data_juicer.ops.common.hawor_func module

data_juicer.ops.common.hawor_func.default_config() → CfgNode

Get a yacs CfgNode object with the default config values.

data_juicer.ops.common.hawor_func.get_config(config_file: str, merge: bool = True, update_cachedir: bool = False) → CfgNode

Read a config file and optionally merge it with the default config file. :param config_file: Path to config file. :type config_file: str :param merge: Whether to merge with the default config or not. :type merge: bool

返回:

Config as a yacs CfgNode object.

返回类型:

CfgNode

data_juicer.ops.common.hawor_func.create_backbone(cfg)

data_juicer.ops.common.hawor_func.normalization_layer(norm: str | None, dim: int, norm_cond_dim: int = -1)

class data_juicer.ops.common.hawor_func.AdaptiveLayerNorm1D(data_dim: int, norm_cond_dim: int)

基类：Module

__init__(data_dim: int, norm_cond_dim: int)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor, t: Tensor) → Tensor

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.FrequencyEmbedder(num_frequencies, max_freq_log2)

基类：Module

__init__(num_frequencies, max_freq_log2)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

data_juicer.ops.common.hawor_func.get_transform(center, scale, res, rot=0)

Generate transformation matrix.

data_juicer.ops.common.hawor_func.transform(pt, center, scale, res, invert=0, rot=0, asint=True)

Transform pixel location to different reference.

data_juicer.ops.common.hawor_func.crop(img, center, scale, res, rot=0)

Crop image according to the supplied bounding box.

data_juicer.ops.common.hawor_func.boxes_2_cs(boxes)

data_juicer.ops.common.hawor_func.exists(val)

data_juicer.ops.common.hawor_func.default(val, d)

data_juicer.ops.common.hawor_func.perspective_projection(points, rotation, translation, focal_length, camera_center, distortion=None)

This function computes the perspective projection of a set of points. Input:

points (bs, N, 3): 3D points rotation (bs, 3, 3): Camera rotation translation (bs, 3): Camera translation focal_length (bs,) or scalar: Focal length camera_center (bs, 2): Camera center

data_juicer.ops.common.hawor_func.avg_rot(rot)

data_juicer.ops.common.hawor_func.rot9d_to_rotmat(x)

Convert 9D rotation representation to 3x3 rotation matrix. Based on Levinson et al., "An Analysis of SVD for Deep Rotation Estimation" Input:

(B,9) or (B,J*9) Batch of 9D rotation (interpreted as 3x3 est rotmat)

Output:

(B,3,3) or (B*J,3,3) Batch of corresponding rotation matrices

data_juicer.ops.common.hawor_func.quat_to_rotmat(quat)

Convert quaternion coefficients to rotation matrix. :param quat: size = [B, 4] 4 <===>(w, x, y, z)

返回:

Rotation matrix corresponding to the quaternion -- size = [B, 3, 3]

data_juicer.ops.common.hawor_func.rot6d_to_rotmat(x: Tensor) → Tensor

Convert 6D rotation representation to 3x3 rotation matrix. Based on Zhou et al., "On the Continuity of Rotation Representations in Neural Networks", CVPR 2019 :param x: (B,6) Batch of 6-D rotation representations. :type x: torch.Tensor

返回:

Batch of corresponding rotation matrices with shape (B,3,3).

返回类型:

torch.Tensor

data_juicer.ops.common.hawor_func.rotmat_to_rot6d(rotmat)

Inverse function of the above. Input:

(B,3,3) Batch of corresponding rotation matrices

Output:

(B,6) Batch of 6-D rotation representations

data_juicer.ops.common.hawor_func.rotation_matrix_to_angle_axis(rotation_matrix)

This function is borrowed from kornia/kornia

Convert 3x4 rotation matrix to Rodrigues vector

参数:

rotation_matrix (Tensor) -- rotation matrix.

返回:

Rodrigues vector transformation.

返回类型:

Tensor

Shape:

• Input: \((N, 3, 4)\)

• Output: \((N, 3)\)

示例

>>> input = torch.rand(2, 3, 4)  # Nx4x4
>>> output = tgm.rotation_matrix_to_angle_axis(input)  # Nx3

data_juicer.ops.common.hawor_func.quaternion_to_angle_axis(quaternion: Tensor) → Tensor

This function is borrowed from kornia/kornia

Convert quaternion vector to angle axis of rotation.

Adapted from ceres C++ library: ceres-solver/include/ceres/rotation.h

参数:

quaternion (torch.Tensor) -- tensor with quaternions.

返回:

tensor with angle axis of rotation.

返回类型:

torch.Tensor

Shape:

• Input: \((*, 4)\) where * means, any number of dimensions

• Output: \((*, 3)\)

示例

>>> quaternion = torch.rand(2, 4)  # Nx4
>>> angle_axis = tgm.quaternion_to_angle_axis(quaternion)  # Nx3

data_juicer.ops.common.hawor_func.rotation_matrix_to_quaternion(rotation_matrix, eps=1e-06)

This function is borrowed from kornia/kornia Convert rotation matrix to 4d quaternion vector This algorithm is based on algorithm described in KieranWynn/pyquaternion

:param rotation_matrix (N, 3, 3)

data_juicer.ops.common.hawor_func.estimate_translation_np(S, joints_2d, joints_conf, focal_length=5000.0, img_size=224.0)

This function is borrowed from nkolot/SPIN

Find camera translation that brings 3D joints S closest to 2D the corresponding joints_2d. Input:

S: (25, 3) 3D joint locations joints: (25, 3) 2D joint locations and confidence

返回:

(3,) camera translation vector

data_juicer.ops.common.hawor_func.estimate_translation(S, joints_2d, focal_length=5000.0, img_size=224.0)

Find camera translation that brings 3D joints S closest to 2D the corresponding joints_2d. Input:

S: (B, 49, 3) 3D joint locations joints: (B, 49, 3) 2D joint locations and confidence

返回:

(B, 3) camera translation vectors

data_juicer.ops.common.hawor_func.get_keypoints_rectangle(keypoints: array, threshold: float) → Tuple[float, float, float]

Compute rectangle enclosing keypoints above the threshold. :param keypoints: Keypoint array of shape (N, 3). :type keypoints: np.array :param threshold: Confidence visualization threshold. :type threshold: float

返回:

Rectangle width, height and area.

返回类型:

Tuple[float, float, float]

data_juicer.ops.common.hawor_func.render_keypoints(img: array, keypoints: array, pairs: List, colors: List, thickness_circle_ratio: float, thickness_line_ratio_wrt_circle: float, pose_scales: List, threshold: float = 0.1, alpha: float = 1.0) → array

Render keypoints on input image. :param img: Input image of shape (H, W, 3) with pixel values in the [0,255] range. :type img: np.array :param keypoints: Keypoint array of shape (N, 3). :type keypoints: np.array :param pairs: List of keypoint pairs per limb. :type pairs: List :param colors: (List): List of colors per keypoint. :param thickness_circle_ratio: Circle thickness ratio. :type thickness_circle_ratio: float :param thickness_line_ratio_wrt_circle: Line thickness ratio wrt the circle. :type thickness_line_ratio_wrt_circle: float :param pose_scales: List of pose scales. :type pose_scales: List :param threshold: Only visualize keypoints with confidence above the threshold. :type threshold: float

返回:

Image of shape (H, W, 3) with keypoints drawn on top of the original image.

返回类型:

(np.array)

data_juicer.ops.common.hawor_func.render_hand_keypoints(img, right_hand_keypoints, threshold=0.1, use_confidence=False, map_fn=<function <lambda>>, alpha=1.0)

data_juicer.ops.common.hawor_func.render_hand_landmarks(img, right_hand_keypoints, threshold=0.1, use_confidence=False, map_fn=<function <lambda>>, alpha=1.0)

data_juicer.ops.common.hawor_func.render_body_keypoints(img: array, body_keypoints: array) → array

Render OpenPose body keypoints on input image. :param img: Input image of shape (H, W, 3) with pixel values in the [0,255] range. :type img: np.array :param body_keypoints: Keypoint array of shape (N, 3); 3 <====> (x, y, confidence). :type body_keypoints: np.array

返回:

Image of shape (H, W, 3) with keypoints drawn on top of the original image.

返回类型:

(np.array)

data_juicer.ops.common.hawor_func.render_openpose(img: array, hand_keypoints: array) → array

Render keypoints in the OpenPose format on input image. :param img: Input image of shape (H, W, 3) with pixel values in the [0,255] range. :type img: np.array :param body_keypoints: Keypoint array of shape (N, 3); 3 <====> (x, y, confidence). :type body_keypoints: np.array

返回:

Image of shape (H, W, 3) with keypoints drawn on top of the original image.

返回类型:

(np.array)

data_juicer.ops.common.hawor_func.render_openpose_landmarks(img: array, hand_keypoints: array) → array

Render keypoints in the OpenPose format on input image. :param img: Input image of shape (H, W, 3) with pixel values in the [0,255] range. :type img: np.array :param body_keypoints: Keypoint array of shape (N, 3); 3 <====> (x, y, confidence). :type body_keypoints: np.array

返回:

Image of shape (H, W, 3) with keypoints drawn on top of the original image.

返回类型:

(np.array)

data_juicer.ops.common.hawor_func.get_pylogger(name='data_juicer.ops.common.hawor_func') → Logger

Initializes multi-GPU-friendly python command line logger.

class data_juicer.ops.common.hawor_func.PreNorm(dim: int, fn: Callable, norm: str = 'layer', norm_cond_dim: int = -1)

基类：Module

__init__(dim: int, fn: Callable, norm: str = 'layer', norm_cond_dim: int = -1)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor, *args, **kwargs)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.FeedForward(dim, hidden_dim, dropout=0.0)

基类：Module

__init__(dim, hidden_dim, dropout=0.0)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.Attention(dim, heads=8, dim_head=64, dropout=0.0)

基类：Module

__init__(dim, heads=8, dim_head=64, dropout=0.0)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.CrossAttention(dim, context_dim=None, heads=8, dim_head=64, dropout=0.0)

基类：Module

__init__(dim, context_dim=None, heads=8, dim_head=64, dropout=0.0)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x, context=None)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.Transformer(dim: int, depth: int, heads: int, dim_head: int, mlp_dim: int, dropout: float = 0.0, norm: str = 'layer', norm_cond_dim: int = -1)

基类：Module

__init__(dim: int, depth: int, heads: int, dim_head: int, mlp_dim: int, dropout: float = 0.0, norm: str = 'layer', norm_cond_dim: int = -1)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor, *args)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.TransformerCrossAttn(dim: int, depth: int, heads: int, dim_head: int, mlp_dim: int, dropout: float = 0.0, norm: str = 'layer', norm_cond_dim: int = -1, context_dim: int | None = None)

基类：Module

__init__(dim: int, depth: int, heads: int, dim_head: int, mlp_dim: int, dropout: float = 0.0, norm: str = 'layer', norm_cond_dim: int = -1, context_dim: int | None = None)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor, *args, context=None, context_list=None)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.DropTokenDropout(p: float = 0.1)

基类：Module

__init__(p: float = 0.1)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.ZeroTokenDropout(p: float = 0.1)

基类：Module

__init__(p: float = 0.1)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x: Tensor)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.TransformerEncoder(num_tokens: int, token_dim: int, dim: int, depth: int, heads: int, mlp_dim: int, dim_head: int = 64, dropout: float = 0.0, emb_dropout: float = 0.0, emb_dropout_type: str = 'drop', emb_dropout_loc: str = 'token', norm: str = 'layer', norm_cond_dim: int = -1, token_pe_numfreq: int = -1)

基类：Module

__init__(num_tokens: int, token_dim: int, dim: int, depth: int, heads: int, mlp_dim: int, dim_head: int = 64, dropout: float = 0.0, emb_dropout: float = 0.0, emb_dropout_type: str = 'drop', emb_dropout_loc: str = 'token', norm: str = 'layer', norm_cond_dim: int = -1, token_pe_numfreq: int = -1)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(inp: Tensor, *args, **kwargs)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.TransformerDecoder(num_tokens: int, token_dim: int, dim: int, depth: int, heads: int, mlp_dim: int, dim_head: int = 64, dropout: float = 0.0, emb_dropout: float = 0.0, emb_dropout_type: str = 'drop', norm: str = 'layer', norm_cond_dim: int = -1, context_dim: int | None = None, skip_token_embedding: bool = False)

基类：Module

__init__(num_tokens: int, token_dim: int, dim: int, depth: int, heads: int, mlp_dim: int, dim_head: int = 64, dropout: float = 0.0, emb_dropout: float = 0.0, emb_dropout_type: str = 'drop', norm: str = 'layer', norm_cond_dim: int = -1, context_dim: int | None = None, skip_token_embedding: bool = False)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(inp: Tensor, *args, context=None, context_list=None)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.MANOTransformerDecoderHead(cfg)

基类：Module

HMR2 Cross-attention based SMPL Transformer decoder

__init__(cfg)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x, **kwargs)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.temporal_attention(in_dim=1280, out_dim=1280, hdim=512, nlayer=6, nhead=4, residual=False)

基类：Module

__init__(in_dim=1280, out_dim=1280, hdim=512, nlayer=6, nhead=4, residual=False)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.PositionalEncoding(d_model, dropout=0.1, max_len=100)

基类：Module

__init__(d_model, dropout=0.1, max_len=100)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x)

Define the computation performed at every call.

Should be overridden by all subclasses.

备注

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class data_juicer.ops.common.hawor_func.TrackDatasetEval(imgfiles, boxes, crop_size=256, dilate=1.0, img_focal=None, img_center=None, normalization=True, item_idx=0, do_flip=False)

基类：Dataset

Track Dataset Class - Load images/crops of the tracked boxes.

__init__(imgfiles, boxes, crop_size=256, dilate=1.0, img_focal=None, img_center=None, normalization=True, item_idx=0, do_flip=False)

data_juicer.ops.common.hawor_func.dataset_config() → CfgNode

Get dataset config file :returns: Dataset config as a yacs CfgNode object. :rtype: CfgNode

data_juicer.ops.common.hawor_func.parse_chunks(frame, boxes, min_len=16)

If a track disappear in the middle, we separate it to different segments to estimate the HPS independently. If a segment is less than 16 frames, we get rid of it for now.

data_juicer.ops.common.hawor_func.interpolate_bboxes(bboxes)

data_juicer.ops.common.hawor_func.aa_to_rotmat(theta: Tensor)

Convert axis-angle representation to rotation matrix. Works by first converting it to a quaternion. :param theta: Tensor of shape (B, 3) containing axis-angle representations. :type theta: torch.Tensor

返回:

Corresponding rotation matrices with shape (B, 3, 3).

返回类型:

torch.Tensor

class data_juicer.ops.common.hawor_func.MANO(*args, joint_regressor_extra: str | None = None, **kwargs)

基类：MANOLayer

__init__(*args, joint_regressor_extra: str | None = None, **kwargs)

Extension of the official MANO implementation to support more joints. :param Same as MANOLayer.: :param joint_regressor_extra: Path to extra joint regressor. :type joint_regressor_extra: str

forward(*args, **kwargs) → MANOOutput

Run forward pass. Same as MANO and also append an extra set of joints if joint_regressor_extra is specified.

query(hmr_output)

data_juicer.ops.common.hawor_func.block_print()

data_juicer.ops.common.hawor_func.enable_print()

data_juicer.ops.common.hawor_func.run_mano(trans, root_orient, hand_pose, is_right=None, betas=None, use_cuda=True)

Forward pass of the SMPL model and populates pred_data accordingly with joints3d, verts3d, points3d.

trans : B x T x 3 root_orient : B x T x 3 body_pose : B x T x J*3 betas : (optional) B x D

data_juicer.ops.common.hawor_func.run_mano_left(trans, root_orient, hand_pose, is_right=None, betas=None, use_cuda=True, fix_shapedirs=True)

Forward pass of the SMPL model and populates pred_data accordingly with joints3d, verts3d, points3d.

trans : B x T x 3 root_orient : B x T x 3 body_pose : B x T x J*3 betas : (optional) B x D

data_juicer.ops.common.hawor_func.run_mano_twohands(init_trans, init_rot, init_hand_pose, is_right, init_betas, use_cuda=True, fix_shapedirs=True)

data_juicer.ops.common.hawor_func.quaternion_mul(q0, q1)

EXPECTS WXYZ :param q0 (, 4) :param q1 (, 4)

data_juicer.ops.common.hawor_func.quaternion_inverse(q, eps=1e-08)

EXPECTS WXYZ :param q (*, 4)

data_juicer.ops.common.hawor_func.quaternion_slerp(t, q0, q1, eps=1e-08)

:param t (, 1) must be between 0 and 1 :param q0 (, 4) :param q1 (*, 4)

data_juicer.ops.common.hawor_func.angle_axis_to_rotation_matrix(angle_axis)

:param angle_axis (, 3) return (, 3, 3)

data_juicer.ops.common.hawor_func.quaternion_to_rotation_matrix(quaternion)

Convert a quaternion to a rotation matrix. Taken from kornia/kornia, based on matthew-brett/transforms3d tensorflow/graphics :param quaternion (N, 4) expects WXYZ order returns rotation matrix (N, 3, 3)

data_juicer.ops.common.hawor_func.angle_axis_to_quaternion(angle_axis)

This function is borrowed from kornia/kornia Convert angle axis to quaternion in WXYZ order :param angle_axis (, 3) :returns quaternion (, 4) WXYZ order

class data_juicer.ops.common.hawor_func.HAWOR(cfg: CfgNode)

基类：LightningModule

__init__(cfg: CfgNode)

Setup HAWOR model :param cfg: Config file as a yacs CfgNode :type cfg: CfgNode

get_parameters()

configure_optimizers() → Optimizer

Setup model and distriminator Optimizers :returns: Model and discriminator optimizers :rtype: Tuple[torch.optim.Optimizer, torch.optim.Optimizer]

forward_step(batch: Dict, train: bool = False) → Dict

Run a forward step of the network :param batch: Dictionary containing batch data :type batch: Dict :param train: Flag indicating whether it is training or validation mode :type train: bool

返回:

Dictionary containing the regression output

返回类型:

Dict

compute_loss(batch: Dict, output: Dict, train: bool = True) → Tensor

Compute losses given the input batch and the regression output :param batch: Dictionary containing batch data :type batch: Dict :param output: Dictionary containing the regression output :type output: Dict :param train: Flag indicating whether it is training or validation mode :type train: bool

返回:

Total loss for current batch

返回类型:

torch.Tensor

tensorboard_logging(batch: Dict, output: Dict, step_count: int, train: bool = True, write_to_summary_writer: bool = True, render_log: bool = True) → None

Log results to Tensorboard :param batch: Dictionary containing batch data :type batch: Dict :param output: Dictionary containing the regression output :type output: Dict :param step_count: Global training step count :type step_count: int :param train: Flag indicating whether it is training or validation mode :type train: bool

forward(batch: Dict) → Dict

Run a forward step of the network in val mode :param batch: Dictionary containing batch data :type batch: Dict

返回:

Dictionary containing the regression output

返回类型:

Dict

training_step(joint_batch: Dict, batch_idx: int) → Dict

Run a full training step :param joint_batch: Dictionary containing image and mocap batch data :type joint_batch: Dict :param batch_idx: Unused. :type batch_idx: int :param batch_idx: Unused. :type batch_idx: torch.Tensor

返回:

Dictionary containing regression output.

返回类型:

Dict

inference(imgfiles, boxes, img_focal, img_center, device='cuda', do_flip=False)

validation_step(batch: Dict, batch_idx: int, dataloader_idx=0) → Dict

Run a validation step and log to Tensorboard :param batch: Dictionary containing batch data :type batch: Dict :param batch_idx: Unused. :type batch_idx: int

返回:

Dictionary containing regression output.

返回类型:

Dict

visualize_tensorboard(images, pred_keypoints, gt_project_j2d, gt_keypoints)

project(points, pred_cam, center, scale, img_focal, img_center, return_full=False)

get_trans(pred_cam, center, scale, img_focal, img_center)

bbox_est(center, scale, img_focal, img_center)