Functions — github.com/akanazawa/human

↓ 162 callersMethodappend

Appends an omega. Args: omega (Bx85): [cams, poses, shapes].

src/omega.py:250

↓ 26 callersMethodload

(self)

src/data_loader_sequence.py:56

↓ 23 callersMethodupdate

(self, vars)

src/discriminators.py:83

↓ 16 callersFunctiondraw_skeleton

joints is 3 x 19. but if not will transpose it. 0: Right heel 1: Right knee 2: Right hip 3: Left hip 4: Left knee 5: Left

src/util/render/render_utils.py:38

↓ 15 callersMethodget_kps

Returns the keypoints at time t. Note that the shape is different for ground truth omegas and predicted omegas. Arg

src/omega.py:67

↓ 14 callersMethodget_shapes

Returns shapes at time t. Args: t (int). Returns: Shapes (Bx10).

src/omega.py:131

↓ 13 callersFunctionfloat_feature

Wrapper for inserting float features into Example proto.

src/datasets/common.py:173

↓ 11 callersMethodencode_jpeg

(self, image)

src/datasets/common.py:62

↓ 11 callersFunctionprocess_tensors

Wrapper function for processing sparse tensors outputted by variable length features. 1. Converts sparse tensors to dense tensors. 2

src/util/data_utils.py:478

↓ 10 callersMethodadd_scalar_summary

Helper function to add scalar summary. Args: summary_name (str). scalar_value (float).

src/trainer_sequence_fc.py:687

↓ 10 callersMethoddecode_jpeg

(self, image_data)

src/datasets/common.py:53

↓ 10 callersMethodget_poses_rot

Returns poses as rotation matrices at time t. Args: t (int). Returns: Poses (Bx24x3x3).

src/omega.py:95

↓ 8 callersMethodappend_batched

Appends multiple omegas. Args: omegas (BxTx85): [cams, poses, shapes].

src/omega.py:237

↓ 8 callersFunctionbatch_rodrigues

Theta is N x 3

src/tf_smpl/batch_lbs.py:42

↓ 8 callersMethodget_cams

Gets cams at time t. Args: t (int). Returns: Cams (Bx3).

src/omega.py:306

↓ 8 callersFunctionint64_feature

Wrapper for inserting int64 features into Example proto.

src/datasets/common.py:166

↓ 8 callersFunctionresize_img

(img, scale_factor)

src/util/common.py:7

↓ 7 callersMethodget_joints

Returns the joints at time t. Args: t (int). Returns: Joints (Bx25x3).

src/omega.py:55

↓ 6 callersFunctioncompute_errors_batched

Computes errors.

src/evaluation/eval.py:114

↓ 6 callersMethodcompute_smpl

Batch computation of vertices, joints, rotation matrices, and keypoints. Due to the overhead added to computation graph, call this on

src/omega.py:263

↓ 6 callersFunctionto_variable

(x)

src/util/render/nmr_renderer.py:243

↓ 5 callersFunctionconvert_to_example_temporal

Builds an Example proto for an image temporal example for N images. Note, no paired mosh data is available, so all poses and shape are None

src/datasets/common.py:187

↓ 5 callersFunctionget_smooth_bbox_params

Computes smooth bounding box parameters from keypoints: 1. Computes bbox by rescaling the person to be around 150 px. 2. Linearly int

src/util/smooth_bbox.py:10

↓ 5 callersMethodget_verts

(self)

src/omega.py:328

↓ 5 callersFunctionmkdir

(dir_path)

src/util/common.py:17

↓ 5 callersFunctionsave_seq_to_test_tfrecord

Saves a sequence to test format for rendering and evaluation. Args: out_name (str): Tfrecord filename. im_paths (list of len

src/datasets/make_test_tfrecords.py:22

↓ 5 callersFunctionto_float_tensor

(x)

src/util/render/nmr_renderer.py:249

↓ 4 callersFunctionbatch_pred_omega

Given B x T x * inputs, computes IEF on them by batching them as BT x *. if use_optcam is True, only outputs 72 or 82 dims. and appe

src/models.py:233

↓ 4 callersMethodcompute_all_phis

Computes the image features for any arbitrary sequence of images. Args: all_images (TxHxWx3). Returns:

src/datasets/resnet_extractor.py:74

↓ 4 callersFunctioncompute_gpu_smpl

(poses, shapes, get_joints=False)

src/evaluation/eval.py:68

↓ 4 callersFunctioncompute_loss_e_kp

L_{KP}=\sum_{i=1}^K v_i||x_i - (\hat{x_i}||_1 Args: kp_gt (NxKx3): Ground truth kp. kp_pred (NxKx2): Predicted kp. n

src/ops.py:41

↓ 4 callersMethodgather

Gathers a subset over time. Args: values (BxTx...): Tensor that we only need a subset of. indices (iterable)

src/omega.py:144

↓ 4 callersMethodmake_omega_pred

(self, use_optcam=False)

src/trainer_sequence_fc.py:481

↓ 4 callersFunctionmkdir

(dir_path)

src/datasets/upenn_to_tfrecords_video.py:432

↓ 4 callersFunctionrepeat_col

(col, num_repeat)

src/util/data_utils.py:560

↓ 4 callersFunctiontf_repeat

Repeats elements of a tensor. Tensorflow implementation of np.repeat. Args: tensor (tensor): Input tensor. repeat (int):

src/util/data_utils.py:765

↓ 4 callersFunctionundo_chumpy

(x)

src/tf_smpl/batch_smpl.py:22

↓ 4 callersFunctionupdate_dict_entries

Appends values in appender to list in accumulator. Args: accumulator (dict): Dict of lists with entries from appenders. appe

src/evaluation/eval_util.py:302

↓ 4 callersMethodview

For debug.

src/trainer_sequence_fc.py:1122

↓ 3 callersFunctionbytes_feature

Wrapper for inserting bytes features into Example proto.

src/datasets/common.py:180

↓ 3 callersFunctioncompute_loss_e_3d

(poses_gt, poses_pred, shapes_gt, shapes_pred, joints_gt, joints_pred, batch_size, has_g

src/ops.py:59

↓ 3 callersFunctioncompute_loss_mse

Computes the l2 loss between 3D params pred and gt for the data that has_gt3d is True. Parameters to compute loss over: 3Djoints: 14

src/ops.py:87

↓ 3 callersMethodcompute_losses_deltas

Computes all delta losses. Args: omegas_dict (dict): Dict mapping delta_t to Omegas. suffix_future (str): Su

src/trainer_sequence_fc.py:848

↓ 3 callersFunctionget_config

()

src/config.py:137

↓ 3 callersFunctionget_dims

(x)

src/util/render/nmr_renderer.py:39

↓ 3 callersFunctionmkdir

(dir_path)

src/datasets/insta_variety_to_tfrecords.py:518

↓ 3 callersFunctionmkdir

(dir_path)

src/datasets/video_in_the_wild_to_tfrecords.py:497

↓ 3 callersMethodpredict_all_images

Wrapper to predict entire sequence. Because of edge padding, images at edges will have low quality predictions since they do

src/evaluation/tester.py:260

↓ 3 callersFunctionremove_pads

Undos padding done by make_square.

src/util/render/nmr_renderer.py:507

↓ 3 callersMethodrotated

vert is N x 3, torch FloatTensor (or Variable)

src/util/render/nmr_renderer.py:176

↓ 2 callersMethod__init__

(self, config, batch_size=None)

src/omega.py:23

↓ 2 callersFunctionalign_by_pelvis

Assumes joints is N x 14 x 3 in LSP order. Then hips are: [3, 2] Takes mid point of these points, then subtracts it.

src/ops.py:167

↓ 2 callersFunctionalign_by_pelvis

Aligns joints by pelvis to be at origin. Assumes hips are index 3 and 2 of joints (14x3) in LSP order. Pelvis is midpoint of hips. Args:

src/evaluation/eval_util.py:158

↓ 2 callersFunctionbatch_orth_proj_idrot

X is N x num_points x 3 camera is N x 3 same as applying orth_proj_idrot to each N

src/tf_smpl/projection.py:16

↓ 2 callersFunctionclean_video

returns None if video is bad ow. returns cleaned/adjusted image_paths and gt2ds

src/datasets/upenn_to_tfrecords_video.py:291

↓ 2 callersFunctionclean_video

returns None if video is bad ow. returns cleaned/adjusted image_paths and gt2ds

src/datasets/insta_variety_to_tfrecords.py:273

↓ 2 callersFunctionclean_video

returns None if video is bad ow. returns cleaned/adjusted image_paths and gt2ds

src/datasets/video_in_the_wild_to_tfrecords.py:274

↓ 2 callersFunctioncompute_error_verts

Computes MPJPE over 6890 surface vertices. Args: verts_gt (Nx6989x3). verts_pred (Nx6989x3). Returns: error_ver

src/evaluation/eval_util.py:140

↓ 2 callersMethodcompute_losses_prior

(self)

src/trainer_sequence_fc.py:989

↓ 2 callersFunctioncompute_tracks

This script basically: 1. Extracts individual frames from mp4 since PoseFlow requires per frame images to be written. 2. Call Alph

extract_tracks.py:127

↓ 2 callersMethoddecode_png

(self, image_data)

src/datasets/common.py:76

↓ 2 callersFunctionencoder_fc3_dropout

3D inference module. 3 MLP layers (last is the output) With dropout on first 2. Input: - x: N x [|img_feat|, |3D_param|] - reuse

src/models.py:80

↓ 2 callersFunctionextend_dict_entries

Extends values in accumulator with appender.

src/evaluation/eval_util.py:278

↓ 2 callersFunctionflip_image

Flipping image and kp. kp is 3 x N! pose is 72D gt3d is 14 x 3

src/util/data_utils.py:601

↓ 2 callersFunctionget_hallucinator_model

(model_type='fc2_res')

src/models.py:26

↓ 2 callersFunctionget_image_encoder

Retrieves encoder fn for image and 3D

src/models.py:12

↓ 2 callersFunctionget_pred_path_name

Gets path name to store cached predictions. If directory doesn't exist, build it automatically. File structure: +-- PRED_DIR

src/evaluation/prediction.py:22

↓ 2 callersFunctionget_seq_labels

(data_dir, split)

src/datasets/insta_variety_to_tfrecords.py:444

↓ 2 callersFunctionget_seq_labels

(data_dir, split, video_list_file)

src/datasets/video_in_the_wild_to_tfrecords.py:445

↓ 2 callersFunctionget_sequences

(data_dir, split)

src/datasets/threedpw/read_3dpw.py:69

↓ 2 callersFunctionget_temporal_encoder

(model_type='AZ_FC2GN')

src/models.py:37

↓ 2 callersFunctionget_upenn2coco

Converts Upenn keypoints to 25 universal keypoints with toes. Note: UPenn does not have "heel". H36M has heel but not ankles.

src/datasets/upenn/read_upenn.py:36

↓ 2 callersFunctionget_verts

(fname)

src/datasets/threedpw/compute_neutral_shape.py:45

↓ 2 callersFunctionhamilton_product

Multiply qa by qb. Args: qa: B X N X 4 quaternions qb: B X N X 4 quaternions Returns: q_mult: B X N X 4

src/util/render/torch_utils.py:76

↓ 2 callersFunctionhmr_ief

Runs HMR-style IEF. Args: phi (Bx2048): Image features. omega_start (Bx85): Starting Omega as input to first IEF. sc

src/models.py:380

↓ 2 callersFunctionhmr_ief

Runs HMR-style IEF. Args: phi (Bx2048): Image features. omega_start (Bx85): Starting Omega as input to first IEF. sc

src/ops.py:270

↓ 2 callersFunctionmake_A

(R, t, name=None)

src/tf_smpl/batch_lbs.py:163

↓ 2 callersMethodmake_fetch_dict

(self, omegas, suffix='')

src/evaluation/tester.py:217

↓ 2 callersMethodmake_omega_pred

(self, use_optcam=False)

src/evaluation/tester.py:154

↓ 2 callersFunctionmake_square

Bc nmr only deals with square image, adds pad to the shorter side.

src/util/render/nmr_renderer.py:491

↓ 2 callersFunctionmake_video

output_path is the final mp4 name img_dir is where the images to make into video are saved.

src/evaluation/run_video.py:205

↓ 2 callersFunctionmean_of_dict_values

Flattens values of dictionary and computes the mean.

src/evaluation/eval_util.py:291

↓ 2 callersFunctionprint_summary

(errors_dict)

src/evaluation/eval.py:330

↓ 2 callersFunctionprocess_image

Processes an image, producing 224x224 crop. Args: im_path (str). bbox_param (3,): [cx, cy, scale]. visualize (bool).

src/evaluation/run_video.py:56

↓ 2 callersFunctionprocess_videos_test

seq_path is a path to the directory of a seq. each video becomes a tfrecord

src/datasets/upenn_to_tfrecords_video.py:376

↓ 2 callersMethodprune_variables

(self, variables, skip_prefixes)

src/trainer_sequence_fc.py:667

↓ 2 callersMethodread_data

Reads data from given filename queue and pre-processes the data. Args: filename_queue: Queue of filename strings.

src/data_loader_sequence.py:267

↓ 2 callersFunctionread_from_example

Returns data from an entry in test tfrecord. Args: serialized_ex (str). Returns: dict. Keys: N (1).

src/datasets/common.py:86

↓ 2 callersFunctionread_labels

Returns: - kp: N x 13 x 3 - is_train

src/datasets/upenn/read_upenn.py:103

↓ 2 callersFunctionread_poses

(path, n_frames=None, is_3d=False, joint_ids=range(32))

src/datasets/h36/read_human36m.py:231

↓ 2 callersFunctionrender_preds

src/evaluation/run_video.py:110

↓ 2 callersFunctionrun_on_video

Main driver. First extracts alphapose/posetrack in track_dir Then runs HMMR.

demo_video.py:194

↓ 2 callersFunctionsave_results

(config, all_dataset_results, json_path='')

src/evaluation/eval.py:341

↓ 2 callersFunctionsave_seq_to_tfrecord

(out_name, im_paths, gt2ds, gt3ds,

src/datasets/h36_to_tfrecords_video.py:329

↓ 2 callersMethodset_cams

Only used for opt_cam

src/omega.py:318

↓ 2 callersMethodupdate_E_vars

(self)

src/trainer_sequence_fc.py:675

↓ 2 callersMethodupdate_instance_vars

(self)

src/omega.py:231

↓ 2 callersFunctionvisualize_img

Visualizes the image with the ground truth keypoints and predicted keypoints on left and image with mesh on right. Keypoints should be i

src/util/render/nmr_renderer.py:265

↓ 2 callersMethodvisualize_strip

Visualizes the delta and hal predictions side-by-side. Each strip contains projected 2D skeleton on left and mesh overlaid on gt imag

src/trainer_sequence_fc.py:1284

↓ 1 callersMethod__call__

verts is |V| x 3 numpy/cuda torch Variable or B x V x 3 cams is 3D [s, tx, ty], numpy/cuda torch Variable or B x 3 cams is NO

src/util/render/nmr_renderer.py:81

Functions299 in github.com/akanazawa/human_dynamics