Learn about Codecs¶

In the keypoint detection task, depending on the algorithm, it is often necessary to generate targets in different formats, such as normalized coordinates, vectors and heatmaps, etc. Similarly, for the model outputs, a decoding process is required to transform them into coordinates.

Encoding and decoding are closely related and inverse each other. In earlier versions of MMPose, encoding and decoding are implemented at different modules, making it less intuitive and unified.

MMPose 1.0 introduced a new module Codec to integrate the encoding and decoding together in a modular and user-friendly form.

Here is a diagram to show where the Codec is:

pose_estimator_en

Basic Concepts¶

A typical codec consists of two parts:

Encoder
Decoder

Encoder¶

The encoder transforms the coordinates in the input image space into the needed target format:

Normalized Coordinates
One-dimensional Vectors
Gaussian Heatmaps

For example, in the Regression-based method, the encoder will be:

def encode(self,
           keypoints: np.ndarray,
           keypoints_visible: Optional[np.ndarray] = None) -> dict:
    """Encoding keypoints from input image space to normalized space.

    Args:
        keypoints (np.ndarray): Keypoint coordinates in shape (N, K, D)
        keypoints_visible (np.ndarray): Keypoint visibilities in shape
            (N, K)

    Returns:
        dict:
        - keypoint_labels (np.ndarray): The normalized regression labels in
            shape (N, K, D) where D is 2 for 2d coordinates
        - keypoint_weights (np.ndarray): The target weights in shape
            (N, K)
    """
    if keypoints_visible is None:
        keypoints_visible = np.ones(keypoints.shape[:2], dtype=np.float32)

    w, h = self.input_size
    valid = ((keypoints >= 0) &
             (keypoints <= [w - 1, h - 1])).all(axis=-1) & (
                 keypoints_visible > 0.5)

    keypoint_labels = (keypoints / np.array([w, h])).astype(np.float32)
    keypoint_weights = np.where(valid, 1., 0.).astype(np.float32)

    encoded = dict(
        keypoint_labels=keypoint_labels, keypoint_weights=keypoint_weights)

    return encoded

The encoded data is converted to Tensor format in PackPoseInputs and packed in data_sample.gt_instance_labels for model calls. By default it will consist of the following encoded fields:

keypoint_labels
keypoint_weights
keypoints_visible_weights

To specify data fields to be packed, you can define the label_mapping_table attribute in the codec. For example, in VideoPoseLifting:

label_mapping_table = dict(
        trajectory_weights='trajectory_weights',
        lifting_target_label='lifting_target_label',
        lifting_target_weight='lifting_target_weight',
)

data_sample.gt_instance_labels are generally used for loss calculation, as demonstrated by loss() in RegressionHead.

def loss(self,
         inputs: Tuple[Tensor],
         batch_data_samples: OptSampleList,
         train_cfg: ConfigType = {}) -> dict:
    """Calculate losses from a batch of inputs and data samples."""

    pred_outputs = self.forward(inputs)

    keypoint_labels = torch.cat(
        [d.gt_instance_labels.keypoint_labels for d in batch_data_samples])
    keypoint_weights = torch.cat([
        d.gt_instance_labels.keypoint_weights for d in batch_data_samples
    ])

    # calculate losses
    losses = dict()
    loss = self.loss_module(pred_outputs, keypoint_labels,
                            keypoint_weights.unsqueeze(-1))

    losses.update(loss_kpt=loss)
    ### Omitted ###

Note

Encoder also defines data to be packed in data_sample.gt_instances and data_sample.gt_fields. Modify instance_mapping_table and field_mapping_table in the codec will specify values to be packed respectively. For default values, please check BaseKeypointCodec.

Decoder¶

The decoder transforms the model outputs into coordinates in the input image space, which is the opposite processing of the encoder.

For example, in the Regression-based method, the decoder will be:

def decode(self, encoded: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
    """Decode keypoint coordinates from normalized space to input image
    space.

    Args:
        encoded (np.ndarray): Coordinates in shape (N, K, D)

    Returns:
        tuple:
        - keypoints (np.ndarray): Decoded coordinates in shape (N, K, D)
        - scores (np.ndarray): The keypoint scores in shape (N, K).
            It usually represents the confidence of the keypoint prediction

    """

    if encoded.shape[-1] == 2:
        N, K, _ = encoded.shape
        normalized_coords = encoded.copy()
        scores = np.ones((N, K), dtype=np.float32)
    elif encoded.shape[-1] == 4:
        # split coords and sigma if outputs contain output_sigma
        normalized_coords = encoded[..., :2].copy()
        output_sigma = encoded[..., 2:4].copy()
        scores = (1 - output_sigma).mean(axis=-1)
    else:
        raise ValueError(
            'Keypoint dimension should be 2 or 4 (with sigma), '
            f'but got {encoded.shape[-1]}')

    w, h = self.input_size
    keypoints = normalized_coords * np.array([w, h])

    return keypoints, scores

By default, the decode() method only performs decoding on a single instance. You can also implement the batch_decode() method to boost the decoding process.

Common Usage¶

The example below shows how to use a codec in your config:

Define the Codec
Generate Targets
Head

Define the Codec¶

Take the Regression-based method to generate normalized coordinates as an example, you can define a codec in your config as follows:

codec = dict(type='RegressionLabel', input_size=(192, 256))

Generate Targets¶

In pipelines, A codec should be passed into GenerateTarget to work as the encoder:

dict(type='GenerateTarget', encoder=codec)

Head¶

In MMPose workflows, we decode the model outputs in Head, which requires a codec to work as the decoder:

head=dict(
    type='RLEHead',
    in_channels=2048,
    num_joints=17,
    loss=dict(type='RLELoss', use_target_weight=True),
    decoder=codec
)

Here is the phase of a config file:

# codec settings
codec = dict(type='RegressionLabel', input_size=(192, 256))                     ## definition ##

# model settings
model = dict(
    type='TopdownPoseEstimator',
    data_preprocessor=dict(
        type='PoseDataPreprocessor',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        bgr_to_rgb=True),
    backbone=dict(
        type='ResNet',
        depth=50,
        init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50'),
    ),
    neck=dict(type='GlobalAveragePooling'),
    head=dict(
        type='RLEHead',
        in_channels=2048,
        num_joints=17,
        loss=dict(type='RLELoss', use_target_weight=True),
        decoder=codec),                                                         ## Head ##
    test_cfg=dict(
        flip_test=True,
        shift_coords=True,
    ))

# base dataset settings
dataset_type = 'CocoDataset'
data_mode = 'topdown'
data_root = 'data/coco/'

backend_args = dict(backend='local')

# pipelines
train_pipeline = [
    dict(type='LoadImage', backend_args=backend_args),
    dict(type='GetBBoxCenterScale'),
    dict(type='RandomFlip', direction='horizontal'),
    dict(type='RandomHalfBody'),
    dict(type='RandomBBoxTransform'),
    dict(type='TopdownAffine', input_size=codec['input_size']),
    dict(type='GenerateTarget', encoder=codec),   ## Generate Target ##
    dict(type='PackPoseInputs')
]
test_pipeline = [
    dict(type='LoadImage', backend_args=backend_args),
    dict(type='GetBBoxCenterScale'),
    dict(type='TopdownAffine', input_size=codec['input_size']),
    dict(type='PackPoseInputs')
]

Supported Codecs¶

Supported codecs are in $MMPOSE/mmpose/codecs/. Here is a list:

RegressionLabel
IntegralRegressionLabel
MSRAHeatmap
UDPHeatmap
MegviiHeatmap
SPR
SimCC
DecoupledHeatmap
ImagePoseLifting
VideoPoseLifting
MotionBERTLabel

RegressionLabel¶

[Github]

The RegressionLabel codec is used to generate normalized coordinates as the regression targets.

Input

Encoding keypoints from input image space to normalized space.

Output

Decoding normalized coordinates from normalized space to input image space.

Related works:

DeepPose
RLE

IntegralRegressionLabel¶

[Github]

The IntegralRegressionLabel codec is used to generate normalized coordinates as the regression targets.

Input

Encoding keypoints from input image space to normalized space, and generate Gaussian heatmaps as well.

Output

Decoding normalized coordinates from normalized space to input image space.

Related works:

MSRAHeatmap¶

[Github]

The MSRAHeatmap codec is used to generate Gaussian heatmaps as the targets.

Input

Encoding keypoints from input image space to output space as 2D Gaussian heatmaps.

Output

Decoding 2D Gaussian heatmaps from output space to input image space as coordinates.

Related works:

UDPHeatmap¶

[Github]

The UDPHeatmap codec is used to generate Gaussian heatmaps as the targets.

Input

Encoding keypoints from input image space to output space as 2D Gaussian heatmaps.

Output

Decoding 2D Gaussian heatmaps from output space to input image space as coordinates.

Related works:

UDP

MegviiHeatmap¶

[Github]

The MegviiHeatmap codec is used to generate Gaussian heatmaps as the targets, which is usually used in Megvii’s works.

Input

Encoding keypoints from input image space to output space as 2D Gaussian heatmaps.

Output

Decoding 2D Gaussian heatmaps from output space to input image space as coordinates.

Related works:

MSPN
RSN

SPR¶

[Github]

The SPR codec is used to generate Gaussian heatmaps of instances’ center, and offsets as the targets.

Input

Encoding keypoints from input image space to output space as 2D Gaussian heatmaps and offsets.

Output

Decoding 2D Gaussian heatmaps and offsets from output space to input image space as coordinates.

Related works:

DEKR

SimCC¶

[Github]

The SimCC codec is used to generate 1D Gaussian representations as the targets.

Input

Encoding keypoints from input image space to output space as 1D Gaussian representations.

Output

Decoding 1D Gaussian representations from output space to input image space as coordinates.

Related works:

DecoupledHeatmap¶

[Github]

The DecoupledHeatmap codec is used to generate Gaussian heatmaps as the targets.

Input

Encoding human center points and keypoints from input image space to output space as 2D Gaussian heatmaps.

Output

Decoding 2D Gaussian heatmaps from output space to input image space as coordinates.

Related works:

CID

ImagePoseLifting¶

[Github]

The ImagePoseLifting codec is used for image 2D-to-3D pose lifting.

Input

Encoding 2d keypoints from input image space to normalized 3d space.

Output

Decoding 3d keypoints from normalized space to input image space.

Related works:

SimpleBaseline3D

VideoPoseLifting¶

[Github]

The VideoPoseLifting codec is used for video 2D-to-3D pose lifting.

Input

Encoding 2d keypoints from input image space to normalized 3d space.

Output

Decoding 3d keypoints from normalized space to input image space.

Related works:

VideoPose3D

MotionBERTLabel¶

[Github]

The MotionBERTLabel codec is used for video 2D-to-3D pose lifting.

Input

Encoding 2d keypoints from input image space to normalized 3d space.

Output

Decoding 3d keypoints from normalized space to input image space.

Related works:

MotionBERT