electron-rare/SMPLer-X
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

clean the hardcode and debug code

Browse Source
This commit is contained in:
yinwanqi
2023-06-15 00:51:29 +08:00
parent ffbe466651
commit 028331d707
11 changed files with 10 additions and 325 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/utils/preprocessing.py
-25
View File
@@ -206,7 +206,6 @@ def process_db_coord(joint_img, joint_cam, joint_valid, do_flip, img_shape, flip
joint_img[:, 1] = joint_img[:, 1] / cfg.input_img_shape[0] * cfg.output_hm_shape[1]
# check truncation
# TODO
joint_trunc = joint_valid * ((joint_img_original[:, 0] > 0) * (joint_img[:, 0] >= 0) * (joint_img[:, 0] < cfg.output_hm_shape[2]) * \
(joint_img_original[:, 1] > 0) *(joint_img[:, 1] >= 0) * (joint_img[:, 1] < cfg.output_hm_shape[1]) * \
(joint_img_original[:, 2] > 0) *(joint_img[:, 2] >= 0) * (joint_img[:, 2] < cfg.output_hm_shape[0])).reshape(-1,
@@ -296,8 +295,6 @@ def process_human_model_output(human_model_param, cam_param, do_flip, img_shape,
leye_pose=zero_pose, reye_pose=zero_pose, expression=expr)
mesh_cam = output.vertices[0].numpy()
joint_cam = output.joints[0].numpy()[smpl_x.joint_idx, :]
### HARDCODE
# joint_cam_orig_ = joint_cam.copy()
# apply camera exrinsic (translation)
# compenstate rotation (translation from origin to root joint was not cancled)
@@ -489,28 +486,6 @@ def process_human_model_output(human_model_param, cam_param, do_flip, img_shape,
pose = pose.reshape(-1)
expr = expr.numpy().reshape(-1)
### ### HARDCODE temp save vis for debug
# POSE: torch.cat((root_pose, body_pose, lhand_pose, rhand_pose, jaw_pose))
# pose_ = pose.numpy().copy()
# root_pose = torch.FloatTensor(pose[0:3]).view(1, 3) # (1,3)
# body_pose = torch.FloatTensor(pose[3:66]).view(-1, 3) # (21,3)
# lhand_pose = torch.FloatTensor(pose[66:111]).view(-1, 3) # (15,3)
# rhand_pose = torch.FloatTensor(pose[111:156]).view(-1, 3) # (15,3)
# jaw_pose = torch.FloatTensor(pose[156:159]).view(-1, 3) # (1,3)
# shape = torch.FloatTensor(shape).view(1, -1) # SMPLX shape parameter
# expr = torch.FloatTensor(expr).view(1, -1) # SMPLX expression parameter
# trans = torch.FloatTensor(trans).view(1, -1) # translation vector
# with torch.no_grad():
# output = smpl_x.layer[gender](betas=shape, body_pose=body_pose.view(1, -1), global_orient=root_pose,
# transl=zero_pose, left_hand_pose=lhand_pose.view(1, -1),
# right_hand_pose=rhand_pose.view(1, -1), jaw_pose=jaw_pose.view(1, -1),
# leye_pose=zero_pose, reye_pose=zero_pose, expression=expr)
# mesh_rot = output.vertices[0].numpy()
# joint_rot = output.joints[0].numpy()[smpl_x.joint_idx, :]
# return mesh_rot, joint_cam, joint_trunc, pose, shape, expr, rotation_valid, coord_valid, expr_valid, mesh_cam_orig, joint_cam_orig_
return joint_img, joint_cam, joint_trunc, pose, shape, expr, rotation_valid, coord_valid, expr_valid, mesh_cam_orig
elif human_model_type == 'smpl':
pose = pose.reshape(-1)
data/AGORA/AGORA.py
+2 -100
View File
@@ -157,14 +157,9 @@ class AGORA(torch.utils.data.Dataset):
else:
db = COCO(osp.join(self.data_path, 'AGORA_validation.json'))
### HARDCODE vis for debug
# count = 0
i = 0
for aid in tqdm.tqdm(list(db.anns.keys())):
# if count > 50:
# continue
# count += 1
i += 1
if self.data_split == 'train' and i % getattr(cfg, 'AGORA_train_sample_interval', 1) != 0:
continue
@@ -418,8 +413,6 @@ class AGORA(torch.utils.data.Dataset):
joint_img[:, 1] = joint_img[:, 1] / 2160 * self.resolution[0]
with open(data['joints_3d_path']) as f:
joint_cam = np.array(json.load(f)).reshape(-1, 3)
### HARDCODE vis for debug
# joint_cam_orig = joint_cam.copy()
with open(data['smplx_param_path'], 'rb') as f:
smplx_param = pickle.load(f, encoding='latin1')
@@ -438,30 +431,6 @@ class AGORA(torch.utils.data.Dataset):
rhand_bbox_size = rhand_bbox[1] - rhand_bbox[0];
face_bbox_size = face_bbox[1] - face_bbox[0];
"""
# for debug
_img = img.numpy().transpose(1,2,0)[:,:,::-1].copy() * 255
if lhand_bbox_valid:
_tmp = lhand_bbox.copy().reshape(2,2)
_tmp[:,0] = _tmp[:,0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1]
_tmp[:,1] = _tmp[:,1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0]
cv2.rectangle(_img, (int(_tmp[0,0]), int(_tmp[0,1])), (int(_tmp[1,0]), int(_tmp[1,1])), (255,0,0), 3)
cv2.imwrite('agora_' + str(idx) + '_lhand.jpg', _img)
if rhand_bbox_valid:
_tmp = rhand_bbox.copy().reshape(2,2)
_tmp[:,0] = _tmp[:,0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1]
_tmp[:,1] = _tmp[:,1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0]
cv2.rectangle(_img, (int(_tmp[0,0]), int(_tmp[0,1])), (int(_tmp[1,0]), int(_tmp[1,1])), (255,0,0), 3)
cv2.imwrite('agora_' + str(idx) + '_rhand.jpg', _img)
if face_bbox_valid:
_tmp = face_bbox.copy().reshape(2,2)
_tmp[:,0] = _tmp[:,0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1]
_tmp[:,1] = _tmp[:,1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0]
cv2.rectangle(_img, (int(_tmp[0,0]), int(_tmp[0,1])), (int(_tmp[1,0]), int(_tmp[1,1])), (255,0,0), 3)
cv2.imwrite('agora_' + str(idx) + '_face.jpg', _img)
#cv2.imwrite('agora_' + str(idx) + '.jpg', _img)
"""
# coordinates
joint_cam = joint_cam - joint_cam[self.joint_set['root_joint_idx'], None, :] # root-relative
joint_cam[self.joint_set['joint_part']['lhand'], :] = joint_cam[self.joint_set['joint_part']['lhand'],
@@ -507,27 +476,6 @@ class AGORA(torch.utils.data.Dataset):
joint_cam_wo_ra[smpl_x.joint_part['face'], :] = joint_cam_wo_ra[smpl_x.joint_part['face'], :] \
+ joint_cam_wo_ra[smpl_x.neck_idx, None,: ] # face root-relative
"""
# for debug
_tmp = joint_img.copy()
_tmp[:,0] = _tmp[:,0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1]
_tmp[:,1] = _tmp[:,1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0]
_img = img.numpy().transpose(1,2,0)[:,:,::-1] * 255
_img = vis_keypoints(_img.copy(), _tmp)
cv2.imwrite('agora_' + str(idx) + '.jpg', _img)
"""
"""
# for debug
_tmp = joint_cam.copy()[:,:2]
_tmp[:,0] = _tmp[:,0] / (cfg.body_3d_size / 2) * cfg.input_img_shape[1] + cfg.input_img_shape[1]/2
_tmp[:,1] = _tmp[:,1] / (cfg.body_3d_size / 2) * cfg.input_img_shape[0] + cfg.input_img_shape[0]/2
_img = np.zeros((cfg.input_img_shape[0], cfg.input_img_shape[1], 3), dtype=np.float32)
_img = vis_keypoints(_img.copy(), _tmp)
cv2.imwrite('agora_' + str(idx) + '_cam.jpg', _img)
"""
# smplx parameters
root_pose = np.array(smplx_param['global_orient'], dtype=np.float32).reshape(
-1) # rotation to world coordinate
@@ -553,9 +501,6 @@ class AGORA(torch.utils.data.Dataset):
'trans': trans}
_, _, _, smplx_pose, smplx_shape, smplx_expr, smplx_pose_valid, _, smplx_expr_valid, _ = process_human_model_output(
smplx_param, cam_param, do_flip, img_shape, img2bb_trans, rot, 'smplx')
### HARDCODE vis for debug
# mesh_rot_, joint_cam_, _, smplx_pose, smplx_shape, smplx_expr, smplx_pose_valid, _, smplx_expr_valid, mesh_orig, joint_cam_orig_ = process_human_model_output(
# smplx_param, cam_param, do_flip, img_shape, img2bb_trans, rot, 'smplx')
smplx_pose_valid = np.tile(smplx_pose_valid[:, None], (1, 3)).reshape(-1)
if not getattr(cfg, 'agora_valid_root_pose', False):
@@ -573,8 +518,6 @@ class AGORA(torch.utils.data.Dataset):
'smplx_pose_valid': smplx_pose_valid, 'smplx_shape_valid': float(smplx_shape_valid),
'smplx_expr_valid': float(smplx_expr_valid), 'is_3D': float(True),
'lhand_bbox_valid': lhand_bbox_valid, 'rhand_bbox_valid': rhand_bbox_valid, 'face_bbox_valid': face_bbox_valid}
### HARDCODE vis for debug
# 'gt_3d_path': data['joints_3d_path'], 'smplx_path': data['smplx_param_path'], 'id': idx}
return inputs, targets, meta_info
else:
# load crop and resize information (for the 4K setting)
@@ -678,41 +621,8 @@ class AGORA(torch.utils.data.Dataset):
eval_result['pa_mpvpe_face'].append(
np.sqrt(np.sum((mesh_out_face_align - mesh_gt_face) ** 2, 1)).mean() * 1000)
### HARDCODE
if vis:
# from utils.vis import vis_keypoints, vis_mesh, save_obj, render_mesh
# # img = (out['img'].transpose(1,2,0)[:,:,::-1] * 255).copy()
# # joint_img = out['joint_img'].copy()
# # joint_img[:,0] = joint_img[:,0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1]
# # joint_img[:,1] = joint_img[:,1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0]
# # for j in range(len(joint_img)):
# # cv2.circle(img, (int(joint_img[j][0]), int(joint_img[j][1])), 3, (0,0,255), -1)
# # cv2.imwrite(str(cur_sample_idx + n) + '.jpg', img)
# img_path = annot['img_path']
# img_id = img_path.split('/')[-1][:-4]
# ann_id = 0
# # ann_id = annot['ann_id']
# img = load_img(img_path)[:, :, ::-1]
# bbox = annot['bbox']
# focal = list(cfg.focal)
# princpt = list(cfg.princpt)
# focal[0] = focal[0] / cfg.input_body_shape[1] * bbox[2]
# focal[1] = focal[1] / cfg.input_body_shape[0] * bbox[3]
# princpt[0] = princpt[0] / cfg.input_body_shape[1] * bbox[2] + bbox[0]
# princpt[1] = princpt[1] / cfg.input_body_shape[0] * bbox[3] + bbox[1]
# img = render_mesh(img, out['smplx_mesh_cam'], smpl_x.face, {'focal': focal, 'princpt': princpt}, mesh_as_vertices=True)
# # img = cv2.resize(img, (512,512))
# cv2.imwrite(osp.join(vis_save_dir, img_id + '_' + str(ann_id) + '.jpg'), img)
# vis_mesh_out = out['smplx_mesh_cam']
# vis_mesh_out = vis_mesh_out - np.dot(smpl_x.layer['neutral'].J_regressor, vis_mesh_out)[
# smpl_x.J_regressor_idx['pelvis'], None, :]
# # vis_mesh_gt = out['smplx_mesh_cam_target']
# # vis_mesh_gt = vis_mesh_gt - np.dot(smpl_x.layer['neutral'].J_regressor, vis_mesh_gt)[smpl_x.J_regressor_idx['pelvis'],None,:]
# # save_obj(vis_mesh_out, smpl_x.face, osp.join(img_id + '_' + str(ann_id) + '.obj'))
# # save_obj(vis_mesh_gt, smpl_x.face, str(cur_sample_idx + n) + '_gt.obj')
img_path = out['img_path']
rel_img_path = img_path.split('..')[-1]
smplx_pred = {}
@@ -785,14 +695,6 @@ class AGORA(torch.utils.data.Dataset):
with open(osp.join(cfg.result_dir, 'predictions', save_name), 'wb') as f:
pickle.dump(save_dict, f)
"""
# for debug
img_path = annot['img_path']
img_path = osp.join(self.data_path, '3840x2160', 'test', img_path.split('/')[-1].split('_')[0] + '.png')
img = cv2.imread(img_path)
img = vis_keypoints(img.copy(), joint_proj)
cv2.imwrite(img_path.split('/')[-1], img)
"""
if getattr(cfg, 'vis', False):
file.close()
data/EHF/EHF.py
-1
View File
@@ -291,7 +291,6 @@ class EHF(torch.utils.data.Dataset):
smplx_pred['expression'] = out['smplx_expr'].reshape(-1,10)
smplx_pred['transl'] = out['cam_trans'].reshape(-1,3)
# import pdb; pdb.set_trace()
np.savez(os.path.join(cfg.vis_dir, f'{self.save_idx}.npz'), **smplx_pred)
# save img path and error
data/Human36M/Human36M.py
-9
View File
@@ -192,15 +192,6 @@ class Human36M(torch.utils.data.Dataset):
smplx_pose_valid, smplx_joint_valid, smplx_expr_valid, smplx_mesh_cam_orig = \
process_human_model_output(smplx_param, cam_param, do_flip, img_shape, img2bb_trans, rot, 'smplx')
"""
# for debug
_tmp = joint_img.copy()
_tmp[:,0] = _tmp[:,0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1]
_tmp[:,1] = _tmp[:,1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0]
_img = img.numpy().transpose(1,2,0)[:,:,::-1] * 255
_img = vis_keypoints(_img, _tmp)
cv2.imwrite('h36m_' + str(idx) + '.jpg', _img)
"""
# reverse ra
smplx_joint_cam_wo_ra = smplx_joint_cam.copy()
smplx_joint_cam_wo_ra[smpl_x.joint_part['lhand'], :] = smplx_joint_cam_wo_ra[smpl_x.joint_part['lhand'], :] \
data/MPII/MPII.py
-10
View File
@@ -131,16 +131,6 @@ class MPII(torch.utils.data.Dataset):
smplx_joint_img, smplx_joint_cam, smplx_joint_trunc, smplx_pose, smplx_shape, smplx_expr, smplx_pose_valid, smplx_joint_valid, smplx_expr_valid, smplx_mesh_cam_orig = process_human_model_output(smplx_param['smplx_param'], smplx_param['cam_param'], do_flip, img_shape, img2bb_trans, rot, 'smplx')
is_valid_fit = True
"""
# for debug
_tmp = joint_img.copy()
_tmp[:,0] = _tmp[:,0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1]
_tmp[:,1] = _tmp[:,1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0]
_img = img.numpy().transpose(1,2,0)[:,:,::-1] * 255
_img = vis_keypoints(_img.copy(), _tmp)
cv2.imwrite('mpii_' + str(idx) + '.jpg', _img)
"""
else:
# dummy values
smplx_joint_img = np.zeros((smpl_x.joint_num,3), dtype=np.float32)
data/MSCOCO/MSCOCO.py
-10
View File
@@ -321,16 +321,6 @@ class MSCOCO(torch.utils.data.Dataset):
img_shape, img2bb_trans, rot, 'smplx')
is_valid_fit = True
"""
# for debug
_tmp = joint_img.copy()
_tmp[:,0] = _tmp[:,0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1]
_tmp[:,1] = _tmp[:,1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0]
_img = img.numpy().transpose(1,2,0)[:,:,::-1] * 255
_img = vis_keypoints(_img, _tmp)
cv2.imwrite('coco_' + str(idx) + '.jpg', _img)
"""
else:
# dummy values
smplx_joint_img = np.zeros((smpl_x.joint_num, 3), dtype=np.float32)
data/PW3D/PW3D.py
-21
View File
@@ -165,10 +165,6 @@ class PW3D(torch.utils.data.Dataset):
# 'left_elbow', 'right_elbow', 'left_wrist', 'right_wrist']
joint_mapper = [1, 2, 4, 5, 7, 8, 12, 15, 16, 17, 18, 19, 20, 21]
### Save vis for debug
# joint_gt_body_to_save = np.zeros((sample_num, len(joint_mapper), 3))
# joint_out_body_root_align_to_save = np.zeros((sample_num, len(joint_mapper), 3))
# joint_out_body_pa_align_to_save = np.zeros((sample_num, len(joint_mapper), 3))
for n in range(sample_num):
@@ -182,11 +178,6 @@ class PW3D(torch.utils.data.Dataset):
mesh_out_align = mesh_out - np.dot(smpl_x.J_regressor, mesh_out)[smpl_x.J_regressor_idx['pelvis'], None, :] \
+ np.dot(smpl.joint_regressor, mesh_gt)[smpl.root_joint_idx, None, :]
# only eval point0-21 since only smpl gt is given
# joint_gt_body = np.dot(smpl.joint_regressor, mesh_gt)[:22, :]
# joint_out_body = np.dot(smpl_x.J_regressor, mesh_out)[:22, :]
# joint_out_body_root_align = np.dot(smpl_x.J_regressor, mesh_out_align)[:22, :]
# only test 14 keypoints
joint_gt_body = np.dot(smpl.joint_regressor, mesh_gt)[joint_mapper, :]
joint_out_body = np.dot(smpl_x.J_regressor, mesh_out)[joint_mapper, :]
@@ -199,18 +190,6 @@ class PW3D(torch.utils.data.Dataset):
joint_out_body_pa_align = rigid_align(joint_out_body, joint_gt_body)
eval_result['pa_mpjpe_body'].append(
np.sqrt(np.sum((joint_out_body_pa_align - joint_gt_body) ** 2, 1)).mean() * 1000)
### Save vis for debug
# joint_gt_body_to_save[n, ...] = joint_gt_body
# joint_out_body_root_align_to_save[n, ...] = joint_out_body_root_align
# joint_out_body_pa_align_to_save[n, ...] = joint_out_body_pa_align
### Save vis for debug
# import numpy as np
# np.save(f'./vis/val_0509_joint_gt_body.npy', joint_gt_body_to_save)
# np.save(f'./vis/val_0509_joint_out_body_root_align.npy', joint_out_body_root_align_to_save)
# np.save(f'./vis/val_0509_joint_out_body_pa_align.npy', joint_out_body_pa_align_to_save)
# import pdb; pdb.set_trace()
return eval_result
data/SHAPY/SHAPY.py
-6
View File
@@ -96,8 +96,6 @@ class SHAPY(HumanDataset):
sample_num = len(outs)
eval_result = {'v2v_t_errors': [], 'point_t_errors': [], 'height': [], 'chest': [], 'waist': [], 'hips': [], 'mass': []}
# sample_num = sample_num // 10 # TODO: debug only
for n in range(sample_num):
annot = annots[cur_sample_idx + n]
out = outs[n]
@@ -107,8 +105,6 @@ class SHAPY(HumanDataset):
# compute v_shaped
betas_fit = torch.tensor(betas_fit.reshape(-1, 10)).cuda()
# betas_fit = torch.tensor(out['smplx_shape_target'].reshape(-1, 10)).cuda() # TODO: debug only
# betas_fit = torch.zeros((1, 10)).cuda() # TODO: debug only
output = self.smplx_layer(
betas=betas_fit,
body_pose=torch.zeros((1, 63)).to(betas_fit.device),
@@ -122,8 +118,6 @@ class SHAPY(HumanDataset):
return_verts=True
)
v_shaped_fit = output.vertices.detach().cpu().numpy().squeeze()
# v_shaped_gt = v_shaped_fit # TODO: debug only
# v_shaped_fit = self.smplx_layer.forward_shape(betas=betas_fit)
image_name = '/'.join(img_path.split('/')[-4:])
self.images_names.append(image_name)
data/UBody/UBody.py
+3 -71
View File
@@ -235,9 +235,7 @@ class UBody_Part(torch.utils.data.Dataset):
video_name = file_name.split('/')[-2]
if 'Trim' in video_name:
video_name = video_name.split('_Trim')[0]
# if video_name in test_video_list:
# # data to use in test
# import pdb; pdb.set_trace()
if video_name not in test_video_list: continue # exclude the train video
img_path = osp.join(self.img_path, file_name)
if not os.path.exists(img_path): continue
@@ -403,15 +401,6 @@ class UBody_Part(torch.utils.data.Dataset):
smplx_cam_trans = np.array(smplx_param['smplx_param']['trans'])
is_valid_fit = True
"""
# for debug
_tmp = joint_img.copy()
_tmp[:,0] = _tmp[:,0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1]
_tmp[:,1] = _tmp[:,1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0]
_img = img.numpy().transpose(1,2,0)[:,:,::-1] * 255
_img = vis_keypoints(_img, _tmp)
cv2.imwrite('coco_' + str(idx) + '.jpg', _img)
"""
# reverse ra
smplx_joint_cam_wo_ra = smplx_joint_cam.copy()
smplx_joint_cam_wo_ra[smpl_x.joint_part['lhand'], :] = smplx_joint_cam_wo_ra[smpl_x.joint_part['lhand'], :] \
@@ -514,25 +503,14 @@ class UBody_Part(torch.utils.data.Dataset):
# smplx coordinates and parameters
smplx_param = data['smplx_param']
# if str(data['ann_id'])=='184516':
# print(data['ann_id'], smplx_param)
if smplx_param is not None:
smplx_joint_img, smplx_joint_cam, smplx_joint_trunc, smplx_pose, smplx_shape, smplx_expr, smplx_pose_valid, smplx_joint_valid, smplx_expr_valid, smplx_mesh_cam_orig = process_human_model_output(
smplx_param['smplx_param'], smplx_param['cam_param'], do_flip, img_shape, img2bb_trans, rot,
'smplx')
is_valid_fit = True
smplx_cam_trans = np.array(smplx_param['smplx_param']['trans'])
# if str(data['ann_id']) == '184516':
# print(data['ann_id'], smplx_pose)
"""
# for debug
_tmp = joint_img.copy()
_tmp[:,0] = _tmp[:,0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1]
_tmp[:,1] = _tmp[:,1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0]
_img = img.numpy().transpose(1,2,0)[:,:,::-1] * 255
_img = vis_keypoints(_img, _tmp)
cv2.imwrite('coco_' + str(idx) + '.jpg', _img)
"""
# reverse ra
smplx_joint_cam_wo_ra = smplx_joint_cam.copy()
smplx_joint_cam_wo_ra[smpl_x.joint_part['lhand'], :] = smplx_joint_cam_wo_ra[smpl_x.joint_part['lhand'], :] \
@@ -755,15 +733,6 @@ class UBody(Dataset):
smplx_cam_trans = np.array(smplx_param['smplx_param']['trans'])
is_valid_fit = True
"""
# for debug
_tmp = joint_img.copy()
_tmp[:,0] = _tmp[:,0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1]
_tmp[:,1] = _tmp[:,1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0]
_img = img.numpy().transpose(1,2,0)[:,:,::-1] * 255
_img = vis_keypoints(_img, _tmp)
cv2.imwrite('coco_' + str(idx) + '.jpg', _img)
"""
# reverse ra
smplx_joint_cam_wo_ra = smplx_joint_cam.copy()
smplx_joint_cam_wo_ra[smpl_x.joint_part['lhand'], :] = smplx_joint_cam_wo_ra[smpl_x.joint_part['lhand'], :] \
@@ -1026,45 +995,10 @@ class UBody(Dataset):
if len(pa_mpjpe_hand)>0:
eval_result['pa_mpjpe_hand'].append(np.mean(pa_mpjpe_hand))
# data_dict = {}
# data_dict['mpvpe_all'] = eval_result['mpvpe_all'][-1]
# data_dict['mpvpe_hand'] = eval_result['mpvpe_hand'][-1]
# data_dict['mpvpe_face'] = eval_result['mpvpe_face'][-1]
# data_dict['mesh'] = mesh_out
# data_dict['mesh_gt'] = mesh_gt
vis = cfg.vis
save_folder = cfg.vis_dir
data_folder = os.path.join(cfg.root_dir, 'dataset', 'UBody', 'images')
if vis:
# from common.utils.vis import vis_keypoints, vis_mesh, save_obj, render_mesh
# img_path = annot['img_path']
# render_img_save_path = img_path.replace(data_folder, f'{save_folder}/render/')
# if os.path.exists(render_img_save_path):
# img = load_img(render_img_save_path)[:, :, ::-1]
# else:
# img = load_img(img_path)[:, :, ::-1]
# ''' for debug
# kpt_path = render_img_save_path.replace('/render/', '/keypoints/')
# kpt_img = img.copy()
# kpt_img = vis_keypoints(kpt_img, joint_proj)
# # kpt_img = vis_keypoints(kpt_img, mesh_gt_proj)
# os.makedirs(os.path.dirname(kpt_path), exist_ok=True)
# cv2.imwrite(kpt_path, kpt_img)
# '''
# bbox = annot['bbox']
# focal = list(cfg.focal)
# princpt = list(cfg.princpt)
# focal[0] = focal[0] / cfg.input_body_shape[1] * bbox[2]
# focal[1] = focal[1] / cfg.input_body_shape[0] * bbox[3]
# princpt[0] = princpt[0] / cfg.input_body_shape[1] * bbox[2] + bbox[0]
# princpt[1] = princpt[1] / cfg.input_body_shape[0] * bbox[3] + bbox[1]
# img = render_mesh(img, out['smplx_mesh_cam'], smpl_x.face, {'focal': focal, 'princpt': princpt})
# os.makedirs(os.path.dirname(render_img_save_path), exist_ok=True)
# cv2.imwrite(render_img_save_path, img)
vis_save_dir = cfg.vis_dir
rel_img_path = img_path.split('..')[-1]
smplx_pred = {}
@@ -1080,8 +1014,6 @@ class UBody(Dataset):
smplx_pred['transl'] = out['gt_smplx_transl'].reshape(-1,3)
smplx_pred['img_path'] = rel_img_path
# import pdb; pdb.set_trace()
npz_path = os.path.join(cfg.vis_dir, f'{cur_sample_idx + n}.npz')
np.savez(npz_path, **smplx_pred)
data/humandata.py
+2 -5
View File
@@ -287,7 +287,6 @@ class HumanDataset(torch.utils.data.Dataset):
smplx_param = {k: v[i] for k, v in smplx.items()}
# TODO: set invalid if None?
smplx_param['root_pose'] = smplx_param.pop('global_orient', None)
smplx_param['shape'] = smplx_param.pop('betas', None)
smplx_param['trans'] = smplx_param.pop('transl', np.zeros(3))
@@ -298,9 +297,8 @@ class HumanDataset(torch.utils.data.Dataset):
# TODO do not fix betas, give up shape supervision
if 'betas_neutral' in smplx_param:
smplx_param['shape'] = smplx_param.pop('betas_neutral')
# smplx_param['shape'] = np.zeros(10, dtype=np.float32)
# # TODO fix shape of poses
# TODO fix shape of poses
if self.__class__.__name__ == 'Talkshow':
smplx_param['body_pose'] = smplx_param['body_pose'].reshape(21, 3)
smplx_param['lhand_pose'] = smplx_param['lhand_pose'].reshape(15, 3)
@@ -466,7 +464,7 @@ class HumanDataset(torch.utils.data.Dataset):
targets = {'joint_img': joint_img_aug, # keypoints2d
'smplx_joint_img': joint_img_aug, #smplx_joint_img, # projected smplx if valid cam_param, else same as keypoints2d
'joint_cam': joint_cam_wo_ra, # joint_cam actually not used in any loss, # raw kps3d probably without ra
'smplx_joint_cam': smplx_joint_cam if (dummy_cord or getattr(cfg, 'debug', False)) else joint_cam_ra, # kps3d with body, face, hand ra
'smplx_joint_cam': smplx_joint_cam if dummy_cord else joint_cam_ra, # kps3d with body, face, hand ra
'smplx_pose': smplx_pose,
'smplx_shape': smplx_shape,
'smplx_expr': smplx_expr,
@@ -680,7 +678,6 @@ class HumanDataset(torch.utils.data.Dataset):
smplx_pred['transl'] = out['gt_smplx_transl'].reshape(-1,3)
smplx_pred['img_path'] = rel_img_path
# import pdb; pdb.set_trace()
npz_path = os.path.join(cfg.vis_dir, f'{self.save_idx}.npz')
np.savez(npz_path, **smplx_pred)
main/SMPLer_X.py
+3 -67
View File
@@ -62,9 +62,8 @@ class Model(nn.Module):
param_net = param_bb + param_neck + param_head
print('#parameters:')
print(f'{param_bb}, {param_neck}, {param_head}, {param_net}')
# import pdb; pdb.set_trace()
# print('#parameters:')
# print(f'{param_bb}, {param_neck}, {param_head}, {param_net}')
def get_camera_trans(self, cam_param):
# camera translation
@@ -242,12 +241,7 @@ class Model(nn.Module):
loss['smplx_orient'] = self.param_loss(pose, targets['smplx_pose'], meta_info['smplx_pose_valid'])[:, :3] * smplx_orient_weight
loss['smplx_pose'] = self.param_loss(pose, targets['smplx_pose'], meta_info['smplx_pose_valid']) * smplx_pose_weight
### debug
# import numpy as np
# check = torch.isnan(loss['smplx_pose']).cpu()
# pause = np.any(check.numpy())
# if pause:
# import pdb; pdb.set_trace()
else:
loss['smplx_pose'] = self.param_loss(pose, targets['smplx_pose'], meta_info['smplx_pose_valid'])[:, 3:] * smplx_pose_weight
@@ -355,55 +349,6 @@ class Model(nn.Module):
loss['smplx_joint_img'] = self.coord_loss(joint_img, smpl_x.reduce_joint_set(targets['smplx_joint_img']),
smpl_x.reduce_joint_set(meta_info['smplx_joint_trunc'])) * net_kps_2d_weight
### save vis for keypoints checking
# import pdb; pdb.set_trace()
if getattr(cfg, 'debug', False):
import numpy as np
import cv2
out = {}
datalist = cfg.trainset_humandata + cfg.trainset_2d + cfg.trainset_3d
dataset = datalist[0]
out['img'] = inputs['img']
np.save(f'./vis/train_{dataset}.npy', targets)
np.save(f'./vis/train_{dataset}_out.npy', out)
for key in ['joint_cam', 'smplx_joint_cam']:
to_save = targets[key].cpu().detach().numpy()
np.save(f'./vis/train_{dataset}_{key}.npy', to_save)
for sample_id in range(5):
print(sample_id)
vis_joint_img = targets['original_joint_img'][sample_id,...].cpu().detach().numpy()
vis_smplx_joint_img = targets['original_smplx_joint_img'][sample_id,...].cpu().detach().numpy()
# get image
image = out['img'][sample_id].cpu().numpy().transpose(1, 2, 0) * 255
image = image.astype(np.uint8).copy()
image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
color = [(0, 0, 255), (0, 255, 0)] # R: pred, G: gt
for set_id, joint_proj in enumerate([vis_joint_img, vis_smplx_joint_img]): #, joint_2d_gt]):
th = 3 - set_id
# restore kps
joint_proj[:, 0] = joint_proj[:, 0] / cfg.output_hm_shape[2] * cfg.input_img_shape[1] #
joint_proj[:, 1] = joint_proj[:, 1] / cfg.output_hm_shape[1] * cfg.input_img_shape[0] #
# R: pred, G: gt
n_kps = joint_proj.shape[0]
for i in range (n_kps):
# import pdb; pdb.set_trace()
kps = joint_proj[i]
# import pdb;pdb.set_trace()
image = cv2.circle(image, (int(kps[0]),int(kps[1])), radius=th, color=color[set_id], thickness=th)
# image = cv2.rectangle(image, (int(lhand_bbox[0]),int(lhand_bbox[1])),
# (int(lhand_bbox[2]),int(lhand_bbox[3])), (0, 255, 0), thickness=th)
# image = cv2.rectangle(image, (int(rhand_bbox[0]),int(rhand_bbox[1])),
# (int(rhand_bbox[2]),int(rhand_bbox[3])), (0, 0, 255), thickness=th)
# image = cv2.rectangle(image, (int(face_bbox[0]),int(face_bbox[1])),
# (int(face_bbox[2]),int(face_bbox[3])), (255, 0, 0), thickness=th)
cv2.imwrite(f'./vis/joint_img/{dataset}_{sample_id}.jpg',image)
import pdb;pdb.set_trace()
return loss
else:
@@ -459,15 +404,6 @@ class Model(nn.Module):
if 'gt_smplx_transl' in meta_info:
out['gt_smplx_transl'] = meta_info['gt_smplx_transl']
### save result for vis and debug
# import numpy as np
# # np.save('./vis/val_exp38_wo_bbox_sup_out.npy', out)
# for key in ['joint_cam', 'smplx_joint_cam']:
# to_save = targets[key].cpu().detach().numpy()
# np.save(f'./vis/val_0517_{key}.npy', to_save)
# import pdb;pdb.set_trace()
return out
def init_weights(m):