset'
    data_mode = 'topdown'
    data_root = 'data/coco/'
    • dataset_type specifies the class name of the dataset. Users can refer to Datasets APIs to find the class name of their desired dataset.

    • data_mode determines the output format of the dataset, with two options available: 'topdown' and 'bottomup'. If data_mode='topdown', the data element represents a single instance with its pose; otherwise, the data element is an entire image containing multiple instances and poses.

    • data_root designates the root directory of the dataset.

  • Data Processing Pipelines

    # pipelines
    train_pipeline = [
        dict(type='RandomFlip', direction='horizontal'),
        dict(type='TopdownAffine', input_size=codec['input_size']),
        dict(type='GenerateTarget', encoder=codec),
    val_pipeline = [
        dict(type='TopdownAffine', input_size=codec['input_size']),

    The train_pipeline and val_pipeline define the steps to process data elements during the training and evaluation phases, respectively. In addition to loading images and packing inputs, the train_pipeline primarily consists of data augmentation techniques and target generator, while the val_pipeline focuses on transforming data elements into a unified format.

  • Data Loaders

    # data loaders
    train_dataloader = dict(
        sampler=dict(type='DefaultSampler', shuffle=True),
    val_dataloader = dict(
        sampler=dict(type='DefaultSampler', shuffle=False, round_up=False),
    test_dataloader = val_dataloader

    This section is crucial for configuring the dataset in the config file. In addition to the basic dataset arguments and pipelines discussed earlier, other important parameters are defined here. The batch_size determines the batch size per GPU; the ann_file indicates the annotation file for the dataset; and data_prefix specifies the image folder. The bbox_file, which supplies detected bounding box information, is only used in the val/test data loader for top-down datasets.

We recommend copying the dataset configuration from provided config files that use the same dataset, rather than writing it from scratch, in order to minimize potential errors. By doing so, users can simply make the necessary modifications as needed, ensuring a more reliable and efficient setup process.

Use a custom dataset

The Customize Datasets guide provides detailed information on how to build a custom dataset. In this section, we will highlight some key tips for using and configuring custom datasets.

  • Determine the dataset class name. If you reorganize your dataset into the COCO format, you can simply use CocoDataset as the value for dataset_type. Otherwise, you will need to use the name of the custom dataset class you added.

  • Specify the meta information config file. Suppose the path of the annotation file is aaa/annotations/xxx.json, and the path of imgs is aaa/train/c.jpg, you should specify the meta information config file as follows:

    train_dataloader = dict(
            # ann file is stored at {data_root}/{ann_file}
            # e.g. aaa/annotations/xxx.json
            # img is stored at {data_root}/{img}/
            # e.g. aaa/train/c.jpg
            # specify dataset meta information

    Note that the argument metainfo must be specified in the val/test data loaders as well.

Use mixed datasets for training

MMPose offers a convenient and versatile solution for training with mixed datasets. Please refer to Use Mixed Datasets for Training.

Browse dataset

tools/analysis_tools/ helps the user to browse a pose dataset visually, or save the image to a designated directory.

python tools/misc/ ${CONFIG} [-h] [--output-dir ${OUTPUT_DIR}] [--max-item-per-dataset ${MAX_ITEM_PER_DATASET}] [--not-show] [--phase ${PHASE}] [--mode ${MODE}] [--show-interval ${SHOW_INTERVAL}]
ARGS Description
CONFIG The path to the config file.
--output-dir OUTPUT_DIR The target folder to save visualization results. If not specified, the visualization results will not be saved.
--not-show Do not show the visualization results in an external window.
--phase {train, val, test} Options for dataset.
--mode {original, transformed} Specify the type of visualized images. original means to show images without pre-processing; transformed means to show images are pre-processed.
--show-interval SHOW_INTERVAL Time interval between visualizing two images.
--max-item-per-dataset Define the maximum item processed per dataset, default to 50

For instance, users who want to visualize images and annotations in COCO dataset use:

python tools/misc/ configs/body_2d_keypoint/topdown_heatmap/coco/ --mode original

The bounding boxes and keypoints will be plotted on the original image. Following is an example: original_coco

The original images need to be processed before being fed into models. To visualize pre-processed images and annotations, users need to modify the argument mode to transformed. For example:

python tools/misc/ configs/body_2d_keypoint/topdown_heatmap/coco/ --mode transformed

Here is a processed sample


The heatmap target will be visualized together if it is generated in the pipeline.

Download dataset via MIM

By using OpenXLab, you can obtain free formatted datasets in various fields. Through the search function of the platform, you may address the dataset they look for quickly and easily. Using the formatted datasets from the platform, you can efficiently conduct tasks across datasets.

If you use MIM to download, make sure that the version is greater than v0.3.8. You can use the following command to update, install, login and download the dataset:

# upgrade your MIM
pip install -U openmim

# install OpenXLab CLI tools
pip install -U openxlab
# log in OpenXLab
openxlab login

# download coco2017 and preprocess by MIM
mim download mmpose --dataset coco2017

Supported datasets

Here is the list of supported datasets, we will continue to update it in the future.


Dataset name Download command
COCO 2017 mim download mmpose --dataset coco2017
MPII mim download mmpose --dataset mpii
AI Challenger mim download mmpose --dataset aic
CrowdPose mim download mmpose --dataset crowdpose


Dataset name Download command
LaPa mim download mmpose --dataset lapa
300W mim download mmpose --dataset 300w
WFLW mim download mmpose --dataset wflw


Dataset name Download command
OneHand10K mim download mmpose --dataset onehand10k
FreiHand mim download mmpose --dataset freihand
HaGRID mim download mmpose --dataset hagrid

Whole Body

Dataset name Download command
Halpe mim download mmpose --dataset halpe


Dataset name Download command
AP-10K mim download mmpose --dataset ap10k


Coming Soon

Read the Docs v: latest
On Read the Docs
Project Home

Free document hosting provided by Read the Docs.