def eval_correspondences_pointcloud(self, pcl_pred_files, pcl_tgt):
''' Calculates correspondence score for point clouds.
Args:
pcl_pred_files (list): list of point cloud prediction files
pcl_tgt (list): list of target point clouds
'''
# Detect NN for GT points
pc_t0 = np.expand_dims(load_pointcloud(pcl_pred_files[0]), axis=0)
ind, _ = get_nearest_neighbors_indices_batch(
pc_t0, np.expand_dims(pcl_tgt[0], axis=0))
ind = ind[0]
# Next time steps
eval_dict = {}
for i in range(len(pcl_tgt)):
pc_t = load_pointcloud(pcl_pred_files[i])
pc_nn_t = pcl_tgt[i][ind]
l2_loss = np.mean(np.linalg.norm(pc_t - pc_nn_t, axis=-1)).item()
eval_dict['l2 %d (pcl)' % i] = l2_loss
return eval_dict
if os.path.exists(mesh_file):
mesh = trimesh.load(mesh_file, process=False)
eval_dict_mesh = evaluator.eval_mesh(mesh, pointcloud_tgt,
normals_tgt, points_tgt,
occ_tgt)
for k, v in eval_dict_mesh.items():
eval_dict[k + ' (mesh)'] = v
else:
print('Warning: mesh does not exist: %s' % mesh_file)
# Evaluate point cloud
if cfg['test']['eval_pointcloud']:
pointcloud_file = os.path.join(pointcloud_dir, '%s.ply' % modelname)
if os.path.exists(pointcloud_file):
pointcloud = load_pointcloud(pointcloud_file)
eval_dict_pcl = evaluator.eval_pointcloud(pointcloud,
pointcloud_tgt)
for k, v in eval_dict_pcl.items():
eval_dict[k + ' (pcl)'] = v
else:
print('Warning: pointcloud does not exist: %s' % pointcloud_file)
# Create pandas dataframe and save
eval_df = pd.DataFrame(eval_dicts)
eval_df.set_index(['idx'], inplace=True)
eval_df.to_pickle(out_file)
# Create CSV file with main statistics
eval_df_class = eval_df.groupby(by=['class name']).mean()
eval_df_class.to_csv(out_file_class)
# Evaluate point cloud
if cfg['test']['eval_pointcloud']:
pointcloud_folder = os.path.join(pointcloud_dir, '%s' % modelname,
'%05d' % start_idx)
if os.path.exists(pointcloud_folder):
ply_files = glob.glob(os.path.join(pointcloud_folder, '*.ply'))
ply_files.sort()
# Eval Pointcloud Correspondences
if cfg['test']['eval_pointcloud_correspondences']:
eval_dict_cor = evaluator.eval_correspondences_pointcloud(
ply_files, pointcloud_tgt)
for k, v in eval_dict_cor.items():
eval_dict[k] = v
# Eval Point Cloud Chamfer distances
for i, ply_file in enumerate(ply_files[:17]):
pointcloud = load_pointcloud(ply_file)
eval_dict_pcl = evaluator.eval_pointcloud(
pointcloud, pointcloud_tgt[i])
for k, v in eval_dict_pcl.items():
eval_dict['%s %d (pcl)' % (k, i)] = v
else:
print('Warning: pointcloud does not exist: %s (%d)' %
(modelname, start_idx))
if it > 0 and (it % 50) == 0:
# Create pandas dataframe and save
eval_df = pd.DataFrame(eval_dicts)
eval_df.set_index(['idx'], inplace=True)
eval_df.to_pickle(out_file_tmp)
# Create CSV file with main statistics
if os.path.exists(mesh_file):
mesh = trimesh.load(mesh_file)
eval_dict_mesh = evaluator.eval_mesh(
mesh, pointcloud_tgt)
for k, v in eval_dict_mesh.items():
eval_dict[k + ' (mesh)'] = v
else:
print('Warning: mesh does not exist: %s' % mesh_file)
if eval_pointcloud:
pointcloud_file = os.path.join(output_dir,
'mesh_{}_{}_{}_{}_{}.ply'.format(split, data_completeness, data_sparsity,
checkpoint, it))
if os.path.exists(pointcloud_file):
pointcloud = load_pointcloud(pointcloud_file).astype(np.float32)
eval_dict_pcl = evaluator.eval_pointcloud(
pointcloud, pointcloud_tgt)
for k, v in eval_dict_pcl.items():
eval_dict[k + ' (pcl)'] = v
else:
print('Warning: pointcloud does not exist: %s' % pointcloud_file)
out_file = os.path.join(output_dir, 'result',
'eval_input_full_{}_{}_{}_{}.pkl'.format(split, data_completeness, data_sparsity, checkpoint))
out_file_class = os.path.join(output_dir, 'result',
'eval_input_{}_{}_{}_{}.csv'.format(split, data_completeness, data_sparsity, checkpoint))
# Create pandas dataframe and save
eval_df = pd.DataFrame(eval_dicts)
eval_df.set_index(['idx'], inplace=True)
