EasyMocap/easymocap/assignment/criterion.py

111 lines
4.0 KiB
Python
Raw Normal View History

'''
@ Date: 2021-05-28 16:36:45
@ Author: Qing Shuai
@ LastEditors: Qing Shuai
2021-06-25 15:54:54 +08:00
@ LastEditTime: 2021-06-25 11:48:57
@ FilePath: /EasyMocapRelease/easymocap/assignment/criterion.py
'''
import numpy as np
class BaseCrit:
def __init__(self, min_conf, min_joints=3) -> None:
self.min_conf = min_conf
self.min_joints = min_joints
self.name = self.__class__.__name__
def __call__(self, keypoints3d, **kwargs):
# keypoints3d: (N, 4)
conf = keypoints3d[..., -1]
conf[conf<self.min_conf] = 0
idx = keypoints3d[..., -1] > self.min_conf
return len(idx) > self.min_joints
class CritWithTorso(BaseCrit):
def __init__(self, torso_idx, min_conf, **kwargs) -> None:
super().__init__(min_conf)
self.idx = torso_idx
self.min_conf = min_conf
def __call__(self, keypoints3d, **kwargs) -> bool:
self.log = '{}'.format(keypoints3d[self.idx, -1])
return (keypoints3d[self.idx, -1] > self.min_conf).all()
class CritLenTorso(BaseCrit):
def __init__(self, src, dst, min_torso_length, max_torso_length, min_conf) -> None:
super().__init__(min_conf)
self.src = src
self.dst = dst
self.min_torso_length = min_torso_length
self.max_torso_length = max_torso_length
def __call__(self, keypoints3d, **kwargs):
"""length of torso"""
# eps = 0.1
# MIN_TORSO_LENGTH = 0.3
# MAX_TORSO_LENGTH = 0.8
if (keypoints3d[[self.src, self.dst], -1] < self.min_conf).all():
# low confidence, skip
return True
length = np.linalg.norm(keypoints3d[self.dst] - keypoints3d[self.src])
self.log = '{}: {:.3f}'.format(self.name, length)
if length < self.min_torso_length or length > self.max_torso_length:
return False
return True
class CritRange(BaseCrit):
def __init__(self, minr, maxr, rate_inlier, min_conf) -> None:
super().__init__(min_conf)
self.min = minr
self.max = maxr
self.rate = rate_inlier
def __call__(self, keypoints3d, **kwargs):
idx = keypoints3d[..., -1] > self.min_conf
k3d = keypoints3d[idx, :3]
crit = (k3d[:, 0] > self.min[0]) & (k3d[:, 0] < self.max[0]) &\
(k3d[:, 1] > self.min[1]) & (k3d[:, 1] < self.max[1]) &\
(k3d[:, 2] > self.min[2]) & (k3d[:, 2] < self.max[2])
self.log = '{}: {}'.format(self.name, k3d)
return crit.sum()/crit.shape[0] > self.rate
class CritMinMax(BaseCrit):
def __init__(self, max_human_length, min_conf) -> None:
super().__init__(min_conf)
self.max_human_length = max_human_length
def __call__(self, keypoints3d, **kwargs):
idx = keypoints3d[..., -1] > self.min_conf
k3d = keypoints3d[idx, :3]
mink = np.min(k3d, axis=0)
maxk = np.max(k3d, axis=0)
length = max(np.abs(maxk - mink))
self.log = '{}: {:.3f}'.format(self.name, length)
2021-06-25 15:54:54 +08:00
return length < self.max_human_length
class CritLimbLength(BaseCrit):
def __init__(self, body_type, max_rate, min_conf) -> None:
super().__init__(min_conf)
self.body_type = body_type
self.max_rate= max_rate
from ..dataset.config import CONFIG
config = CONFIG[body_type]
self.skeleton = config['skeleton']
def __call__(self, keypoints3d, **kwargs):
valid = True
for (i, j), info in self.skeleton.items():
if keypoints3d[i, 3] < self.min_conf or keypoints3d[j, 3] < self.min_conf:
continue
l_mean = info['mean']
l_est = np.linalg.norm(keypoints3d[i, :3] - keypoints3d[j, :3])
if l_mean > 0.15: # 超过十五厘米的 用均值判断
l_std = info['std']
rate = abs(l_est - l_mean)/l_mean
if rate > self.max_rate:
valid = False
break
else:
if l_est > 0.3:
valid = False
break
return valid