Spaces:
Sleeping
Sleeping
File size: 8,785 Bytes
d7e58f0 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 |
import os
import os.path as osp
import numpy as np
import torch
import cv2
import json
import copy
import csv
from pycocotools.coco import COCO
from config.config import cfg
from util.human_models import smpl_x
from util.transforms import world2cam, cam2pixel, rigid_align
from humandata import HumanDataset
from util.transforms import rigid_align, rigid_align_batch
class EgoBody_Kinect(HumanDataset):
def __init__(self, transform, data_split):
super(EgoBody_Kinect, self).__init__(transform, data_split)
if self.data_split == 'train':
filename = 'data/preprocessed_npz/multihuman_data/egobody_kinect_train_multi_080824.npz'
self.annot_path_cache = 'data/preprocessed_npz/cache/egobody_kinect_train_cache_080824.npz'
self.sample_interval = 10
else:
filename = 'data/preprocessed_npz/egobody_kinect_test_230503_043_fix_betas_multi.npz'
self.annot_path_cache = 'data/preprocessed_npz/egobody_kinect_test_230503_043_fix_betas_multi_cache_100.npz'
self.sample_interval = 100
self.use_betas_neutral = getattr(cfg, 'egobody_fix_betas', False)
self.img_dir = 'data/osx_data/EgoBody'
self.annot_path = filename
self.use_cache = getattr(cfg, 'use_cache', False)
self.img_shape = (1080, 1920) # (h, w)
self.cam_param = {}
# check image shape
img_path = osp.join(self.img_dir,
np.load(self.annot_path)['image_path'][0])
img_shape = cv2.imread(img_path).shape[:2]
assert self.img_shape == img_shape, 'image shape is incorrect: {} vs {}'.format(
self.img_shape, img_shape)
# load data or cache
if self.use_cache and osp.isfile(self.annot_path_cache):
print(
f'[{self.__class__.__name__}] loading cache from {self.annot_path_cache}'
)
self.datalist = self.load_cache(self.annot_path_cache)
else:
if self.use_cache:
print(
f'[{self.__class__.__name__}] Cache not found, generating cache...'
)
self.datalist = self.load_data(train_sample_interval=self.sample_interval)
if self.use_cache:
self.save_cache(self.annot_path_cache, self.datalist)
def evaluate(self, outs, cur_sample_idx):
annots = self.datalist
sample_num = len(outs)
eval_result = {
'pa_mpvpe_all': [],
'pa_mpvpe_l_hand': [],
'pa_mpvpe_r_hand': [],
'pa_mpvpe_hand': [],
'pa_mpvpe_face': [],
'mpvpe_all': [],
'mpvpe_l_hand': [],
'mpvpe_r_hand': [],
'mpvpe_hand': [],
'mpvpe_face': []
}
vis = getattr(cfg, 'vis', False)
vis_save_dir = cfg.vis_dir
for n in range(sample_num):
annot = annots[cur_sample_idx + n]
out = outs[n]
mesh_gt = out['smplx_mesh_cam_target']
mesh_out = out['smplx_mesh_cam']
ann_idx = out['gt_ann_idx']
img_path = []
for ann_id in ann_idx:
img_path.append(annots[ann_id]['img_path'])
eval_result['img_path'] = img_path
eval_result['ann_idx'] = ann_idx
# MPVPE from all vertices
mesh_out_align = \
mesh_out - np.dot(
smpl_x.J_regressor, mesh_out).transpose(1,0,2)[:, smpl_x.J_regressor_idx['pelvis'], None, :] + \
np.dot(smpl_x.J_regressor, mesh_gt).transpose(1,0,2)[:, smpl_x.J_regressor_idx['pelvis'], None, :]
eval_result['mpvpe_all'].extend(
np.sqrt(np.sum(
(mesh_out_align - mesh_gt)**2, -1)).mean(-1) * 1000)
mesh_out_align = rigid_align_batch(mesh_out, mesh_gt)
eval_result['pa_mpvpe_all'].extend(
np.sqrt(np.sum(
(mesh_out_align - mesh_gt)**2, -1)).mean(-1) * 1000)
# MPVPE from hand vertices
mesh_gt_lhand = mesh_gt[:, smpl_x.hand_vertex_idx['left_hand'], :]
mesh_out_lhand = mesh_out[:, smpl_x.hand_vertex_idx['left_hand'], :]
mesh_gt_rhand = mesh_gt[:, smpl_x.hand_vertex_idx['right_hand'], :]
mesh_out_rhand = mesh_out[:, smpl_x.hand_vertex_idx['right_hand'], :]
mesh_out_lhand_align = \
mesh_out_lhand - \
np.dot(smpl_x.J_regressor, mesh_out).transpose(1,0,2)[:, smpl_x.J_regressor_idx['lwrist'], None, :] + \
np.dot(smpl_x.J_regressor, mesh_gt).transpose(1,0,2)[:, smpl_x.J_regressor_idx['lwrist'], None, :]
mesh_out_rhand_align = \
mesh_out_rhand - \
np.dot(smpl_x.J_regressor, mesh_out).transpose(1,0,2)[:, smpl_x.J_regressor_idx['rwrist'], None, :] + \
np.dot(smpl_x.J_regressor, mesh_gt).transpose(1,0,2)[:, smpl_x.J_regressor_idx['rwrist'], None, :]
eval_result['mpvpe_l_hand'].extend(
np.sqrt(np.sum(
(mesh_out_lhand_align - mesh_gt_lhand)**2, -1)).mean(-1) *
1000)
eval_result['mpvpe_r_hand'].extend(
np.sqrt(np.sum(
(mesh_out_rhand_align - mesh_gt_rhand)**2, -1)).mean(-1) *
1000)
eval_result['mpvpe_hand'].extend(
(np.sqrt(np.sum(
(mesh_out_lhand_align - mesh_gt_lhand)**2, -1)).mean(-1) *
1000 +
np.sqrt(np.sum(
(mesh_out_rhand_align - mesh_gt_rhand)**2, -1)).mean(-1) *
1000) / 2.)
mesh_out_lhand_align = rigid_align_batch(mesh_out_lhand, mesh_gt_lhand)
mesh_out_rhand_align = rigid_align_batch(mesh_out_rhand, mesh_gt_rhand)
eval_result['pa_mpvpe_l_hand'].extend(
np.sqrt(np.sum(
(mesh_out_lhand_align - mesh_gt_lhand)**2, -1)).mean(-1) *
1000)
eval_result['pa_mpvpe_r_hand'].extend(
np.sqrt(np.sum(
(mesh_out_rhand_align - mesh_gt_rhand)**2, -1)).mean(-1) *
1000)
eval_result['pa_mpvpe_hand'].extend(
(np.sqrt(np.sum(
(mesh_out_lhand_align - mesh_gt_lhand)**2, -1)).mean(-1) *
1000 +
np.sqrt(np.sum(
(mesh_out_rhand_align - mesh_gt_rhand)**2, -1)).mean(-1) *
1000) / 2.)
vis = False
if vis:
import mmcv
img = (out['img']).transpose(0,2,3,1)
img = mmcv.imdenormalize(
img=img[0],
mean=np.array([123.675, 116.28, 103.53]),
std=np.array([58.395, 57.12, 57.375]),
to_bgr=True).astype(np.uint8)
from detrsmpl.core.visualization.visualize_keypoints2d import visualize_kp2d
import ipdb;ipdb.set_trace()
visualize_kp2d(
out['smplx_joint_proj'][0][None],
image_array=img[None].copy(),
disable_limbs=True,
overwrite=True,
output_path='./figs/pred2d'
)
from pytorch3d.io import save_obj
save_obj('temp.obj',verts=out['smplx_mesh_cam'][0],faces=torch.tensor([]))
# MPVPE from face vertices
mesh_gt_face = mesh_gt[:, smpl_x.face_vertex_idx, :]
mesh_out_face = mesh_out[:, smpl_x.face_vertex_idx, :]
mesh_out_face_align = \
mesh_out_face - \
np.dot(smpl_x.J_regressor, mesh_out).transpose(1,0,2)[:, smpl_x.J_regressor_idx['neck'], None, :] + \
np.dot(smpl_x.J_regressor, mesh_gt).transpose(1,0,2)[:, smpl_x.J_regressor_idx['neck'], None, :]
eval_result['mpvpe_face'].extend(
np.sqrt(np.sum(
(mesh_out_face_align - mesh_gt_face)**2, -1)).mean(-1) * 1000)
mesh_out_face_align = rigid_align_batch(mesh_out_face, mesh_gt_face)
eval_result['pa_mpvpe_face'].extend(
np.sqrt(np.sum(
(mesh_out_face_align - mesh_gt_face)**2, -1)).mean(-1) * 1000)
# for k,v in eval_result.items():
# if k != 'img_path' and k != 'ann_idx':
# # import ipdb;ipdb.set_trace()
# if len(v)>1:
# eval_result[k] = np.concatenate(v,axis=0)
# else:
# eval_result[k] = np.array(v)
return eval_result |