def tbnail_raw(inputs, outputs):
r = 1 / tbnail_scale
imgs = [
cv.resize(rgbg2srgb(imread(x), maximum=65535), None, fx=r, fy=r)
for x in inputs[:4]
]
for img, o in zip(imgs, outputs[:4]):
imwrite(o, img)
imwrite(
outputs[4],
cv.resize(rgbg2linref(imread(inputs[4]), maximum=65535) * 255,
None,
fx=r,
fy=r).astype(np.uint8))
def crop(input_path, output_path, other=1):
img = imread(input_path)
h, w, _ = img.shape
bd = [int(h * bds[0]), int(h * bds[1]), int(w * bds[2]), int(w * bds[3])]
if img is None:
print(bd)
cv.imwrite(output_path, img[bd[0]:bd[1], bd[2]:bd[3]])
def get5rgbc(inputs, outputs, black_level=64):
"""inputs: gt, ab, f
outputs: gt, ab, f, fo, m"""
check_exif(inputs)
rgbs = [raw_read_rgb(x) for x in inputs]
h, w, c = rgbs[0].shape
bd = [int(h * bds[0]), int(h * bds[1]), int(w * bds[2]), int(w * bds[3])]
rgbsc = [x[bd[0]:bd[1], bd[2]:bd[3]] for x in rgbs]
for out, rgb in zip(outputs[:3], rgbsc):
imwrite(out, rgb)
# return
raws = [rawpy.imread(x) for x in inputs[1:3]]
raws[1].raw_image_visible[:] = np.clip(
raws[1].raw_image_visible.astype(np.int64) -
raws[0].raw_image_visible.astype(np.int64) + black_level, 0,
1023).astype(np.uint16)
fo = cv.resize(raws[1].postprocess(use_camera_wb=True,
no_auto_bright=True),
None,
fx=0.5,
fy=0.5)
imwrite(outputs[3], fo[bd[0]:bd[1], bd[2]:bd[3]])
rgbs = [rgbg2rgb(imread(x)) for x in inputs[:2]]
m = (np.clip((rgbs[0].astype(np.float32) - rgbs[1].astype(np.float32)) /
(rgbs[0] + 1), 0, 1) * 255).astype(np.uint8)
imwrite(outputs[4], m[bd[0]:bd[1], bd[2]:bd[3]])
def tbnail_rgb(inputs, outputs):
for i, o in zip(inputs, outputs):
im = imread(i)
if im is None:
print(i)
continue
# print(im.shape,im.dtype)
imwrite(o, cv.resize(im, None, fx=1 / 16, fy=1 / 16))
def tbnail_raw(inputs, outputs):
r = 1 / 16
imgs = [
cv.resize(rgbg2srgb(imread(x), maximum=65535), None, fx=r, fy=r)
for x in inputs[:5]
]
for img, o in zip(imgs, outputs[:5]):
imwrite(o, img)
def get_frame(self,
frame_number,
dtype=np.float32,
sfactor=1.0,
image_type='rgb'):
return io.imread(self.frame_fullnames[frame_number],
dtype=dtype,
sfactor=sfactor,
image_type=image_type)
def _set_correspondences(img1, img2, field_img_path='./demo/data/field.png'):
field_img = io_utils.imread(field_img_path)
h2, w2 = field_img.shape[0:2]
W, H = 104.73, 67.74
fig, ax = plt.subplots(nrows=2, ncols=2)
ax[0, 0].imshow(img1)
ax[1, 0].imshow(img2)
ax[0, 1].imshow(field_img)
ax[1, 1].imshow(field_img)
ax[0, 0].axis('off')
ax[1, 0].axis('off')
ax[0, 1].axis('off')
ax[1, 1].axis('off')
points2d_img1 = []
points2d_img2 = []
points3d = []
def onclick(event):
# print('button=%d, x=%d, y=%d, xdata=%f, ydata=%f' %
# (event.button, event.x, event.y, event.xdata, event.ydata))
x, y = event.xdata, event.ydata
print("Click : ", x, y)
if event.inaxes.axes.get_position(
).x0 < 0.5 and event.inaxes.axes.get_position().y0 > 0.5:
# if event.inaxes in ax[0, 0]:
ax[0, 0].plot(x, y, 'r.', ms=10)
points2d_img1.append([x, y])
elif event.inaxes.axes.get_position(
).x0 < 0.5 and event.inaxes.axes.get_position().y0 < 0.5:
# elif event.inaxes in ax[1, 0]:
ax[1, 0].plot(x, y, 'r.', ms=10)
points2d_img2.append([x, y])
else:
ax[0, 1].plot(x, y, 'b+', ms=10)
ax[1, 1].plot(x, y, 'b+', ms=10)
points3d.append([x, 0, y])
fig.canvas.draw()
cid = fig.canvas.mpl_connect('button_press_event', onclick)
plt.show()
points2d_img1 = np.array(points2d_img1)
points2d_img2 = np.array(points2d_img2)
points3d = np.array(points3d)
# Scale to H, W of soccer field
points3d[:, 0] = ((points3d[:, 0] - w2 / 2.) / w2) * W
points3d[:, 2] = ((points3d[:, 2] - h2 / 2.) / h2) * H
print(points3d, points2d_img1, points2d_img2)
return points2d_img1, points2d_img2, points3d
def get_frame(self,
vid_name,
frame_number,
dtype=np.float32,
image_type='rgb',
sfactor=1.0,
flip=False):
print("[#] Gather Frame : ", vid_name)
return io.imread(self.frame_fullnames[vid_name][frame_number],
dtype=dtype,
sfactor=sfactor,
image_type=image_type,
flip=flip)
def get_mask_from_detectron(self,
vid_name,
frame_number,
dtype=np.float32,
image_type='rgb',
sfactor=1.0,
flip=False):
print("[#] Gather Detectron : ", vid_name)
return io.imread(self.mask_fullnames[vid_name][frame_number],
dtype=dtype,
sfactor=sfactor,
image_type=image_type,
flip=flip)[..., 0]
def get5rawc(inputs, outputs):
"""inputs: gt, ab, f
outputs: gt, ab, f, fo, m"""
raws = [imread(x) for x in inputs]
h, w, _ = raws[0].shape
bd = [int(h * bds[0]), int(h * bds[1]), int(w * bds[2]), int(w * bds[3])]
rawsc = [x[bd[0]:bd[1], bd[2]:bd[3]] for x in raws]
for out, raw in zip(outputs[:3], rawsc):
imwrite(out, raw)
fo = np.clip(rawsc[2].astype(np.int32) - rawsc[1].astype(np.int32), 0,
65535).astype(np.uint16)
imwrite(outputs[3], fo)
m = (np.clip((rawsc[0].astype(np.float32) - rawsc[1].astype(np.float32)) /
(rawsc[0] + 1), 0, 1) * 65535).astype(np.uint16)
imwrite(outputs[4], m)
def _set_correspondences(
img,
field_img_path='/home/krematas/Mountpoints/grail/data/Singleview/Soccer/field.png'
):
field_img = io.imread(field_img_path)
h2, w2 = field_img.shape[0:2]
W, H = 104.73, 67.74
fig, ax = plt.subplots(1, 2)
ax[0].imshow(img)
ax[1].imshow(field_img)
ax[0].axis('off')
ax[1].axis('off')
points2d = []
points3d = []
def onclick(event):
# print('button=%d, x=%d, y=%d, xdata=%f, ydata=%f' %
# (event.button, event.x, event.y, event.xdata, event.ydata))
x, y = event.xdata, event.ydata
if event.inaxes.axes.get_position().x0 < 0.5:
ax[0].plot(x, y, 'r.', ms=10)
points2d.append([x, y])
else:
ax[1].plot(x, y, 'b+', ms=10)
points3d.append([x, 0, y])
plt.show()
cid = fig.canvas.mpl_connect('button_press_event', onclick)
plt.show()
points2d = np.array(points2d)
points3d = np.array(points3d)
points3d[:, 0] = ((points3d[:, 0] - w2 / 2.) / w2) * W
points3d[:, 2] = ((points3d[:, 2] - h2 / 2.) / h2) * H
return points2d, points3d
def get5rawc(inputs, outputs, bds):
"""inputs: gt, ab, f
outputs: gt, ab, f, fo, m"""
raws = [imread(x) for x in inputs]
w_start, h_start, w_end, h_end = [int(x) for x in bds[:4]]
# print(h_start, w_start, h_end, w_end)
h_offset = (h_end - h_start) // 32 * 32
w_offset = (w_end - w_start) // 32 * 32
rawsc = [
x[h_start:h_start + h_offset, w_start:w_start + w_offset] for x in raws
]
for out, raw in zip(outputs[:3], rawsc):
imwrite(out, raw)
fo = np.maximum(rawsc[2].astype(np.int32) - rawsc[1].astype(np.int32),
0).astype(np.uint16)
imwrite(outputs[3], fo)
tran = np.maximum(rawsc[1].astype(np.int32) - rawsc[0].astype(np.int32),
0).astype(np.uint16)
imwrite(outputs[4], tran)
def get_mask_from_detectron(self, frame_number):
return io.imread(
join(self.path_to_dataset, 'detectron',
self.frame_basenames[frame_number] + '.png'))[:, :, 0]
field = gloo.Program(vertex_tex, fragment_tex)
field.bind(vertices_field)
field['position'] = vertices_field
field['u_texture'] = data.get(join(opt.path_to_data, 'texture.png'))
field['transform'] = trackball
all_programs = []
for fid, fname in enumerate(filenames):
(basename, ext) = file_utils.extract_basename(fname)
print('Loading model {0}/{1}: {2}'.format(fid, len(filenames), basename))
path_to_pc = join(opt.path_to_data, 'scene3d')
img = io.imread(join(path_to_pc, '{0}.jpg'.format(basename)),
dtype=np.float32)
vertices, indices = objload(join(path_to_pc, '{0}.obj'.format(basename)))
vertices['texcoord'][:, 1] = 1.0 - vertices['texcoord'][:, 1]
tex_program = gloo.Program(vertex_tex, fragment_tex)
tex_program.bind(vertices)
tex_program['u_texture'] = img
tex_program['transform'] = trackball
all_programs.append(tex_program)
trackball.theta, trackball.phi, trackball.zoom = -10, 0, 15
window = app.Window(width=512, height=512, color=(0.30, 0.30, 0.35, 1.00))
description='Track camera given an initial estimate')
parser.add_argument('--dataset', default='kth-0', help='Dataset folder')
parser.add_argument('--frame',
type=int,
default=0,
help='Specific frame to estimate camera parameters')
opt, _ = parser.parse_known_args()
path_to_data = file_utils.get_platform_datadir(
join('Singleview/Soccer/', opt.dataset))
path_to_labels = join(path_to_data, 'cnn', 'youtube', 'labels')
path_to_masks = join(path_to_data, 'cnn', 'youtube', 'masks')
label_files = listdir(path_to_masks)
label_files = [i for i in label_files if '_dcrf_2.png' in i]
label_files.sort()
n_files = len(label_files)
for i in range(n_files):
mask = io.imread(join(path_to_masks, label_files[i])) / 255.0
h, w = mask.shape[:2]
lbl = np.zeros((2, h, w), dtype=np.float32)
lbl[0, :, :] = mask[:, :, 0]
lbl[1, :, :] = mask[:, :, 0]
savename = label_files[i].replace('_dcrf_2.png', '_r.npy')
np.save(join(path_to_labels, savename), lbl)
cam = cam_utils.Camera('camera0', db[0].calib['00118']['A'],
db[0].calib['00118']['R'], db[0].calib['00118']['T'],
db[0].shape[0], db[0].shape[1])
# initialize figure
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
# camera0
x = cam.get_position().item(0, 0)
y = cam.get_position().item(1, 0)
z = cam.get_position().item(2, 0)
ax.scatter(x, y, z, label='camera0', marker='o')
# demo field for x-y plane
field = io.imread('./demo/data/demo_field.png')
plot_field()
# set axis names and position
ax.set_xlabel('X', fontsize=25, rotation=0)
ax.set_ylabel('Y', fontsize=25, rotation=0)
ax.set_zlabel('Z', fontsize=25, rotation=0)
ax.xaxis._axinfo['juggled'] = (0, 0, 0)
ax.yaxis._axinfo['juggled'] = (1, 1, 1)
ax.zaxis._axinfo['juggled'] = (2, 2, 2)
# scale
ax.set_xlim3d(-100, 100)
ax.set_ylim3d(0, 200)
ax.set_zlim3d(-100, 100)
db.refine_poses(keypoint_thresh=7, score_thresh=0.4, neck_thresh=0.4)
file_utils.mkdir(os.path.join(db.path_to_dataset, 'players', 'masks'))
file_utils.mkdir(os.path.join(db.path_to_dataset, 'players', 'labels'))
# Merge masks for players
for sel_frame in tqdm(range(db.n_frames)):
basename = db.frame_basenames[sel_frame]
poses = db.poses[basename]
for i in range(len(poses)):
fname = os.path.join(
db.path_to_dataset, 'players', 'cnn_masks',
'{0}_{1:05d}.png'.format(db.frame_basenames[sel_frame], i))
cnn_mask = io.imread(fname)
fname = os.path.join(
db.path_to_dataset, 'players', 'pose_masks',
'{0}_{1:05d}.png'.format(db.frame_basenames[sel_frame], i))
pose_mask = io.imread(fname)
if len(cnn_mask.shape) == 1:
cnn_mask = np.ones_like(pose_mask)
mask = cnn_mask[:, :, 0] * pose_mask[:, :, 0] * 255
cv2.imwrite(
os.path.join(
db.path_to_dataset, 'players', 'masks',
'{0}_{1:05d}.png'.format(db.frame_basenames[sel_frame], i)),
mask)
label = {'mask': mask, 'depth': mask, 'billboard': mask}
