import numpy as np
scale = 70.
sys.path.append("/Users/Alex/Dropbox/Science/Datta lab/Posture Tracking/hsmm-particlefilters/renderer")
path_to_behavior_data = "/Users/Alex/Dropbox/Science/Datta lab/Posture Tracking/Test Data"
from load_data import load_behavior_data
cmds.promptDialog(title="Which Image?", message="Image number:")
index = cmds.promptDialog(query=True, text=True)
index = int(index)
print index+1
images = load_behavior_data(path_to_behavior_data, index+1, "images")
image = images[-1]
width,height = image.shape
mesh_name = "mouse_mesh"
cmds.polyPlane(name=mesh_name,
width=width, height=height,
subdivisionsWidth=width,
subdivisionsHeight=height)
cmds.move(width/2., 0, height/2., relative=True)
for i in range(width*height):
translation = cmds.xform(mesh_name +".vtx[%d]" % (i), query=True, worldSpace=True, t=True)
x,y = translation[0], translation[2]
x,y = int(x), int(y)
from __future__ import division from load_data import load_behavior_data import scipy.ndimage as ndimage from os.path import join, dirname import numpy as np num_frames = 8000 # total number of images in the behavior file num_iterations_dilation = 1 blur_sigma = (2,2) source_behavior_data = "../Test Data/Mouse No Median Filter, No Dilation" dest_behavior_data = "../Test Data/Blurred Edge" path_to_behavior_data = join(dirname(__file__),'..',source_behavior_data) images = load_behavior_data(path_to_behavior_data, num_frames, 'images') out_images = np.zeros_like(images) for i,img in enumerate(images): mask = img == 0 dilation = ndimage.morphology.binary_dilation(mask, iterations=num_iterations_dilation) mask = dilation-mask blurred_img = ndimage.filters.median_filter(img, (5,5)) img[mask] = blurred_img[mask] out_images[i] = img out_images.tofile(join(dest_behavior_data, "Extracted Mouse Images.int16binary"))
def test_single_mouse():
path_to_behavior_data = os.path.join(os.path.dirname(__file__), "..", "Test Data/Blurred Edge")
# which_img = 30
# which_img = 731
which_img = 5
from load_data import load_behavior_data
image = load_behavior_data(path_to_behavior_data, which_img + 1, "images")[-1]
image = image.T[::-1, :].astype("float32")
num_particles = 128 ** 2
numCols = 32
numRows = 32
scenefile = os.path.join(os.path.dirname(__file__), "data/mouse_mesh_low_poly3.npz")
useFramebuffer = False
ms = MouseScene(
scenefile,
mouse_width=80,
mouse_height=80,
scale_width=18.0,
scale_height=200.0,
scale_length=18.0,
numCols=numCols,
numRows=numRows,
useFramebuffer=useFramebuffer,
)
ms.gl_init()
# Figure out the number of passes we'll be making
num_passes = int(num_particles / ms.num_mice)
# Let's fill in our particles
particle_data = np.zeros((num_particles, 8 + ms.num_bones * 3))
# Set the horizontal offsets
position_val = 0
particle_data[1:, :2] = np.random.normal(loc=0, scale=1, size=(num_particles - 1, 2))
# Set the vertical offset
particle_data[1:, 2] = np.random.normal(loc=0.0, scale=3.0, size=(num_particles - 1,))
# Set the angles (yaw and roll)
theta_val = 0
particle_data[1:, 3] = theta_val + np.random.normal(loc=0, scale=3, size=(num_particles - 1,))
particle_data[1:, 4] = np.random.normal(loc=0, scale=0.01, size=(num_particles - 1,))
# Set the scales (width, length, height)
particle_data[0, 5] = np.max(ms.scale_width)
particle_data[0, 6] = np.max(ms.scale_length)
particle_data[0, 7] = np.max(ms.scale_height)
particle_data[1:, 5] = np.random.normal(loc=18, scale=2, size=(num_particles - 1,))
particle_data[1:, 6] = np.random.normal(loc=18, scale=2, size=(num_particles - 1,))
particle_data[1:, 7] = np.abs(np.random.normal(loc=200.0, scale=10, size=(num_particles - 1,)))
# Grab the baseline joint rotations
rot = ms.get_joint_rotations().copy()
rot = np.tile(rot, (num_passes, 1, 1))
particle_data[:, 8::3] = rot[:, :, 0]
particle_data[:, 9::3] = rot[:, :, 1]
particle_data[:, 10::3] = rot[:, :, 2]
# Add noise to the baseline rotations (just the pitch and yaw for now)
# particle_data[1:,8::3] += np.random.normal(scale=20, size=(num_particles-1, ms.num_bones))
particle_data[1:, 9 + 6 :: 3] += np.random.normal(scale=20, size=(num_particles - 1, ms.num_bones - 2))
particle_data[1:, 10::3] += np.random.normal(scale=20, size=(num_particles - 1, ms.num_bones))
likelihoods, posed_mice = ms.get_likelihood(
image, x=position_val, y=position_val, theta=theta_val, particle_data=particle_data, return_posed_mice=True
)
# L = ms.likelihood.T.ravel()
particle_rotations = np.hstack((particle_data[:, 9::3], particle_data[:, 10::3]))
real_rotations = np.hstack((rot[:, :, 1], rot[:, :, 2]))
rotation_diffs = np.sum((particle_rotations - real_rotations) ** 2.0, 1)
figure()
plot(rotation_diffs, likelihoods, ".k")
ylabel("Likelihood")
xlabel("Rotation angle differences")
title("Rotation angle difference versus likelihood")
binrange = (0, 3000)
num_bins = 10
bins = np.linspace(binrange[0], binrange[1], num_bins)
index = np.digitize(rotation_diffs, bins)
means = [np.mean(likelihoods[index == i]) for i in range(num_bins)]
errs = [np.std(likelihoods[index == i]) for i in range(num_bins)]
errorbar(bins, means, yerr=errs, linewidth=2)
# savefig("/Users/Alex/Desktop/angle vs likelihood.png")
# figure(); imshow(ms.likelihood)
# figure(); imshow(ms.data); colorbar()
# figure(); imshow(ms.diffmap); colorbar()
# Find the five best mice
idx = np.argsort(likelihoods)
fivebest = np.hstack(posed_mice[idx[-5:]])
# Show first the raw mouse, then my hand-posed mouse, and then the five best poses
figure()
title("Five best (best, far right)")
imshow(np.hstack((ms.mouse_img, posed_mice[0], fivebest)))
plt.clim(0, 300)
vlines(ms.mouse_width, 0, ms.mouse_height, linewidth=3, color="w")
text(
ms.mouse_width / 2.0,
ms.mouse_width * 0.9,
"Real Mouse",
horizontalalignment="center",
verticalalignment="center",
color="white",
)
vlines(ms.mouse_width * 2, 0, ms.mouse_height, linewidth=3, color="w")
text(
ms.mouse_width * 1.5,
ms.mouse_width * 0.9,
"Unposed Mouse",
horizontalalignment="center",
verticalalignment="center",
color="white",
)
return ms, rotation_diffs, likelihoods, particle_data, posed_mice
