def calResiduals(self):
### Compute the affine matrix
self.getAffineT()
self.residuals = np.zeros(self.gtShapes.shape)
num = self.gtShapes.shape[0]
for i in range(num):
# Project to meanshape coordinary
T = self.real2mss[i]
bndBox = self.bndBoxs[i]
err = self.gtShapes[i] - self.initShapes[i]
err = np.divide(err, (bndBox[2], bndBox[3]))
err = Affine.transPntsForwardWithSameT(err, T)
self.residuals[i, :] = err
def genFea(self, train_set, pointIdx, sampleIdxs, feaTypes):
sampleNum = len(sampleIdxs)
feaNum = feaTypes.shape[0]
pdFea = np.zeros((sampleNum, feaNum), dtype=np.float32)
coord_a = np.zeros((feaNum, 2))
coord_b = np.zeros((feaNum, 2))
angle_cos = np.cos(feaTypes[:, [1, 3]])
angle_sin = np.sin(feaTypes[:, [1, 3]])
ms_x_ratio = angle_cos * feaTypes[:, [0, 2]]
ms_y_ratio = angle_sin * feaTypes[:, [0, 2]]
augNum = train_set.augNum
for i, idx in enumerate(sampleIdxs):
T = train_set.ms2reals[idx]
bndBox = train_set.bndBoxs[idx]
imgData = train_set.imgDatas[int(idx / augNum)]
initShape = train_set.initShapes[idx]
# numpy image: H x W x C
imgH, imgW = imgData.shape
# print('imgH:{}, imgW:{}'.format(imgH, imgW))
w, h = bndBox[2:4]
coord_a[:, 0] = ms_x_ratio[:, 0] * w
coord_a[:, 1] = ms_y_ratio[:, 0] * h
coord_b[:, 0] = ms_x_ratio[:, 1] * w
coord_b[:, 1] = ms_y_ratio[:, 1] * h
# convert meanshape coord into real coord
coord_a = Affine.transPntsForwardWithSameT(coord_a, T)
coord_b = Affine.transPntsForwardWithSameT(coord_b, T)
coord_a = coord_a + initShape[pointIdx]
coord_b = coord_b + initShape[pointIdx]
# TODO use other interpolations
coord_a = np.floor(coord_a)
coord_b = np.floor(coord_b)
# print('coord_a shape: {}'.format(coord_a.shape))
# Check with the image size
coord_a[coord_a < 0] = 0
# print('coord_a[:,0]>imgW-1: {}'.format(coord_a[:,0]>imgW-1))
# print('coord_a[:,0]: {}'.format(coord_a[:,0]))
coord_a[coord_a[:, 0] > imgW - 1, 0] = imgW - 1
coord_a[coord_a[:, 1] > imgH - 1, 1] = imgH - 1
coord_b[coord_b < 0] = 0
coord_b[coord_b[:, 0] > imgW - 1, 0] = imgW - 1
coord_b[coord_b[:, 1] > imgH - 1, 1] = imgH - 1
# Construct the idx list for get the elements
idx_a = np.transpose(coord_a).tolist()
idx_a[0], idx_a[1] = idx_a[1], idx_a[0]
idx_b = np.transpose(coord_b).tolist()
# print('idx_a: {}'.format(idx_a))
idx_b[0], idx_b[1] = idx_b[1], idx_b[0]
idx_a = np.array(idx_a, dtype=np.int64)
# print(('idx_aa: {}'.format(idx_a[1])))
idx_b = np.array(idx_b, dtype=np.int64)
### get the diff
# print('i: {}, ind:{}'.format(i, idx))
# print('imgData shape', imgData.shape)
# print('idx_a[0]: {}\n idx_a[1]: {} \n idx_b[0]:{}\n idx_b[1]: {}'.format(idx_a[0], idx_a[1], idx_b[0], idx_b[1]))
# print('idx_a[0] shape: {}'.format(idx_a[0].shape))
pdFea[i, :] = np.subtract(imgData[idx_a[0], idx_a[1]],
imgData[idx_b[0], idx_b[1]],
dtype=np.int64)
return pdFea
