def _test_blendswapdata(self, data):
normal_tensor = ops.warp2d(normalized=True,
border_mode='value',
border_value=-1,
input=data['image_pair'][:, 3:, :, :],
displacements=data['flow'])
warped = normal_tensor.eval()
if (warped.dtype == np.float32):
import ijremote
ijremote.setImage("image_pair", data['image_pair'][:, 0:3, :, :])
ijremote.setImage("warped_image", warped)
def testing_ds_api(dataset):
dataset = dataset.make_initializable_iterator()
sess = tf.InteractiveSession()
sess.run(dataset.initializer)
next_batch = dataset.get_next()
final_img_batch, final_lbl_batch = combine_batches_from_datasets(next_batch)
next_batch_forward = final_img_batch[:,0,:,:,:]
next_batch_backward = final_img_batch[:,1,:,:,:]
forward, backward = sess.run([next_batch_forward,next_batch_backward])
ij.setHost('tcp://linus:13463')
ij.setImage('myimage_f',np.transpose(forward,[0,3,1,2]))
ij.setImage('myimage_b',np.transpose(backward,[0,3,1,2]))
def _test_blendswapdata(self, data):
depth_tensor = ops.flow_to_depth(inverse_depth=True,
normalized_flow=True,
flow=data['flow'],
intrinsics=data['intrinsics'],
rotation=data['motion'][:, 0:3],
translation=data['motion'][:, 3:])
depth = depth_tensor.eval()
import ijremote
ijremote.setImage("image_pair", data['image_pair'])
ijremote.setImage("depth", depth)
ijremote.setImage("depth_gt", data['depth'])
self.assertAllClose(depth, data['depth'])
def _test_blendswapdata(self, data):
flow_tensor = ops.depth_to_flow(inverse_depth=True,
normalize_flow=True,
depth=data['depth'],
intrinsics=data['intrinsics'],
rotation=data['motion'][:, 0:3],
translation=data['motion'][:, 3:])
flow = flow_tensor.eval()
if data['depth'].dtype == np.float32:
import ijremote
#ijremote.setHost('tcp://clancy:13463')
ijremote.setImage("image_pair", data['image_pair'])
ijremote.setImage("flow", flow)
ijremote.setImage("flow_gt", data['flow'])
self.assertAllClose(flow, data['flow'])
def _test_blendswapdata(self, data):
normal_tensor = ops.depth_to_normals(inverse_depth=True,
depth=data['depth'],
intrinsics=data['intrinsics'])
normal = normal_tensor.eval()
if (normal.dtype == np.float32):
import ijremote
ijremote.setImage("image_pair", data['image_pair'])
ijremote.setImage("normal", normal)
ijremote.setImage("normal_gt", data['normal'])
import vis
img1, _ = vis.numpy_imagepair_to_PIL_images(data['image_pair'][1])
visualize_points_with_normals('points', data['depth'][1],
data['intrinsics'][1], img1,
normal[1])
visualize_points_with_normals('points_gt', data['depth'][1],
data['intrinsics'][1], img1,
data['normal'][1])
normal_no_nan = normal.copy()
normal_no_nan[np.isnan(normal)] = 0
normal_gt_no_nan = data['normal'].copy()
normal_gt_no_nan[np.isnan(data['normal'])] = 0
gt_flow = predictor.optical_floww
else:
predictor.preprocess()
gt_flow, gt_depth_change = predictor.read_gt(FLAGS.FLOW,FLAGS.DISPARITY_CHNG)
pr_flow, loss = predictor.predict()
print(pr_flow.shape)
if not loss is None:
print(loss)
if FLAGS.PARSING_PTB == False:
# show gt flows
if FLAGS.SHOW_GT_FLOWS == True:
ij.setImage('gt_flow_uv',np.transpose(gt_flow,[2,0,1]))
# warp with gt predited flow values
if FLAGS.SHOW_GT_WARPED_RESULT == True:
# gt_flow = np.pad(gt_flow,((4,4),(0,0),(0,0)),'constant')
flow = predictor.warp(predictor.img2_arr,gt_flow)
result = flow.eval()[0].astype(np.uint8)
predictor.show_image(result,'warped_img_gt')
# show inv depth values for both images
if FLAGS.SHOW_GT_DEPTHS == True:
ij.setImage('gt_inv_depth1',predictor.inv_depth1)
ij.setImage('gt_inv_depth2',predictor.inv_depth2)
predictor = FlowPredictor()
predictor.preprocess(FLAGS.IMG1, FLAGS.IMG2, FLAGS.DISPARITY1,
FLAGS.DISPARITY2)
gt_flow, gt_depth_change = predictor.read_gt(FLAGS.FLOW, FLAGS.DISPARITY_CHNG)
pr_flow, pr_depth_change = predictor.predict()
# show gt images
if FLAGS.SHOW_GT_IMGS == True:
predictor.init_img1.show()
predictor.init_img2.show()
# show gt flows
if FLAGS.SHOW_GT_FLOWS == True:
ij.setImage('gt_flow_u', gt_flow[:, :, 0])
ij.setImage('gt_flow_v', gt_flow[:, :, 1])
# warp with gt predited flow values
if FLAGS.SHOW_GT_WARPED_RESULT == True:
gt_flow = np.pad(gt_flow, ((4, 4), (0, 0), (0, 0)), 'constant')
flow = predictor.warp(predictor.img2_arr, gt_flow)
result = flow.eval()[0].astype(np.uint8)
predictor.show_image(result, 'warped_img_gt')
# show inv depth values for both images
if FLAGS.SHOW_GT_DEPTHS == True:
ij.setImage('gt_inv_depth1', predictor.inv_depth1)
ij.setImage('gt_inv_depth2', predictor.inv_depth2)
# show predicted depth change