gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.7)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
tf.train.start_queue_runners(sess=sess)
colorlist_pred = ['#660000', '#b30000', '#ff0000', '#ff4d4d', '#ff9999']
colorlist_gt = ['#000066', '#0000b3', '#0000ff', '#4d4dff', '#9999ff']
# Plots all 5 fingers
def plot_fingers(points, plt_specs, c, ax):
for i in range(5):
start, end = i*4+1, (i+1)*4+1
to_plot = np.concatenate((points[start:end], points[0:1]), axis=0)
ax.plot(to_plot[:,0], to_plot[:,1], to_plot[:,2], plt_specs, color=c[i])
fig = None
util_vae = EvalUtil()
# iterate dataset
for i in range(dataset.num_samples):
# Get required data
keypoint_xyz21, keypoint_vis21, keypoint_scale, img_crop, hand_side = \
sess.run([data['keypoint_xyz21'], data['keypoint_vis21'], data['keypoint_scale'], data['image_crop'], data['hand_side']])
keypoint_xyz21 = np.squeeze(keypoint_xyz21)
keypoint_vis21 = np.squeeze(keypoint_vis21)
keypoint_scale = np.squeeze(keypoint_scale)
img_crop = np.squeeze(img_crop)
# Get VAE prediction
img_res = cv.resize(img_crop, end_size)
img_pyt = img_res.transpose(2, 0, 1).reshape((1, 3, end_size[0], end_size[1]))
img_crop_var = Variable(np2pyt(img_pyt), volatile=True).cuda()
# retrained version: PoseNet
last_cpt = tf.train.latest_checkpoint(PATH_TO_POSENET_SNAPSHOTS)
assert last_cpt is not None, "Could not locate snapshot to load. Did you already train the network and set the path accordingly?"
load_weights_from_snapshot(sess,
last_cpt,
discard_list=['Adam', 'global_step', 'beta'])
else:
# load weights used in the paper
net.init(sess,
weight_files=[
'./weights/handsegnet-rhd.pickle',
'./weights/posenet-rhd-stb.pickle'
],
exclude_var_list=['PosePrior', 'ViewpointNet'])
util = EvalUtil()
# iterate dataset
for i in range(dataset.num_samples):
# get prediction
keypoints_scoremap_v,\
scale_crop_v, center_v, kp_uv21_gt, kp_vis = sess.run([keypoints_scoremap, scale_crop, center, data['keypoint_uv21'], data['keypoint_vis21']])
keypoints_scoremap_v = np.squeeze(keypoints_scoremap_v)
kp_uv21_gt = np.squeeze(kp_uv21_gt)
kp_vis = np.squeeze(kp_vis)
# detect keypoints
coord_hw_pred_crop = detect_keypoints(np.squeeze(keypoints_scoremap_v))
coord_hw_pred = trafo_coords(coord_hw_pred_crop, center_v, scale_crop_v,
256)
coord_uv_pred = np.stack([coord_hw_pred[:, 1], coord_hw_pred[:, 0]], 1)
]) # left hand (true for all samples provided)
evaluation = tf.placeholder_with_default(True, shape=())
# build network
net = ColorHandPose3DNetwork()
hand_scoremap_tf, image_crop_tf, scale_tf, center_tf, \
keypoints_scoremap_tf, keypoint_coord3d_tf = net.inference_crop(image_tf, hand_side_tf, evaluation)
# Start TF
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.8)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
# initialize network
net.init(sess)
# Feed image list through network
eval2d = EvalUtil()
eval3d = EvalUtil()
import tqdm
import cv2
for i in tqdm.tqdm(range(len(data))):
sample = data[i]
image_v = sample['image']
hand_scoremap_v, image_crop_v, scale_v, center_v, \
keypoints_scoremap_v, keypoint_coord3d_v = sess.run([hand_scoremap_tf, image_crop_tf, scale_tf, center_tf,
keypoints_scoremap_tf, keypoint_coord3d_tf],
feed_dict={image_tf: image_v})
hand_scoremap_v = np.squeeze(hand_scoremap_v)
image_crop_v = np.squeeze(image_crop_v)