def transform_points_to_voxels(self,
data_dict=None,
config=None,
voxel_generator=None):
if data_dict is None:
from spconv.utils import VoxelGenerator
voxel_generator = VoxelGenerator(
voxel_size=config.VOXEL_SIZE,
point_cloud_range=self.point_cloud_range,
max_num_points=config.MAX_POINTS_PER_VOXEL,
max_voxels=config.MAX_NUMBER_OF_VOXELS[self.mode])
grid_size = (self.point_cloud_range[3:6] -
self.point_cloud_range[0:3]) / np.array(
config.VOXEL_SIZE)
self.grid_size = np.round(grid_size).astype(np.int64)
self.voxel_size = config.VOXEL_SIZE
return partial(self.transform_points_to_voxels,
voxel_generator=voxel_generator)
points = data_dict['points']
voxels, coordinates, num_points = voxel_generator.generate(points)
if not data_dict['use_lead_xyz']:
voxels = voxels[..., 3:] # remove xyz in voxels(N, 3)
data_dict['voxels'] = voxels
data_dict['voxel_coords'] = coordinates
data_dict['voxel_num_points'] = num_points
return data_dict
def noise_robustness(batch_size, ret):
"""
Add noise points around every ground truth box for SECOND and Point-RCNN.
The idea is proposed in TANet (AAAI 2020), https://arxiv.org/pdf/1912.05163.pdf.
How to use:
For now, this function should be added to line 179 of pcdet/datasets/dataset.py
TODO: Integrate with pcdet/datasets/dataset.py
Change no_of_pts to number of uniformly generated noise points desired
around each GT bounding box.
:param batch_size:
:param ret:
:return:
"""
if True:
np.random.seed(0) # Numpy module.
voxels_flag = True
no_of_pts = 100
voxel_list = []
coords_list = []
num_points_list = []
for k in range(batch_size):
for i_box in range(len(ret['gt_boxes'][k])):
bbox = ret['gt_boxes'][k][i_box]
cx = bbox[0]
cy = bbox[1]
cz = bbox[2]
l = bbox[3]
w = bbox[4]
h = bbox[5]
z1 = np.random.uniform(cx + l // 2, cx + 3 * l,
(no_of_pts // 2, 1))
z2 = np.random.uniform(cx - l // 2, cx - 3 * l,
(no_of_pts // 2, 1))
z = np.concatenate([z1, z2], 0)
y1 = np.random.uniform(cy + w // 2, cy + 3 * w,
(no_of_pts // 2, 1))
y2 = np.random.uniform(cy - w // 2, cy - 3 * w,
(no_of_pts // 2, 1))
y = np.concatenate([y1, y2], 0)
x1 = np.random.uniform(cz + h // 2, cz + 3 * h,
(no_of_pts // 2, 1))
x2 = np.random.uniform(cz - h // 2, cz - 3 * h,
(no_of_pts // 2, 1))
x = np.concatenate([x1, x2], 0)
r = np.ones([no_of_pts, 1])
b = np.zeros([no_of_pts, 1]) + k
noise = np.concatenate([b, z, y, x, r], 1)
ret['points'] = np.concatenate([ret['points'], noise], 0)
if voxels_flag:
voxel_generator = VoxelGenerator(
voxel_size=[0.05, 0.05, 0.1],
point_cloud_range=[0, -40, -3, 70.4, 40, 1],
max_num_points=5,
max_voxels=40000,
)
batch_mask = ret['points'][:, 0] == k
points_extract = ret['points'][batch_mask, :]
voxels, coordinates, num_points = voxel_generator.generate(
points_extract[:, 1:])
voxel_list.append(voxels)
coords_list.append(coordinates)
num_points_list.append(num_points)
if voxels_flag:
ret['voxels'] = np.concatenate(voxel_list, axis=0)
ret['voxel_num_points'] = np.concatenate(num_points_list, axis=0)
coors = []
for i, coor in enumerate(coords_list):
coor_pad = np.pad(coor, ((0, 0), (1, 0)),
mode='constant',
constant_values=i)
coors.append(coor_pad)
ret['voxel_coords'] = np.concatenate(coors, axis=0)
return ret