small_transform = tf.constant(config.camera_params['cam_transform_02_inv'], dtype = tf.float32) X2_pooled = tf.nn.max_pool(X2, ksize=[1,5,5,1], strides=[1,1,1,1], padding="SAME") net = global_agg_net.Nets(X1, X2_pooled, phase_rgb, phase, fc_keep_prob) output_vectors, weight_summaries = net.build() # se(3) -> SE(3) for the whole batch # output_vectors_ft = tf.map_fn(lambda x:transform_functions.RV2RM(expected_transforms[x], output_vectors[x]), elems=tf.range(0, batch_size, 1), dtype=tf.float32) # output_vectors_ft = tf.reshape(output_vectors_ft, shape=(batch_size, 6)) predicted_transforms = tf.map_fn(lambda x:exponential_map_single(output_vectors[x]), elems=tf.range(0, batch_size * time_step, 1), dtype=tf.float32) # predicted_transforms = tf.concat([expected_transforms[:, :3, :3], tf.reshape(predicted_transforms[:, :3, 3], shape=[batch_size, 3, 1])], axis=-1) # transforms depth maps by the predicted transformation depth_maps_predicted, cloud_pred = tf.map_fn(lambda x:at3._simple_transformer(X2_pooled[x,:,:,0]*40.0 + 40.0, predicted_transforms[x], K_final, small_transform), elems = tf.range(0, batch_size * time_step, 1), dtype = (tf.float32, tf.float32)) # transforms depth maps by the expected transformation depth_maps_expected, cloud_exp = tf.map_fn(lambda x:at3._simple_transformer(X2_pooled[x,:,:,0]*40.0 + 40.0, expected_transforms[x], K_final, small_transform), elems = tf.range(0, batch_size * time_step, 1), dtype = (tf.float32, tf.float32)) # photometric loss between predicted and expected transformation photometric_loss = tf.nn.l2_loss(tf.subtract((depth_maps_expected[:,10:-10,10:-10] - 40.0)/40.0, (depth_maps_predicted[:,10:-10,10:-10] - 40.0)/40.0)) # point cloud distance between point clouds cloud_loss = model_utils.get_cd_loss(cloud_pred, cloud_exp) # earth mover's distance between point clouds emd_loss = model_utils.get_emd_loss(cloud_pred, cloud_exp) # regression loss output_vectors_exp = tf.map_fn(lambda x: transform_functions.convert(expected_transforms[x]), elems=tf.range(0, batch_size, 1), dtype=tf.float32) output_vectors_exp = tf.squeeze(output_vectors_exp) tr_loss = tf.norm(output_vectors[:, :3] - output_vectors_exp[:, :3], axis=1)
ksize=[1, 5, 5, 1],
strides=[1, 1, 1, 1],
padding="SAME")
output_vectors, weight_summaries = global_agg_net.End_Net_Out(
X1, phase_rgb, X2_pooled, phase, keep_prob)
# se(3) -> SE(3) for the whole batch
predicted_transforms = tf.map_fn(
lambda x: exponential_map_single(output_vectors[x]),
elems=tf.range(0, batch_size, 1),
dtype=tf.float32)
# transforms depth maps by the predicted transformation
depth_maps_predicted, cloud_pred = tf.map_fn(lambda x: at3._simple_transformer(
X2_pooled[x, :, :, 0] * 40.0 + 40.0, predicted_transforms[x], K_final,
small_transform),
elems=tf.range(0, batch_size, 1),
dtype=(tf.float32, tf.float32))
# transforms depth maps by the expected transformation
depth_maps_expected, cloud_exp = tf.map_fn(lambda x: at3._simple_transformer(
X2_pooled[x, :, :, 0] * 40.0 + 40.0, expected_transforms[x], K_final,
small_transform),
elems=tf.range(0, batch_size, 1),
dtype=(tf.float32, tf.float32))
# photometric loss between predicted and expected transformation
photometric_loss = tf.nn.l2_loss(
tf.subtract((depth_maps_expected[:, 10:-10, 10:-10] - 40.0) / 40.0,
(depth_maps_predicted[:, 10:-10, 10:-10] - 40.0) / 40.0))