def main():
from core.kitti import KITTI, ALL_OBJECTS, CARS_ONLY
DS_DIR = '/home/salam/datasets/KITTI/training'
reader = KITTI(DS_DIR, CARS_ONLY)
ids = reader.get_ids('train')
for t in [ids[4]]:
img = reader.get_image(t)
pts, ref = reader.get_velo(t, use_fov_filter=False)
boxes_2D = reader.get_boxes_2D(t)
boxes_3D = reader.get_boxes_3D(t)
P2 = reader.get_calib(t)[2]
for b in boxes_3D:
b.text = f"{np.random.uniform(0.0, 1.0, (1,))[0]:0.2f}"
imshow(bev(pts, pred_boxes=boxes_3D, title="GT"))
return fmap
rv2bev_2x = process_fmap(rv2bev_2x)
bev_2x = process_fmap(bev_2x)
rv2bev_8x = process_fmap(rv2bev_8x)
bev_8x = process_fmap(bev_8x)
bevout = process_fmap(bevout)
rv_high = process_fmap(rv_high)
rv_low = process_fmap(rv_low)
f2 = process_fmap(f2)
f4 = process_fmap(f4)
outmap = np.squeeze(np.concatenate((obj3d, geo), axis=-1))
decoded_boxes = target_encoder.decode(outmap, 0.2)
filtered_boxes = nms(decoded_boxes)
pts, _ = kitti.get_velo(ids[0], use_fov_filter=True)
boxes = kitti.get_boxes_3D(ids[0])
cnvs = bev(pts=pts.T,
gt_boxes=boxes,
pred_boxes=filtered_boxes)
cnvs = np.rot90(cnvs)
plt.imshow(cnvs)
plt.axis('off')
buf = io.BytesIO()
plt.savefig(buf,
format='png',
bbox_inches='tight',
pad_inches=0)
buf.seek(0)
plot = tf.image.decode_png(buf.getvalue(), channels=4)
plot = tf.expand_dims(plot, 0)
P2,
parts=4)
print('world_pts.shape', world_pts.shape)
print('nearest.shape ', nearest.shape)
geo_feature = nearest - world_pts
nearest_projected = project(P2, nearest).astype(
np.int32).T // image_downsampling_factor
return {
'mapping': nearest_projected.reshape((bev_length, bev_width, 2)),
'geo_feature': geo_feature.reshape((bev_length, bev_width, 3))
}
for i, id in enumerate(train_ids + val_ids):
pc, _ = kitti.get_velo(id,
workspace_lim=((-40, 40), (-1, 3), (0, 70)),
use_fov_filter=True)
_, _, P2 = kitti.get_calib(id)
# img = kitti.get_image(id)
# img = cv2.resize(img, (311, 94))
# print('img.shape', img.shape)
bev_shape = (448, 512)
world_pts2x, nearest2x = get_nearest_neighbour(pc, bev_shape[1] // 2,
bev_shape[0] // 2, 2, P2)
# mapping2x[:,:,(0,1)] = mapping2x[:,:,(1,0)]
world_pts4x, nearest4x = get_nearest_neighbour(pc, bev_shape[1] // 4,
bev_shape[0] // 4, 4, P2)
# mapping4x[:,:,(0,1)] = mapping4x[:,:,(1,0)]
world_pts8x, nearest8x = get_nearest_neighbour(pc, bev_shape[1] // 8,
def is_box_valid(limit, box_bev_pts):
(x1, y1), (x2, y2) = limit
limit_p = Polygon([(0, 0), (x1, y1), (x2, y2)])
box_p = Polygon([(box_bev_pts[0][0], box_bev_pts[0][1]),
(box_bev_pts[1][0], box_bev_pts[1][1]),
(box_bev_pts[2][0], box_bev_pts[2][1]),
(box_bev_pts[3][0], box_bev_pts[3][1])])
if limit_p.intersection(box_p).area > 0:
return False
return True
t = '000013' #ids[4]
boxes = reader.get_boxes_3D(t)
pts, ref = reader.get_velo(t, use_fov_filter=False)
# open3d(pts, boxes)
# imshow(bev(pts, pred_boxes=boxes, title="GT"))
lidar_src = (0, 0)
border_1 = (70, 40), (70, -40)
border_2 = np.array([40, 0]), np.array([40, 70])
border_3 = np.array([-40, 0]), np.array([-40, 70])
circles = []
d1_clr, d2_clr = (255, 0, 0), (0, 255, 0)
box_limits = []
all_limits = []
rand_num = 30
random_cars_dir = 'data_utils/aug_utils/annotations/cars/'
cur_metrics = compute_metric(model, pc, (obj, geo), True)
for metric_name, metric_val in cur_metrics.items():
with file_writer.as_default():
tf.summary.scalar(metric_name,
data=metric_val,
step=train_steps)
if train_steps % configs['vis_every'] is 0 and train_steps is not 0:
for var in model.trainable_variables:
with hist_writer.as_default():
tf.summary.histogram(var.name, var, step=train_steps)
if train_steps % configs['vis_every'] is 0 and train_steps is not 0:
for chosen_id in configs['chosen_ids']:
pts, _ = kitti.get_velo(chosen_id, use_fov_filter=False)
boxes = kitti.get_boxes_3D(chosen_id)
if pts.shape[1] != 3:
pts = pts.T
pc = np.expand_dims(pc_encoder.encode(pts), axis=0)
tar = target_encoder.encode(boxes)
obj_map, obj_mask = tar[0][..., 0], tar[0][..., 1]
obj_map = np.expand_dims(np.expand_dims(np.squeeze(obj_map),
axis=0),
axis=-1)
obj_mask = np.expand_dims(np.expand_dims(np.squeeze(obj_mask),
axis=0),
axis=-1)
with img_writer.as_default():
import deepdish as dd
from core.kitti import KITTI, ALL_OBJECTS, CARS_ONLY
from data_utils.augmentation import PointCloudAugmenter
DS_DIR = '/home/salam/datasets/KITTI/training'
reader = KITTI(DS_DIR, CARS_ONLY)
ids = reader.get_ids('train')
for id in ids:
boxes = reader.get_boxes_3D(id)
pts, _ = reader.get_velo(id, use_fov_filter=False)
for i, box in enumerate(boxes):
pts_inside_ids = PointCloudAugmenter.find_containing_points(box, pts.T)
if len(pts_inside_ids) > 5:
pts_inside = pts[:, pts_inside_ids].T
data = {
'frame_id': id,
'num_points': len(pts_inside_ids),
'pts': pts_inside,
'box_x': box.x,
'box_y': box.y,
'box_z': box.z,
'box_w': box.w,
'box_l': box.l,
'box_h': box.h,