def get_pos_only_prediction(model, trace, pred_hor):
indices_input_trace = np.linspace(0,
len(trace) - 1,
M_WINDOW + 1,
dtype=int)
subsampled_trace = trace[indices_input_trace]
encoder_pos_inputs = subsampled_trace[-M_WINDOW - 1:-1]
decoder_pos_inputs = subsampled_trace[-1:]
model_prediction = model.predict([
np.array(
transform_batches_cartesian_to_normalized_eulerian(
[encoder_pos_inputs])),
np.array(
transform_batches_cartesian_to_normalized_eulerian(
[decoder_pos_inputs]))
])[0][int(pred_hor / MODEL_SAMPLING_RATE)]
model_prediction = transform_normalized_eulerian_to_cartesian(
model_prediction)
yaw_pred, pitch_pred = cartesian_to_eulerian(model_prediction[0],
model_prediction[1],
model_prediction[2])
yaw_pred, pitch_pred = transform_the_radians_to_original(
yaw_pred, pitch_pred)
return create_head_map(np.array([yaw_pred, pitch_pred]), cartesian=False)
def get_CVPR18_prediction(model, trace, saliency, pred_hor):
indices_input_trace = np.linspace(0,
len(trace) - 1,
M_WINDOW + 1,
dtype=int)
indices_input_saliency = np.linspace(0,
len(saliency) - 1,
M_WINDOW + H_WINDOW,
dtype=int)
subsampled_trace = trace[indices_input_trace]
subsampled_saliency = saliency[indices_input_saliency]
encoder_pos_inputs = subsampled_trace[-M_WINDOW - 1:-1]
decoder_pos_inputs = subsampled_trace[-1:]
# encoder_sal_inputs = subsampled_saliency[:M_WINDOW]
decoder_sal_inputs = subsampled_saliency[M_WINDOW:]
model_prediction = model.predict([
np.array([encoder_pos_inputs]),
np.array([decoder_pos_inputs]),
np.array([decoder_sal_inputs])
])[0][int(pred_hor / MODEL_SAMPLING_RATE)]
yaw_pred, pitch_pred = cartesian_to_eulerian(model_prediction[0],
model_prediction[1],
model_prediction[2])
yaw_pred, pitch_pred = transform_the_radians_to_original(
yaw_pred, pitch_pred)
return create_head_map(np.array([yaw_pred, pitch_pred]), cartesian=False)
def recover_original_angles_from_xyz_trace(xyz_trace):
angles_per_video = []
for sample in xyz_trace:
restored_yaw, restored_pitch = cartesian_to_eulerian(sample[1], sample[2], sample[3])
restored_yaw, restored_pitch = transform_the_radians_to_original(restored_yaw, restored_pitch)
angles_per_video.append(np.array([restored_yaw, restored_pitch]))
return np.array(angles_per_video)
def transform_batches_cartesian_to_normalized_eulerian(positions_in_batch):
positions_in_batch = np.array(positions_in_batch)
eulerian_batches = [[
cartesian_to_eulerian(pos[0], pos[1], pos[2]) for pos in batch
] for batch in positions_in_batch]
eulerian_batches = np.array(eulerian_batches) / np.array(
[2 * np.pi, np.pi])
return eulerian_batches
def recover_original_angles_from_quaternions_trace(quaternions_trace):
angles_per_video = []
orig_vec = np.array([1, 0, 0])
for sample in quaternions_trace:
quat_rot = Quaternion(sample[1:])
sample_new = quat_rot.rotate(orig_vec)
restored_yaw, restored_pitch = cartesian_to_eulerian(sample_new[0], sample_new[1], sample_new[2])
restored_yaw, restored_pitch = transform_the_radians_to_original(restored_yaw, restored_pitch)
angles_per_video.append(np.array([restored_yaw, restored_pitch]))
return np.array(angles_per_video)
def transform_back_predicted_angles(sample):
restored_yaw, restored_pitch = cartesian_to_eulerian(sample[0], sample[1], sample[2])
restored_yaw, restored_pitch = transform_the_radians_to_original(restored_yaw, restored_pitch)
return restored_yaw, restored_pitch
