def scale_int(s, config, num_overlaps):
shape = config['shape']
ext = feature_extent(s, config)
#introduce noise to ext
ext_noise = config['ext_noise']
ext = np.random.normal(ext, ext_noise*ext)
extsize = ext*2
shapesize = shape[0]
if extsize <= shapesize:
scale = 1
else:
scale = int(np.floor(extsize/shapesize) + 1)
newshape = [i * scale for i in shape]
holo = LMHologram(coordinates=coordinates(newshape))
holo.instrument.properties = config['instrument']
# ... calculate hologram
frame = np.random.normal(0, config['noise'], newshape)
holo.particle = s
holo.particle.x_p += (scale-1)*100
holo.particle.y_p += (scale-1)*100
holo.particle = add_overlaps(ext, num_overlaps, config).append(holo.particle)
frame += holo.hologram().reshape(newshape)
frame = np.clip(100 * frame, 0, 255).astype(np.uint8)
#decimate
frame = frame[::scale, ::scale]
return frame, scale
def format_yolo(sample, config):
'''Returns a string of YOLO annotations'''
(h, w) = config['shape']
fmt = '{}' + 4 * ' {:.6f}' + '\n'
annotation = ''
for sphere in sample:
stype = classify(sphere, config)
diameter = 2. * feature_extent(sphere, config)
x_p = sphere.x_p / w
y_p = sphere.y_p / h
w_p = diameter / w
h_p = diameter / h
annotation += fmt.format(stype, x_p, y_p, w_p, h_p)
return annotation
def scale_float(s, config, num_overlaps):
shape = config['shape']
ext = feature_extent(s, config)
#introduce noise to ext
ext_noise = config['ext_noise']
ext = np.random.normal(ext, ext_noise*ext)
extsize = ext*2
shapesize = shape[0]
scale = float(extsize)/float(shapesize)
newshape = [int(extsize)]*2
holo = LMHologram(coordinates=coordinates(newshape))
holo.instrument.properties = config['instrument']
# ... calculate hologram
frame = np.random.normal(0, config['noise'], newshape)
s.x_p += (scale-1)*100.
s.y_p += (scale-1)*100.
totalspheres = add_overlaps(ext, num_overlaps, config)
totalspheres.append(s)
holo.lorenzmie.particle = totalspheres
frame += holo.hologram().reshape(newshape)
frame = np.clip(100 * frame, 0, 255).astype(np.uint8)
#reshape
#frame = cv2.resize(frame, tuple(shape))
return frame, scale
def localizer_accuracy(configuration='yolov5_test',
nframes=None,
version=None,
plot=False,
savedir='./results/',
overwrite=False):
basedir = os.path.dirname(os.path.abspath(__file__)).split('eval')[0]
path = basedir + 'cfg_yolov5/{}.json'.format(configuration)
with open(path, 'r') as f:
config = json.load(f)
file_header = os.path.abspath(config['directory'])
eval_dir = file_header + '/eval'
mtd_config = config.copy()
mtd_config['directory'] = eval_dir
if not nframes:
nframes = config['nframes_eval']
mtd_config['nframes'] = nframes
mtd_config['particle']['nspheres'] = [1, 1]
mtd_config['overwrite'] = overwrite
print('making data')
makedata(config=mtd_config)
print('data made')
localizer = Localizer(configuration=configuration, version=version)
imgpath_fmt = config['directory'] + '/eval/images/image{}.png'
parampath_fmt = config['directory'] + '/eval/params/image{}.json'
df = pd.DataFrame(columns=[
'img_num', 'x_true', 'y_true', 'ext_true', 'num_detections', 'x_pred',
'y_pred', 'ext_pred', 'conf'
])
style = dict(fill=False, linewidth=3, edgecolor='r')
for n in range(nframes):
print('frame {}'.format(n), end='\r')
img = cv2.imread(imgpath_fmt.format(str(n).zfill(5)))
with open(parampath_fmt.format(str(n).zfill(5)), 'r') as f:
params = json.load(f)[0]
sphere = Sphere()
sphere.loads(params)
params = ast.literal_eval(params)
ext = 2. * feature_extent(sphere, mtd_config)
results = localizer.detect([img])[0]
if plot:
fig, ax = plt.subplots()
ax.imshow(img, cmap='gray')
for feature in results:
corner, w, h = feature['bbox']
ax.add_patch(Rectangle(xy=corner, width=w, height=h, **style))
plt.show()
resultsdict = {
'img_num': n,
'x_true': params['x_p'],
'y_true': params['y_p'],
'ext_true': ext,
'num_detections': len(results),
'x_pred': None,
'y_pred': None,
'ext_pred': None,
'conf': None
}
if len(results) == 1 and not results[0]['edge']:
r = results[0]
p_ext = max(r['bbox'][1:])
resultsdict['x_pred'] = r['x_p']
resultsdict['y_pred'] = r['y_p']
resultsdict['conf'] = r['conf']
resultsdict['ext_pred'] = p_ext
df = df.append(resultsdict, ignore_index=True)
print(df)
saveheader = savedir + configuration
if version:
saveheader += '_v{}'.format(version)
savepath = saveheader + '_eval.csv'
df.to_csv(savepath)
truepos = df[df.num_detections == 1]
falseneg = df[df.num_detections == 0]
falsepos = df[df.num_detections > 1]
numfalsepos = int(falsepos.num_detections.sum() - len(falsepos))
numtruepos = len(truepos) + len(falsepos)
numfalseneg = len(falseneg)
print(
'{} true positive detections, {} false positive (additional) detections, {} false negative detections'
.format(numtruepos, numfalsepos, numfalseneg))
inplane_err_sq = (truepos.x_true - truepos.x_pred)**2 + (truepos.y_true -
truepos.y_pred)**2
inplane_RMSE = np.sqrt(inplane_err_sq.sum() / len(inplane_err_sq))
inplane_err = np.sqrt(inplane_err_sq)
fig, ax = plt.subplots()
loc = ax.scatter(truepos.x_true,
truepos.y_true,
c=np.log(inplane_err),
cmap='Spectral')
ax.set_xlabel(r'$x_p$ [px]')
ax.set_xlabel(r'$y_p$ [px]')
ax.grid(alpha=0.3)
fig.colorbar(loc, label='log(In-plane error [px])')
ax.annotate('In-plane RMSE: {}px'.format('%.1f' % inplane_RMSE),
xy=(0.05, 0.95),
xycoords='axes fraction',
bbox=dict(facecolor='white', edgecolor='black', alpha=0.5))
fig.tight_layout()
fig.savefig(saveheader + '_inplane_err.png')
plt.show()
ext_err_sq = (truepos.ext_true - truepos.ext_pred)**2
ext_RMSE = np.sqrt(ext_err_sq.sum() / len(ext_err_sq))
ext_percent = np.sqrt(ext_err_sq) / truepos.ext_true
print(ext_percent)
ext_perror = ext_percent.mean() * 100
fig, ax = plt.subplots()
ax.plot(truepos.ext_true, truepos.ext_true, c='r')
ax.scatter(truepos.ext_true, truepos.ext_pred, alpha=0.3, c='b')
ax.annotate('{}% Extent Error'.format('%.1f' % ext_perror),
xy=(0.05, 0.95),
xycoords='axes fraction',
bbox=dict(facecolor='white', edgecolor='black', alpha=0.5))
ax.set_xlabel('True feature size [px]')
ax.set_ylabel('Predicted bounding box size [px]')
ax.grid(alpha=0.3)
fig.tight_layout()
fig.savefig(saveheader + '_ext_err.png')
plt.show()
def catch_accuracy(loc='yolov5_test', est='test', nframes=None, version=None, plot=False, weights='best', savedir='./results/'):
basedir = os.path.dirname(os.path.abspath(__file__)).split('eval')[0]
path = basedir + 'cfg_yolov5/{}.json'.format(loc)
with open(path, 'r') as f:
config = json.load(f)
file_header = os.path.abspath(config['directory'])
eval_dir = file_header + '/eval'
mtd_config = config.copy()
mtd_config['directory'] = eval_dir
if not nframes:
nframes = config['nframes_eval']
mtd_config['nframes'] = nframes
mtd_config['particle']['nspheres'] = [1,1]
#mtd_config['overwrite'] = True
makedata(config = mtd_config)
localizer = Localizer(configuration=loc, version=version)
estimator = Estimator(configuration=est, weights=weights)
catch = CATCH(localizer=localizer, estimator=estimator)
imgpath_fmt = config['directory']+'/eval/images/image{}.png'
parampath_fmt = config['directory']+'/eval/params/image{}.json'
df = pd.DataFrame(columns = ['img_num', 'x_true', 'y_true', 'ext_true', 'num_detections', 'x_pred', 'y_pred', 'ext_pred', 'conf', 'z_pred', 'a_pred', 'n_pred', 'z_true', 'a_true', 'n_true'])
style = dict(fill=False, linewidth=3, edgecolor='r')
for n in range(nframes):
img = cv2.imread(imgpath_fmt.format(str(n).zfill(4)))
with open(parampath_fmt.format(str(n).zfill(4)), 'r') as f:
params = json.load(f)[0]
sphere = Sphere()
sphere.loads(params)
params = ast.literal_eval(params)
ext = 2. * feature_extent(sphere, mtd_config)
results = catch.analyze([img])
if plot:
fig, ax = plt.subplots()
ax.imshow(img, cmap='gray')
for feature in results:
corner, w, h = feature['bbox']
ax.add_patch(Rectangle(xy=corner, width=w, height=h, **style))
plt.show()
resultsdict = {'img_num':n, 'x_true':params['x_p'], 'y_true':params['y_p'], 'z_true':params['z_p'], 'a_true': params['a_p'], 'n_true':params['n_p'], 'ext_true':ext, 'num_detections':len(results), 'x_pred':None, 'y_pred':None, 'z_pred': None, 'a_pred': None, 'n_pred': None, 'ext_pred':None, 'conf':None, 'edge':None}
if len(results) == 1:
r = results.iloc[0]
p_ext = max(r['bbox'][1:])
resultsdict['x_pred'] = r['x_p']
resultsdict['y_pred'] = r['y_p']
resultsdict['a_pred'] = r['a_p']
resultsdict['z_pred'] = r['z_p']
resultsdict['n_pred'] = r['n_p']
resultsdict['conf'] = r['conf']
resultsdict['ext_pred'] = p_ext
resultsdict['edge'] = r['edge']
df = df.append(resultsdict, ignore_index=True)
print(df)
if weights=='best':
wstr = ''
else:
wstr=weights
saveheader = savedir+'est_{}{}_loc_{}'.format(est, wstr, loc)
if version:
saveheader += 'v{}'.format(version)
savepath = saveheader + '_eval.csv'
df.to_csv(savepath)
truepos = df[df.num_detections==1]
falseneg = df[df.num_detections==0]
falsepos = df[df.num_detections>1]
numfalsepos = int(falsepos.num_detections.sum() - len(falsepos))
numtruepos = len(truepos) + len(falsepos)
numfalseneg = len(falseneg)
print('{} true positive detections, {} false positive (additional) detections, {} false negative detections'.format(numtruepos, numfalsepos, numfalseneg))
z_rmse = np.sqrt(((truepos.z_pred - truepos.z_true) **2).mean(axis=0))
a_rmse = np.sqrt(((truepos.a_pred - truepos.a_true) **2).mean(axis=0))
n_rmse = np.sqrt(((truepos.n_pred - truepos.n_true) **2).mean(axis=0))
matplotlib.use('TkAgg')
fig, axes = plt.subplots(1,3, figsize=(10,5))
names = ['z_p', 'a_p', 'n_p']
for ax, name in list(zip(axes, names)):
ax.grid(alpha=0.3)
ax.set_xlabel(r'True ${}$'.format(name))
ax.set_ylabel(r'Pred ${}$'.format(name))
ax1, ax2, ax3 = axes
ax1.scatter(df.z_true, df.z_pred, c='b', alpha=0.4)
ax1.plot(df.z_true, df.z_true, c='r')
ax1.annotate('RMSE = {} px'.format('%.3f'%z_rmse), xy=(0.05, 0.95), xycoords='axes fraction')
ax2.scatter(df.a_true, df.a_pred, c='b', alpha=0.4)
ax2.plot(df.a_true, df.a_true, c='r')
ax2.annotate(r'RMSE = {}$\mu m$ '.format('%.3f'%a_rmse), xy=(0.05, 0.95), xycoords='axes fraction')
ax3.scatter(df.n_true, df.n_pred, c='b', alpha=0.4)
ax3.plot(df.n_true, df.n_true, c='r')
ax3.annotate('RMSE = {}'.format('%.3f'%n_rmse), xy=(0.05, 0.95), xycoords='axes fraction')
fig.tight_layout()
figsavepath = saveheader+'_eval.png'
fig.savefig(figsavepath)
plt.show()