inplanes = sess.run(sess.graph.get_tensor_by_name('inplanes:0'))
priors = sess.run(sess.graph.get_tensor_by_name('priors:0'))
variances = sess.run(sess.graph.get_tensor_by_name('variances:0'))
priors = np.concatenate((priors, variances), axis=1)
# Get tensor handles
tensor_in = sess.graph.get_tensor_by_name('input:0')
tensor_loc = sess.graph.get_tensor_by_name('locations:0')
tensor_cla = sess.graph.get_tensor_by_name('class_probs:0')
tensor_view = sess.graph.get_tensor_by_name('view_probs:0')
tensor_inpl = sess.graph.get_tensor_by_name('inplane_probs:0')
if len(models) == 1: # If single-object network
models = ['obj_{:02d}'.format(sequence)] # Overwrite model name
bench = load_sixd(sixd_base, nr_frames=nr_frames, seq=sequence)
input_shape = (1, 299, 299, 3)
print('Models:', models)
print('Views:', len(views))
print('Inplanes:', len(inplanes))
print('Priors:', priors.shape)
print('Precomputing projections for each used model...')
model_map = bench.models # Mapping from name to model3D instance
for model_name in models:
m = model_map[model_name]
m.projections = precompute_projections(views, inplanes, bench.cam, m)
# Process each frame separately
for f in bench.frames:
inplanes = sess.run(sess.graph.get_tensor_by_name('inplanes:0'))
priors = sess.run(sess.graph.get_tensor_by_name('priors:0'))
variances = sess.run(sess.graph.get_tensor_by_name('variances:0'))
priors = np.concatenate((priors, variances), axis=1)
# Get tensor handles
tensor_in = sess.graph.get_tensor_by_name('input:0')
tensor_loc = sess.graph.get_tensor_by_name('locations:0')
tensor_cla = sess.graph.get_tensor_by_name('class_probs:0')
tensor_view = sess.graph.get_tensor_by_name('view_probs:0')
tensor_inpl = sess.graph.get_tensor_by_name('inplane_probs:0')
if len(models) == 1: # If single-object network
models = ['obj_{:02d}'.format(sequence)] # Overwrite model name
bench = load_sixd(sixd_base, nr_frames=nr_frames, seq=sequence)
input_shape = (1, 299, 299, 3)
print('Models:', models)
print('Views:', len(views))
print('Inplanes:', len(inplanes))
print('Priors:', priors.shape)
print('Precomputing projections for each used model...')
model_map = bench.models # Mapping from name to model3D instance
for model_name in models:
m = model_map[model_name]
m.projections = precompute_projections(views, inplanes, bench.cam, m)
# Process each frame separately
for f in bench.frames:
print(
'===========================> If this step takes too long, you can enable the --vis_fast flag to use fast rendering (real-time).'
)
while (writer.running()):
pass
writer.stop()
final_result = writer.results()
# Till now all the results from detections are in final_result.
# # Output results in json file
write_json(final_result, args.outputpath)
'''
Evaluate final_result*******************************************************
'''
print("Loading ground truth benchmark dataset...")
bench_info = load_sixd(sixd_base, seq=obj_id, nr_frames=0)
diameter = bench_info.diameter[obj_id]
frames_of_ground_truth = bench_info.frames
# Metrics Initialization
add_errs = []
adds = []
proj_2d_errs = []
ious = []
np.set_printoptions(suppress=True)
# for f in tqdm(final_result, ncols=80, ascii=True):
for idx, f in enumerate(final_result):
imgname = f['imgname']
imgname = int(imgname[0:-4]) # throw '.png'
gt_frame = frames_of_ground_truth[imgname]
assert imgname == gt_frame.nr
# embed()
from refiner.refiner import Refiner, Refinable
from rendering.utils import perturb_pose, trans_rot_err
from timeit import default_timer as timer
from docopt import docopt
args = docopt(__doc__)
sixd_base = args["--dataset"]
obj = args["--object"]
network = args["--network"]
max_rot_pert = float(args["--max_rot_pert"]) / 180. * np.pi
max_trans_pert = float(args["--max_trans_pert"])
iterations = int(args["--iterations"])
bench = load_sixd(sixd_base, nr_frames=1, seq=obj)
croppings = yaml.load(open('config/croppings.yaml', 'r'))
if 'linemod' in network:
dataset_name = 'linemod'
elif 'tejani' in network:
dataset_name = 'tejani'
else:
raise Exception('Could not determine dataset')
with tf.Session() as session:
architecture = Architecture(network_file=network, sess=session)
ren = Renderer((640, 480), bench.cam)
refiner = Refiner(architecture=architecture, ren=ren, session=session)
for frame in bench.frames:
print(
'===========================> If this step takes too long, you can enable the --vis_fast flag to use fast rendering (real-time).'
)
while (writer.running()):
pass
writer.stop()
final_result = writer.results()
# Till now all the results from detections are in final_result.
# # Output results in json file
write_json(final_result, args.outputpath)
'''
Evaluate final_result*******************************************************
'''
print("Loading ground truth of OCCLUSION dataset...")
bench_info = load_sixd(sixd_base, seq=2, nr_frames=0)
diameter = bench_info.diameter[obj_id]
frames_of_ground_truth = bench_info.frames
# Metrics Initialization
add_errs = []
adds = []
proj_2d_errs = []
ious = []
# for f in tqdm(final_result, ncols=80, ascii=True):
for idx, f in enumerate(final_result):
imgname = f['imgname']
imgname = int(imgname[0:-4]) # throw '.png'
gt_frame = frames_of_ground_truth[imgname]
assert imgname == gt_frame.nr
for ground_truth in gt_frame.gt:
gt_obj_id = ground_truth[0]
"""
from docopt import docopt
import numpy as np
from rendering.utils import precompute_projections
import glob
from utils.sixd import load_sixd
import os
import math
def calc_azimuth(x, y):
two_pi = 2.0 * math.pi
return (math.atan2(y, x) + two_pi) % two_pi
args = docopt(__doc__)
sixd_base = 'dataset/' + args["--dataset"]
sequences = args["--sequence"].split()
pts = np.load("numpy_files/viewpoints.npy")
sequences = range(1, 21)
models = ['obj_{:06d}'.format(int(x))
for x in sequences] # Overwrite model name
bench = load_sixd(sixd_base, sequences=sequences)
print('Models:', models)
print('Precomputing projections for each used model...')
# image_files = glob.glob(os.path.join(os.path.expanduser("~"), 'data', "images/*.jpg")) # put coco images as background in this folder
model_map = bench.models # Mapping from name to model3D instance
precompute_projections(args["--dataset"], pts, bench.cam, model_map, models)