def _generate(self,
feature_map_shape,
scales,
aspect_ratios,
input_size,
device='cuda'):
"""
"""
# shape(A,)
ratios_sqrt = torch.sqrt(aspect_ratios)
heights = scales * ratios_sqrt * self.base_anchor_size.to(device)
widths = scales / ratios_sqrt * self.base_anchor_size.to(device)
anchor_stride = [
input_size[0] / feature_map_shape[0],
input_size[1] / feature_map_shape[1]
]
y_ctrs = torch.arange(
feature_map_shape[0], device=device
) * anchor_stride[0] + self.anchor_offset[0].to(device).float()
x_ctrs = torch.arange(
feature_map_shape[1], device=device
) * anchor_stride[1] + self.anchor_offset[1].to(device)
y_ctrs = y_ctrs.float()
x_ctrs = x_ctrs.float()
# meshgrid
# shape(H*W,)
x_ctrs, y_ctrs = ops.meshgrid(x_ctrs, y_ctrs)
# shape(K*A,)
heights, x_ctrs = ops.meshgrid(heights, x_ctrs)
widths, y_ctrs = ops.meshgrid(widths, y_ctrs)
xmin = x_ctrs - 0.5 * (widths - 1)
ymin = y_ctrs - 0.5 * (heights - 1)
xmax = x_ctrs + 0.5 * (widths - 1)
ymax = y_ctrs + 0.5 * (heights - 1)
# (x,y,h,w)
return torch.stack([xmin, ymin, xmax, ymax], dim=1)
def generate_voxels(self):
"""
generate all voxels in ground plane
"""
lattice_dims = self.grid_dims / self.voxel_size
self.lattice_dims = lattice_dims
x_inds = torch.arange(0, lattice_dims[0]).cuda()
y_inds = torch.arange(0, lattice_dims[1]).cuda()
z_inds = torch.arange(0, lattice_dims[2]).cuda()
z_inds, x_inds = ops.meshgrid(z_inds, x_inds)
y_inds1, z_inds = ops.meshgrid(y_inds, z_inds)
y_inds2, x_inds = ops.meshgrid(y_inds, x_inds)
y_inds = y_inds1
corner_coords = torch.stack([x_inds, y_inds, z_inds], dim=-1).float()
corner_coords *= self.voxel_size
center_offset = torch.tensor([0.5 * self.voxel_size] *
3).type_as(corner_coords)
center_coords = corner_coords + center_offset
high_interval = self.high_interval
# bugs here
# y_offset = (high_interval[0] + high_interval[1]) / 2
y_offset = high_interval[0]
original_offset = [-0.5 * self.grid_dims[0], y_offset, self.z_offset]
# original_offset = self.original_offset
original_offset = torch.tensor(original_offset).type_as(center_coords)
center_coords = center_coords + original_offset
# tmp_coords = torch.tensor(
# [[-16.53, 2.39, 58.49]]).type_as(center_coords)
# center_coords = torch.cat([center_coords, tmp_coords])
self.voxel_centers = center_coords
return center_coords
def generate(self, feature_map_list, im_shape):
"""
Args:
feature_map_list, list of (stride, ratio)
Returns:
anchors
"""
anchors_list = []
scales, aspect_ratios = ops.meshgrid(self.scales, self.aspect_ratios)
for feature_map_shape in feature_map_list:
anchors_list.append(
self._generate(feature_map_shape, scales, aspect_ratios,
im_shape))
return torch.cat(anchors_list, dim=0)
def generate(self, feature_map_list, input_size, device='cuda'):
"""
Args:
feature_map_list, list of (stride, ratio)
Returns:
anchors
"""
anchors_list = []
scales, aspect_ratios = ops.meshgrid(self.scales.to(device),
self.aspect_ratios.to(device))
for feature_map_shape in feature_map_list:
anchors_list.append(
self._generate(feature_map_shape, scales, aspect_ratios,
input_size, device))
return torch.cat(anchors_list, dim=0)
def nms_map(self, smoothed_fg_mask):
"""
supress the neibor
"""
directions = [-1, 0, 1]
shape = smoothed_fg_mask.shape
orig_index = (torch.arange(shape[0]).cuda().long(),
torch.arange(shape[1]).cuda().long())
orig_index = ops.meshgrid(orig_index[1], orig_index[0])
orig_index = [orig_index[1], orig_index[0]]
dest_indexes = []
for i in directions:
for j in directions:
dest_index = (orig_index[0] + directions[i],
orig_index[1] + directions[j])
dest_indexes.append(dest_index)
nms_filter = torch.ones_like(smoothed_fg_mask).byte()
orig_fg_mask = smoothed_fg_mask
# pad fg mask first to prevent out of boundary
padded_smoothed_fg_mask = torch.zeros(
(shape[0] + 1, shape[1] + 1)).type_as(smoothed_fg_mask)
padded_smoothed_fg_mask[:-1, :-1] = smoothed_fg_mask
# import ipdb
# ipdb.set_trace()
for dest_index in dest_indexes:
nms_filter = nms_filter & (
orig_fg_mask >=
padded_smoothed_fg_mask[dest_index].view_as(orig_fg_mask))
# surpress
smoothed_fg_mask[~nms_filter] = 0
return smoothed_fg_mask
def _assign_classification_targets(self, match, gt_labels, inter_boxes,
proposals):
"""
Generate cls map for segmentation loss
Args:
pass
Returns:
cls_map_targets
"""
# hard code
pooling_size = 8
# get label for each bbox
batch_size = match.shape[0]
offset = torch.arange(0, batch_size) * gt_labels.size(1)
match += offset.view(batch_size, 1).type_as(match)
cls_targets_batch = gt_labels.view(-1)[match.view(-1)].view(
batch_size, match.shape[1])
# set bg
cls_targets_batch[match == -1] = 0
# generate map according to label for segmentation loss
h = proposals[:, :, 3] - proposals[:, :, 1]
w = proposals[:, :, 2] - proposals[:, :, 0]
sub_bin_w = w / pooling_size
sub_bin_h = h / pooling_size
# pass
offset_xmin = inter_boxes[:, :, 0] - proposals[:, :, 0]
offset_xmax = inter_boxes[:, :, 2] - proposals[:, :, 0]
offset_ymin = inter_boxes[:, :, 1] - proposals[:, :, 1]
offset_ymax = inter_boxes[:, :, 3] - proposals[:, :, 1]
offset_xmin_ind = (offset_xmin / sub_bin_w).round().int()
offset_ymin_ind = (offset_ymin / sub_bin_h).round().int()
offset_xmax_ind = (offset_xmax / sub_bin_w).round().int()
offset_ymax_ind = (offset_ymax / sub_bin_h).round().int()
# select not empty bbox from inter boxes
# cond = (inter_boxes[:, :, 2] - inter_boxes[:, :, 0] + 1 > 0) & (
# inter_boxes[:, :, 3] - inter_boxes[:, :, 1] + 1 > 0)
# import ipdb
# ipdb.set_trace()
num = pooling_size * pooling_size
offset_xmin_ind = offset_xmin_ind.unsqueeze(-1).expand(-1, -1, num)
offset_ymin_ind = offset_ymin_ind.unsqueeze(-1).expand(-1, -1, num)
offset_xmax_ind = offset_xmax_ind.unsqueeze(-1).expand(-1, -1, num)
offset_ymax_ind = offset_ymax_ind.unsqueeze(-1).expand(-1, -1, num)
x = torch.range(0, pooling_size - 1).type_as(offset_xmin_ind)
y = torch.range(0, pooling_size - 1).type_as(offset_xmin_ind)
xx, yy = ops.meshgrid(x, y)
coord = torch.stack([xx, yy], dim=-1)
coord = coord.expand(inter_boxes.shape[0], inter_boxes.shape[1], -1,
-1)
# shape(N,M,49)
cond = (coord[:, :, :, 0] <
offset_xmax_ind) & (coord[:, :, :, 0] >= offset_xmin_ind) & (
coord[:, :, :, 1] < offset_ymax_ind) & (coord[:, :, :, 1]
>= offset_ymin_ind)
# torch.nonzero(cond)
# cls_gate_map shape(N,M,49)
# cond = cond.view(cond.shape[0], cond.shape[1], pooling_size,
# pooling_size)
cls_gate_map = cond.int()
# reverse when bg
# cls_gate_map[cls_targets_batch == 0] = (
# 1 - cls_gate_map)[cls_targets_batch == 0]
return cls_gate_map.long()