depth_threshold = 15
inlier_thresh = 0.02
cam_height = 1.70
plotting = False
plot_range = slice(0, -1)
u_crop_min_num = 1
u_crop_min_den = 6
u_crop_max_num = 5
u_crop_max_den = 6
v_crop_min_num = 4
v_crop_min_den = 7
v_crop_max_num = 1
v_crop_max_den = 1
config = load_obj('{}/config'.format(dir))
print(config)
pretrained_depth_path = glob.glob(
'{}/**depth**best-loss-val_seq-**-test_seq-{}**.pth'.format(dir,
''))[0]
pretrained_pose_path = glob.glob(
'{}/**pose**best-loss-val_seq-**-test_seq-{}**.pth'.format(dir, ''))[0]
config['data_dir'] = path_to_dset_downsized + config[
'img_resolution'] + '_res/' #
## uncomment for KITTI to oxford
# config['estimator'] = 'orbslam'
config['load_stereo'] = False
config['augment_motion'] = False
import sys
sys.path.append('../')
import numpy as np
import matplotlib.pyplot as plt
from utils.learning_helpers import save_obj, load_obj
import os
import matplotlib
# Removes the XWindows backend (useful for producing plots via tmux without -X)
matplotlib.use('Agg', warn=False)
os.makedirs('figures', exist_ok=True)
path_to_ws = '/home/brandon-wagstaff/learned_scale_recovery/'
scaled_file = path_to_ws + 'results/202007111233-kitti-scaled-good/scale_factor'
scale = load_obj(scaled_file)
plt.rc('text', usetex=True)
plt.rc('font', family='serif')
f, axarr = plt.subplots(2, sharex=True, sharey=True)
axarr[0].tick_params(labelsize=22)
axarr[1].tick_params(labelsize=22)
axarr[0].plot(scale[1][0:-1500:3], linewidth=2, rasterized=True)
axarr[1].plot(scale[19][0:-1500:3], linewidth=2, rasterized=True)
axarr[0].set_ylabel('Scale Factor', fontsize=22)
axarr[1].set_ylabel('Scale Factor', fontsize=22)
axarr[1].set_xlabel('Minibatch', fontsize=22)
axarr[0].set_title('First Epoch', fontsize=19)
axarr[1].set_title('Final Epoch', fontsize=19)
axarr[0].grid()
axarr[1].grid()
from validate import compute_trajectory as tt
from pose_evaluation_utils import dump_pose_seq_TUM, rot2quat, pose_vec_to_mat, compute_ate
from glob import glob
import os
import matplotlib.pyplot as plt
seq = '10_02'
path_to_ws = '/home/brandonwagstaff/learned_scale_recovery/'
model_dir = 'results/202101281256-3-iter-unscaled-good'
path_to_dset_downsized = '/media/datasets/KITTI-odometry-downsized-stereo/'
out_folder = 'pose_data/our_predictions/{}'.format(seq[0:2])
# out_folder = 'pose_data/sfmlearner_results/{}'.format(seq[0:2])
os.makedirs(out_folder, exist_ok=True)
gt_folder = 'pose_data/ground_truth/{}/'.format(seq[0:2])
dir = path_to_ws + model_dir
config = load_obj('{}/config'.format(dir))
config[
'data_dir'] = path_to_dset_downsized + config['img_resolution'] + '_res/'
config['img_per_sample'] = 2
test_dset = KittiLoaderPytorch(
config, [[seq], [seq], [seq]],
mode='test',
transform_img=get_data_transforms(config)['test'])
test_dset_loaders = torch.utils.data.DataLoader(test_dset,
batch_size=config['minibatch'],
shuffle=False,
num_workers=6)
ts = test_dset.raw_ts[0]
import sys
sys.path.append('../')
import numpy as np
import matplotlib.pyplot as plt
from utils.learning_helpers import save_obj, load_obj
import os
os.makedirs('figures', exist_ok=True)
path_to_ws = '/home/brandon-wagstaff/learned_scale_recovery/'
scaled_file = path_to_ws + 'results/202008041237-oxford-to-kitti-1-epoch-scaled/scale_factor'
unscaled_file = path_to_ws + 'results/202008041450-oxford-to-kitti-1-epoch-unscaled/scale_factor'
scale = load_obj(scaled_file)
unscale = load_obj(unscaled_file)
scaled_average = []
for i in range(0, len(scale[1]), 100):
scaled_average.append(np.mean(scale[1][i:i + 100]))
plt.rc('text', usetex=True)
plt.rc('font', family='serif')
f, axarr = plt.subplots(1)
axarr.tick_params(labelsize=19)
axarr.plot(scale[1][0:-1000:3], linewidth=2) #,color='red',linestyle='--')
axarr.plot(unscale[1][0:-1000:3], linewidth=2) #,color='gold',linestyle='-')
axarr.set_ylabel('Scale Factor', fontsize=19)
axarr.set_xlabel('Minibatch', fontsize=19)
MAX_DEPTH = 80
path_to_ws = '/home/brandonwagstaff/learned_scale_recovery/' ##update this
path_to_dset_downsized = '/media/datasets/KITTI-eigen-split/'
dir = path_to_ws + 'results/final_models/eigen-depth-eval-scaled-202102141219'
pretrained_plane_dir = 'results/plane-model-eigen-202101201842'
cam_height = 1.70
median_scaling = False #align scale of predicted depth with ground truth using median depth
plane_rescaling = False #align scale using ground plane detection and known camera height
post_process = True #use the standard post-processing that flips images, recomputes depth, and merges with unflipped depth
benchmark = 'eigen' ### eigen_benchmark for improved gt, 'eigen' for standard benchmark
splits_dir = '{}/data/splits/{}'.format(path_to_ws, benchmark)
config = load_obj('{}/config'.format(dir))
pretrained_depth_path = glob.glob(
'{}/**depth**best-loss-val_seq-**-test_seq-{}**.pth'.format(dir,
''))[0]
pretrained_pose_path = glob.glob(
'{}/**pose**best-loss-val_seq-**-test_seq-{}**.pth'.format(dir, ''))[0]
config['data_dir'] = path_to_dset_downsized + config[
'img_resolution'] + '_res/' #
config['minibatch'] = 6
config['load_pretrained'] = True
_, models, device = data_and_model_loader(config,
pretrained_depth_path,
pretrained_pose_path,
seq=None)