"""
Copyright (c) Facebook, Inc. and its affiliates.
This source code is licensed under the MIT license found in the
LICENSE file in the root directory of this source tree.
"""
import numpy as np
import io
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import axes3d
from matplotlib import cm
from PIL import Image
from visdom import Visdom
from tools.utils import NumpySeedFix
# the visdom connection handle
viz = None
def get_visdom_env(cfg):
if len(cfg.visdom_env) == 0:
visdom_env = cfg.exp_dir
else:
visdom_env = cfg.visdom_env
return visdom_env
def get_visdom_connection(server='http://localhost', port=8097):
global viz
if viz is None:
viz = Visdom(server=server, port=port)
return viz
def denorm_image_trivial(im):
im = im - im.min()
im = im / (im.max()+1e-7)
return im
def ensure_im_width(img, basewidth):
wpercent = (basewidth/float(img.size[0]))
hsize = int((float(img.size[1])*float(wpercent)))
img = img.resize((basewidth, hsize), Image.ANTIALIAS)
return img
def fig2data(fig, size=None):
"""
Convert a Matplotlib figure to a numpy array
"""
buf = io.BytesIO()
plt.savefig(buf, format='png', bbox_inches='tight', pad_inches=0)
buf.seek(0)
im = Image.open(buf).convert('RGB')
if size:
im = im.resize(size)
return np.array(im)
def show_projections(p,
visdom_env=None,
visdom_win=None,
v=None,
image_path=None,
image=None,
title='projs',
cmap__='gist_ncar',
markersize=None,
sticks=None,
stickwidth=2,
stick_color=None,
plot_point_order=False,
bbox=None,
):
if image is None:
try:
im = Image.open(image_path).convert('RGB')
im = np.array(im).transpose(2, 0, 1)
except:
im = None
print('!cant load image %s' % image_path)
else:
im = image
nkp = int(p.shape[2])
pid = np.linspace(0., 1., nkp)
if v is not None:
okp = v > 0
else:
okp = np.ones(nkp) == 1
possible_markers = ['o', 'x', 'd']
markers = [possible_markers[i % len(possible_markers)]
for i in range(len(p))]
if markersize is None:
msz = 50
if nkp > 40:
msz = 5
markersizes = [msz]*nkp
else:
markersizes = [markersize]*nkp
fig = plt.figure(figsize=[11, 11])
if im is not None:
plt.imshow(im.transpose((1, 2, 0)))
plt.axis('off')
if sticks is not None:
if stick_color is not None:
linecol = stick_color
else:
linecol = [0., 0., 0.]
for p_ in p:
for stick in sticks:
if v is not None:
if v[stick[0]] > 0 and v[stick[1]] > 0:
linestyle = '-'
else:
continue
else:
linestyle = '-'
plt.plot(p_[0, stick], p_[1, stick], linestyle,
color=linecol, linewidth=stickwidth, zorder=1)
for p_, marker, msz in zip(p, markers, markersizes):
plt.scatter(p_[0, okp], p_[1, okp], msz, pid[okp],
cmap=cmap__, linewidths=2, marker=marker, zorder=2,
vmin=0., vmax=1.)
if plot_point_order:
for ii in np.where(okp)[0]:
plt.text(p_[0, ii], p_[1, ii], '%d' %
ii, fontsize=int(msz*0.25))
if bbox is not None:
import matplotlib.patches as patches
# Create a Rectangle patch
rect = patches.Rectangle((bbox[0], bbox[1]), bbox[2], bbox[3],
linewidth=1, edgecolor='r', facecolor='none')
plt.gca().add_patch(rect)
if im is None:
plt.gca().invert_yaxis()
plt.axis('equal')
plt.gca().axes.get_xaxis().set_visible(False)
plt.gca().axes.get_yaxis().set_visible(False)
# plt.gca().set_frame_on(False)
plt.gca().set_axis_off()
else: # remove all margins
plt.gca().axes.get_xaxis().set_visible(False)
plt.gca().axes.get_yaxis().set_visible(False)
plt.gca().set_frame_on(False)
plt.gca().set_axis_off()
# return fig
improj = np.array(fig2data(fig))
if visdom_env is not None:
viz = get_visdom_connection()
viz.image(np.array(improj).transpose(2, 0, 1),
env=visdom_env, opts={'title': title}, win=visdom_win)
plt.close(fig)
return improj
def extend_to_3d_skeleton_simple(ptcloud, sticks, line_resol=10, rgb=None):
H36M_TO_MPII_PERM = [3, 2, 1, 4, 5, 6,
0, 8, 9, 10, 16, 15, 14, 11, 12, 13]
rgb_now = rgb.T if rgb is not None else None
ptcloud_now = ptcloud.T
ptcloud = ptcloud.T
rgb = rgb.T if rgb is not None else rgb
if ptcloud_now.shape[1] == 16: # MPII
sticks_new = []
for stick in sticks:
if stick[0] in H36M_TO_MPII_PERM and stick[1] in H36M_TO_MPII_PERM:
s1 = H36M_TO_MPII_PERM.index(int(stick[0]))
s2 = H36M_TO_MPII_PERM.index(int(stick[1]))
sticks_new.append([s1, s2])
sticks = sticks_new
for sticki, stick in enumerate(sticks):
alpha = np.linspace(0, 1, line_resol)[:, None]
linepoints = ptcloud[stick[0], :][None, :] * alpha + \
ptcloud[stick[1], :][None, :] * (1. - alpha)
ptcloud_now = np.concatenate((ptcloud_now, linepoints), axis=0)
if rgb is not None:
linergb = rgb[stick[0], :][None, :] * alpha + \
rgb[stick[1], :][None, :] * (1.-alpha)
rgb_now = np.concatenate(
(rgb_now, linergb.astype(np.int32)), axis=0)
if rgb is not None:
rgb_now = rgb_now.T
return ptcloud_now.T, rgb_now
def visdom_plot_pointclouds(viz, pcl, visdom_env, title,
plot_legend=True, markersize=2,
nmax=5000, sticks=None, win=None):
if sticks is not None:
pcl = {k: extend_to_3d_skeleton_simple(v, sticks)[0]
for k, v in pcl.items()}
legend = list(pcl.keys())
cmap = 'tab10'
npcl = len(pcl)
rgb = (cm.get_cmap(cmap)(np.linspace(0, 1, 10))
[:, :3]*255.).astype(np.int32).T
rgb = np.tile(rgb, (1, int(np.ceil(npcl/10))))[:, 0:npcl]
rgb_cat = {k: np.tile(rgb[:, i:i+1], (1, p.shape[1])) for
i, (k, p) in enumerate(pcl.items())}
rgb_cat = np.concatenate(list(rgb_cat.values()), axis=1)
pcl_cat = np.concatenate(list(pcl.values()), axis=1)
if pcl_cat.shape[1] > nmax:
with NumpySeedFix():
prm = np.random.permutation(
pcl_cat.shape[1])[0:nmax]
pcl_cat = pcl_cat[:, prm]
rgb_cat = rgb_cat[:, prm]
win = viz.scatter(pcl_cat.T, env=visdom_env,
opts={'title': title, 'markersize': markersize,
'markercolor': rgb_cat.T}, win=win)
# legend
if plot_legend:
dummy_vals = np.tile(
np.arange(npcl)[:, None], (1, 2)).astype(np.float32)
title = "%s_%s" % (title, legend)
opts = dict(title=title, legend=legend, width=400, height=400)
viz.line(dummy_vals.T, env=visdom_env, opts=opts)
return win
def matplot_plot_point_cloud(ptcloud, pointsize=20, azim=90, elev=90,
figsize=(8, 8), title=None, sticks=None, lim=None,
cmap='gist_ncar', ax=None, subsample=None,
flip_y=False):
if lim is None:
lim = np.abs(ptcloud).max()
nkp = int(ptcloud.shape[1])
pid = np.linspace(0., 1., nkp)
rgb = (cm.get_cmap(cmap)(pid)[:, :3]*255.).astype(np.int32)
if subsample is not None:
with NumpySeedFix():
prm = np.random.permutation(nkp)[0:subsample]
pid = pid[prm]
rgb = rgb[prm, :]
ptcloud = ptcloud[:, prm]
if flip_y:
ptcloud[1, :] = -ptcloud[1, :]
if ax is not None:
fig = None
else:
fig = plt.figure(figsize=figsize)
ax = fig.add_subplot(111, projection='3d')
ax.view_init(elev=elev, azim=azim)
if sticks is not None:
for stick in sticks:
line = ptcloud[:, [stick[0], stick[1]]]
xs, ys, zs = line
ax.plot(xs, ys, zs, color='black')
xs, ys, zs = ptcloud
ax.scatter(xs, ys, zs, s=pointsize, c=pid, marker='.', cmap=cmap)
ax.set_xlim(-lim, lim)
ax.set_ylim(-lim, lim)
ax.set_zlim(-lim, lim)
ax.set_zticklabels([])
ax.set_yticklabels([])
ax.set_xticklabels([])
plt.axis('off')
if title is not None:
ax.set_title(title)
plt.show()
return fig
# old functions to replace:
def enlarge_box(box, perc, imsz):
boxw, boxh = box[2]-box[0], box[3]-box[1]
box[0] -= boxw*perc
box[1] -= boxh*perc
box[2] += boxw*perc
box[3] += boxh*perc
imh, imw = imsz
box = np.maximum(np.minimum(box, np.array([imw, imh, imw, imh])), 0.)
return box
