"""
This file contains functions that are used to perform data augmentation.
"""
import torch
import numpy as np
from skimage.transform import rotate, resize
import cv2
from torchvision.transforms import Normalize, ToTensor, Compose
from lib.core import constants
def get_normalization():
normalize_img = Compose([ToTensor(),
Normalize(mean=constants.IMG_NORM_MEAN,
std=constants.IMG_NORM_STD)
])
return normalize_img
def get_transform(center, scale, res, rot=0):
"""Generate transformation matrix."""
h = 200 * scale + 1e-6
t = np.zeros((3, 3))
t[0, 0] = float(res[1]) / h
t[1, 1] = float(res[0]) / h
t[0, 2] = res[1] * (-float(center[0]) / h + .5)
t[1, 2] = res[0] * (-float(center[1]) / h + .5)
t[2, 2] = 1
if not rot == 0:
rot = -rot # To match direction of rotation from cropping
rot_mat = np.zeros((3,3))
rot_rad = rot * np.pi / 180
sn,cs = np.sin(rot_rad), np.cos(rot_rad)
rot_mat[0,:2] = [cs, -sn]
rot_mat[1,:2] = [sn, cs]
rot_mat[2,2] = 1
# Need to rotate around center
t_mat = np.eye(3)
t_mat[0,2] = -res[1]/2
t_mat[1,2] = -res[0]/2
t_inv = t_mat.copy()
t_inv[:2,2] *= -1
t = np.dot(t_inv,np.dot(rot_mat,np.dot(t_mat,t)))
return t
def transform(pt, center, scale, res, invert=0, rot=0, asint=True):
"""Transform pixel location to different reference."""
t = get_transform(center, scale, res, rot=rot)
if invert:
t = np.linalg.inv(t)
new_pt = np.array([pt[0]-1, pt[1]-1, 1.]).T
new_pt = np.dot(t, new_pt)
if asint:
return new_pt[:2].astype(int)+1
else:
return new_pt[:2]+1
def transform_pts(pts, center, scale, res, invert=0, rot=0, asint=True):
"""Transform pixel location to different reference."""
t = get_transform(center, scale, res, rot=rot)
if invert:
t = np.linalg.inv(t)
pts = np.concatenate((pts, np.ones_like(pts)[:, [0]]), axis=-1)
new_pt = pts.T
new_pt = np.dot(t, new_pt)
if asint:
return new_pt[:2, :].T.astype(int)
else:
return new_pt[:2, :].T
def crop(img, center, scale, res, rot=0):
"""Crop image according to the supplied bounding box."""
# Upper left point
ul = np.array(transform([1, 1], center, scale, res, invert=1))-1
# Bottom right point
br = np.array(transform([res[0]+1,
res[1]+1], center, scale, res, invert=1))-1
# Padding so that when rotated proper amount of context is included
pad = int(np.linalg.norm(br - ul) / 2 - float(br[1] - ul[1]) / 2)
if not rot == 0:
ul -= pad
br += pad
new_shape = [br[1] - ul[1], br[0] - ul[0]]
if len(img.shape) > 2:
new_shape += [img.shape[2]]
new_img = np.zeros(new_shape)
# Range to fill new array
new_x = max(0, -ul[0]), min(br[0], len(img[0])) - ul[0]
new_y = max(0, -ul[1]), min(br[1], len(img)) - ul[1]
# Range to sample from original image
old_x = max(0, ul[0]), min(len(img[0]), br[0])
old_y = max(0, ul[1]), min(len(img), br[1])
try:
new_img[new_y[0]:new_y[1], new_x[0]:new_x[1]] = img[old_y[0]:old_y[1],
old_x[0]:old_x[1]]
except:
print("invlid bbox, fill with 0")
if not rot == 0:
# Remove padding
new_img = rotate(new_img, rot)
new_img = new_img[pad:-pad, pad:-pad]
new_img = resize(new_img, res)
return new_img
def crop_j2d(j2d, center, scale, res, rot=0):
"""Crop image according to the supplied bounding box."""
# Upper left point
# crop_j2d = np.array(transform_pts(j2d, center, scale, res, invert=0))
b = scale * 200
points2d = j2d - (center - b/2)
points2d = points2d * (res[0] / b)
return points2d
def crop_crop(img, center, scale, res, rot=0):
"""Crop image according to the supplied bounding box."""
# Upper left point
ul = np.array(transform([1, 1], center, scale, res, invert=1))-1
# Bottom right point
br = np.array(transform([res[0]+1,
res[1]+1], center, scale, res, invert=1))-1
# Padding so that when rotated proper amount of context is included
pad = int(np.linalg.norm(br - ul) / 2 - float(br[1] - ul[1]) / 2)
if not rot == 0:
ul -= pad
br += pad
new_shape = [br[1] - ul[1], br[0] - ul[0]]
if len(img.shape) > 2:
new_shape += [img.shape[2]]
new_img = np.zeros(new_shape)
if new_img.shape[0] > img.shape[0]:
p = (new_img.shape[0] - img.shape[0]) / 2
p = int(p)
new_img = cv2.copyMakeBorder(img, p, p, p, p, cv2.BORDER_REPLICATE)
# Range to fill new array
new_x = max(0, -ul[0]), min(br[0], len(img[0])) - ul[0]
new_y = max(0, -ul[1]), min(br[1], len(img)) - ul[1]
# Range to sample from original image
old_x = max(0, ul[0]), min(len(img[0]), br[0])
old_y = max(0, ul[1]), min(len(img), br[1])
new_img[new_y[0]:new_y[1], new_x[0]:new_x[1]] = img[old_y[0]:old_y[1],
old_x[0]:old_x[1]]
if not rot == 0:
# Remove padding
new_img = rotate(new_img, rot)
new_img = new_img[pad:-pad, pad:-pad]
new_img = resize(new_img, res)
return new_img
def uncrop(img, center, scale, orig_shape, rot=0, is_rgb=True):
"""'Undo' the image cropping/resizing.
This function is used when evaluating mask/part segmentation.
"""
res = img.shape[:2]
# Upper left point
ul = np.array(transform([1, 1], center, scale, res, invert=1))-1
# Bottom right point
br = np.array(transform([res[0]+1,res[1]+1], center, scale, res, invert=1))-1
# size of cropped image
crop_shape = [br[1] - ul[1], br[0] - ul[0]]
new_shape = [br[1] - ul[1], br[0] - ul[0]]
if len(img.shape) > 2:
new_shape += [img.shape[2]]
new_img = np.zeros(orig_shape, dtype=np.uint8)
# Range to fill new array
new_x = max(0, -ul[0]), min(br[0], orig_shape[1]) - ul[0]
new_y = max(0, -ul[1]), min(br[1], orig_shape[0]) - ul[1]
# Range to sample from original image
old_x = max(0, ul[0]), min(orig_shape[1], br[0])
old_y = max(0, ul[1]), min(orig_shape[0], br[1])
img = resize(img, crop_shape, interp='nearest')
new_img[old_y[0]:old_y[1], old_x[0]:old_x[1]] = img[new_y[0]:new_y[1], new_x[0]:new_x[1]]
return new_img
def rot_aa(aa, rot):
"""Rotate axis angle parameters."""
# pose parameters
R = np.array([[np.cos(np.deg2rad(-rot)), -np.sin(np.deg2rad(-rot)), 0],
[np.sin(np.deg2rad(-rot)), np.cos(np.deg2rad(-rot)), 0],
[0, 0, 1]])
# find the rotation of the body in camera frame
per_rdg, _ = cv2.Rodrigues(aa)
# apply the global rotation to the global orientation
resrot, _ = cv2.Rodrigues(np.dot(R,per_rdg))
aa = (resrot.T)[0]
return aa
def flip_img(img):
"""Flip rgb images or masks.
channels come last, e.g. (256,256,3).
"""
img = np.fliplr(img)
return img
def flip_kp(kp):
"""Flip keypoints."""
if len(kp) == 24:
flipped_parts = constants.J24_FLIP_PERM
elif len(kp) == 49:
flipped_parts = constants.J49_FLIP_PERM
kp = kp[flipped_parts]
kp[:,0] = - kp[:,0]
return kp
def flip_pose(pose):
"""Flip pose.
The flipping is based on SMPL parameters.
"""
flipped_parts = constants.SMPL_POSE_FLIP_PERM
pose = pose[flipped_parts]
# we also negate the second and the third dimension of the axis-angle
pose[1::3] = -pose[1::3]
pose[2::3] = -pose[2::3]
return pose
def crop_img(img, center, scale, res, val=255):
"""Crop image according to the supplied bounding box."""
# Upper left point
ul = np.array(transform([1, 1], center, scale, res, invert=1))-1
# Bottom right point
br = np.array(transform([res[0]+1,
res[1]+1], center, scale, res, invert=1))-1
new_shape = [br[1] - ul[1], br[0] - ul[0]]
if len(img.shape) > 2:
new_shape += [img.shape[2]]
new_img = np.ones(new_shape) * val
# Range to fill new array
new_x = max(0, -ul[0]), min(br[0], len(img[0])) - ul[0]
new_y = max(0, -ul[1]), min(br[1], len(img)) - ul[1]
# Range to sample from original image
old_x = max(0, ul[0]), min(len(img[0]), br[0])
old_y = max(0, ul[1]), min(len(img), br[1])
new_img[new_y[0]:new_y[1], new_x[0]:new_x[1]] = img[old_y[0]:old_y[1],
old_x[0]:old_x[1]]
new_img = resize(new_img, res)
return new_img
def boxes_2_cs(boxes):
x1, y1, x2, y2 = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
w, h = x2-x1, y2-y1
cx, cy = x1+w/2, y1+h/2
size = np.stack([w, h]).max(axis=0)
centers = np.stack([cx, cy], axis=1)
scales = size / 200
return centers, scales
def box_2_cs(box):
x1,y1,x2,y2 = box[:4].int().tolist()
w, h = x2-x1, y2-y1
cx, cy = x1+w/2, y1+h/2
size = max(w, h)
center = [cx, cy]
scale = size / 200
return center, scale
def est_intrinsics(img_shape):
h, w, c = img_shape
img_center = torch.tensor([w/2., h/2.]).float()
img_focal = torch.tensor(np.sqrt(h**2 + w**2)).float()
return img_center, img_focal