EquiFusion / src /utils /plot.py
ceccurr's picture
EquiFusion demo
e6d357a verified
Raw History Blame Contribute Delete
8.91 kB
import matplotlib.animation as animation
import matplotlib.pyplot as plt
import numpy as np
from functools import partial
from .numpy import get_fig_as_nparray, numpy_img_to_tensor
RADIUS = np.array([1,1,1])*0.025 #in m
RESOPTIONS = { # high res
# lower res
"figsize": 12, "linewidth": 2,
}
def create_pose(ax, plots, vals, limbseq, left_right_limb, pred=True, num_multiple_preds=1, update=False, random_color=False, color=None, **kwargs):
# Start and endpoints of our representation
I = np.array([touple[0] for touple in limbseq])
J = np.array([touple[1] for touple in limbseq])
# Left / right indicator
LR = np.array([left_right_limb[a] or left_right_limb[b] for a, b in limbseq])
if pred:
lcolor = "#D22B2B"
if num_multiple_preds >1:
lcolor = "#0000FF"
if random_color:
import random
lcolor = "#"+''.join([random.choice('ABCDEF0123456789') for i in range(6)])
if color is not None:
lcolor = color
rcolor = lcolor #"#2ecc71"
else:
rcolor = "#383838" # dark grey
lcolor = rcolor #"#8e8e8e"
if random_color:
import random
lcolor = "#"+''.join([random.choice('ABCDEF0123456789') for i in range(6)])
rcolor = lcolor
elif color is not None:
lcolor = color
rcolor = lcolor
rang = len(I)*num_multiple_preds if pred else 0
for i in np.arange(len(I)):
x = np.array([vals[I[i], 0], vals[J[i], 0]])
y = np.array([vals[I[i], 1], vals[J[i], 1]])
z = np.array([vals[I[i], 2], vals[J[i], 2]])
if not update:
if i == 0:
plots.append(ax.plot(x, y, z, c=rcolor if LR[i] else lcolor, lw=RESOPTIONS["linewidth"],
label=['GT' if not pred else 'Pred'])) #lw=2, linestyle='--',
else:
plots.append(ax.plot(x, y, z, lw=RESOPTIONS["linewidth"], c=lcolor if LR[i] else rcolor)) #lw=2, linestyle='--',
elif update:
plots[rang+i][0].set_xdata(x)
plots[rang+i][0].set_ydata(y)
plots[rang+i][0].set_3d_properties(z)
plots[rang+i][0].set_color(lcolor if LR[i] else rcolor)
return plots
def center_around_hip(vals, ax):
global RADIUS
rx, ry, rz = RADIUS[0], RADIUS[1], RADIUS[2]
# remember that y and z are switched in the plot
xroot, zroot, yroot = vals[0,0], vals[0,1], vals[0,2]
#uncomment folowing lines to set a limit to the plot (relevant if grid is)
ax.set_xlim3d([-rx+xroot*2, rx+xroot*2])
ax.set_ylim3d([-ry+yroot*2, ry+yroot*2])
ax.set_zlim3d([-rz+zroot*2, rz+zroot*2])
def update(num,data_gt,plots_gt,fig,ax, data_pred=None, center_pose=True, return_img=False, **kwargs):
gt_vals=data_gt[num]
ax.set_title(ax.get_title(loc='center').split("\n")[0] + f"\nframe {num+1}/{len(data_gt)}")
plots_gt=create_pose(ax,plots_gt,gt_vals,pred=False,update=True, **kwargs)
if data_pred is not None:
if kwargs["multiple_data_pred"]:
for i,prediction in enumerate(data_pred):
gt_plots=create_pose(ax,plots_gt,prediction[num],num_multiple_preds=i+1, pred=True,update=True, **kwargs)
else:
vals=data_pred[num]
plots_gt=create_pose(ax,plots_gt,vals,pred=True,update=True, **kwargs)
if center_pose:
center_around_hip(gt_vals, ax)
if return_img:
img = get_fig_as_nparray(fig)
return img
return plots_gt
def get_np_frames_3d_projection(poses_reshaped3d, limbseq, left_right_limb, data_pred=None, multiple_data_pred=False, xyz_range=None, is_range_sym=False, center_pose=False, units="mm",
as_tensor=False, if_as_overlapping_image=False, orientation_like=None, title=None, center_like=None, if_hide_grid=False, color=None, fig=None):
assert len(poses_reshaped3d.shape) == 3, poses_reshaped3d.shape
assert units in ["mm", "bins"]
assert orientation_like is None or orientation_like in ["motron", "h36m", "somof", "3dpw", "freeman", "best"]
vals = poses_reshaped3d.copy()
if xyz_range is not None:
xyz_range = xyz_range.clone()
timesteps = vals.shape[0]
if fig is None:
fig = plt.figure(figsize=(RESOPTIONS["figsize"], RESOPTIONS["figsize"])) # change here for speed & lightweight, 20
ax = plt.axes(projection='3d')
if title is not None:
ax.set_title(title)
# vals[:,:,0] = poses_reshaped3d[:,:,2].copy()
# vals[:,:,2] = poses_reshaped3d[:,:,0].copy()
if units == "mm":
vals /= 1000 # from mm to meters
if data_pred is not None:
data_pred = data_pred.copy()
if units == "mm":
data_pred/= 1000
ax.set_xlabel("x")
ax.set_ylabel("z")
ax.set_zlabel("y")
global RADIUS
if xyz_range is not None:
assert center_pose == False
if units=="mm":
xyz_range /= 1000
if not is_range_sym:
ax.set_xlim3d([0, xyz_range[0]])
ax.set_ylim3d([0, xyz_range[2]])
ax.set_zlim3d([0, xyz_range[1]])
else:
ax.set_xlim3d([-xyz_range[0]/2, xyz_range[0]/2])
ax.set_ylim3d([-xyz_range[2]/2, xyz_range[2]/2])
ax.set_zlim3d([-xyz_range[1]/2, xyz_range[1]/2])
elif center_like is not None:
if units=="mm":
RADIUS = (center_like[0].max(axis=0)/1000)//2 # symmetric radius distance from body center
RADIUS[RADIUS==0] = 3
center_around_hip(center_like[0]/1000, ax)
else:
RADIUS = center_like[0].max(axis=0)//2 # symmetric radius distance from body center
RADIUS[RADIUS==0] = 1
center_around_hip(center_like[0], ax)
else:
RADIUS = vals[0].max(axis=0)//2 # symmetric radius distance from body center
RADIUS[RADIUS==0] = 3
center_around_hip(vals[0], ax)
ax.set_box_aspect([1,1,1])
if if_hide_grid:
ax.grid(False)
plt.axis('off')
# following to remove ticks
# # following to remove axis labels
if orientation_like is not None:
if orientation_like =="h36m":
ax.view_init(0, -70) # View angle from cameras in h36m, in altitude(angle xzplane with y axis), azimuth
elif orientation_like =="motron":
ax.view_init(12, 48) # motron h36m view angle
elif orientation_like =="freeman" or orientation_like =="best":
ax.invert_zaxis()
ax.view_init(azim=-60, elev=30)
elif orientation_like =="3dpw":
ax.invert_zaxis()
ax.view_init(azim=-90, elev=0) # somof view angle
elif orientation_like =="somof":
ax.view_init(azim=-90, elev=0) # somof view angle
gt_plots=[]
pred_plots=[]
gt_plots=create_pose(ax,gt_plots,vals[0],pred=False,update=False, limbseq=limbseq, left_right_limb=left_right_limb, color=color)
if data_pred is not None:
if multiple_data_pred:
import random
colors = ["#"+''.join([random.choice('ABCDEF0123456789') for i in range(6)]) for i in range(len(data_pred))]
for i,prediction in enumerate(data_pred):
gt_plots=create_pose(ax,gt_plots,prediction[0],pred=True,num_multiple_preds=i+1, update=False, limbseq=limbseq, left_right_limb=left_right_limb, color=colors[i])
else:
gt_plots=create_pose(ax,gt_plots,data_pred[0],pred=True,update=False, limbseq=limbseq, left_right_limb=left_right_limb, color=color)
if if_as_overlapping_image:
def add_timestep(t):
if data_pred is not None and t < data_pred.shape[-3]:
if multiple_data_pred:
for i,prediction in enumerate(data_pred):
__ = create_pose(ax,gt_plots,prediction[t],pred=True,num_multiple_preds=i+1, update=False, color=colors[i], limbseq=limbseq, left_right_limb=left_right_limb)
else:
__ =create_pose(ax,gt_plots,data_pred[t],pred=True,update=False, limbseq=limbseq, left_right_limb=left_right_limb, color=color)
__ = create_pose(ax,gt_plots,vals[t],pred=False,update=False, limbseq=limbseq, left_right_limb=left_right_limb, color=color)
for t in range(timesteps):
add_timestep(t)
frames = [get_fig_as_nparray(fig)]
else:
frames = [update(t,vals,gt_plots,fig,ax, data_pred=data_pred, center_pose=center_pose, return_img=True, limbseq=limbseq, left_right_limb=left_right_limb, color=color, multiple_data_pred=multiple_data_pred) for t in range(len(vals))]
if as_tensor:
frames = [numpy_img_to_tensor(f) for f in frames]
plt.close() # necessary for memory if training)
return frames#, (fig,ax)