mmrotate.core.bbox.assigners.atss_kld_assigner 源代码
# Copyright (c) OpenMMLab. All rights reserved.
import torch
from mmcv.ops import points_in_polygons
from mmdet.core.bbox.assigners.assign_result import AssignResult
from mmdet.core.bbox.assigners.base_assigner import BaseAssigner
from mmrotate.core.bbox.utils import GaussianMixture
from ..builder import ROTATED_BBOX_ASSIGNERS
from ..transforms import gt2gaussian
[文档]@ROTATED_BBOX_ASSIGNERS.register_module()
class ATSSKldAssigner(BaseAssigner):
"""Assign a corresponding gt bbox or background to each bbox.
Each proposals will be assigned with `0` or a positive integer
indicating the ground truth index.
- 0: negative sample, no assigned gt
- positive integer: positive sample, index (1-based) of assigned gt
Args:
topk (float): Number of bbox selected in each level.
use_reassign (bool, optional): If true, it is used to reassign samples.
"""
def __init__(self, topk, use_reassign=False):
self.topk = topk
self.use_reassign = use_reassign
[文档] def assign(self,
bboxes,
num_level_bboxes,
gt_bboxes,
gt_bboxes_ignore=None,
gt_labels=None):
"""Assign gt to bboxes.
The assignment is done in following steps
1. compute iou between all bbox (bbox of all pyramid levels) and gt
2. compute center distance between all bbox and gt
3. on each pyramid level, for each gt, select k bbox whose center
are closest to the gt center, so we total select k*l bbox as
candidates for each gt
4. get corresponding iou for the these candidates, and compute the
mean and std, set mean + std as the iou threshold
5. compute the mean aspect ratio of all gts, and set
exp((-mean aspect ratio / 4) * (mean + std) as the iou threshold
6. select these candidates whose iou are greater than or equal to
the threshold as positive
7. limit the positive sample's center in gt
Args:
bboxes (Tensor): Bounding boxes to be assigned, shape(n, 4).
num_level_bboxes (List): num of bboxes in each level
gt_bboxes (Tensor): Groundtruth boxes, shape (k, 4).
gt_bboxes_ignore (Tensor, optional): Ground truth bboxes that are
labelled as `ignored`, e.g., crowd boxes in COCO.
gt_labels (Tensor, optional): Label of gt_bboxes, shape (k, ).
Returns:
:obj:`AssignResult`: The assign result.
"""
INF = 100000000
num_gt, num_bboxes = gt_bboxes.size(0), bboxes.size(0)
overlaps = self.kld_overlaps(gt_bboxes, bboxes)
overlaps = overlaps.transpose(1, 0)
# assign 0 by default
assigned_gt_inds = overlaps.new_full((num_bboxes, ),
0,
dtype=torch.long)
if num_gt == 0 or num_bboxes == 0:
# No ground truth or boxes, return empty assignment
max_overlaps = overlaps.new_zeros((num_bboxes, ))
if num_gt == 0:
# No truth, assign everything to background
assigned_gt_inds[:] = 0
if gt_labels is None:
assigned_labels = None
else:
assigned_labels = overlaps.new_full((num_bboxes, ),
-1,
dtype=torch.long)
return AssignResult(
num_gt, assigned_gt_inds, max_overlaps, labels=assigned_labels)
# compute center distance between all bbox and gt
# the center of poly
gt_bboxes_hbb = self.get_horizontal_bboxes(gt_bboxes)
gt_cx = (gt_bboxes_hbb[:, 0] + gt_bboxes_hbb[:, 2]) / 2.0
gt_cy = (gt_bboxes_hbb[:, 1] + gt_bboxes_hbb[:, 3]) / 2.0
gt_points = torch.stack((gt_cx, gt_cy), dim=1)
bboxes = bboxes.reshape(-1, 9, 2)
# y_first False
pts_x = bboxes[:, :, 0::2]
pts_y = bboxes[:, :, 1::2]
pts_x_mean = pts_x.mean(dim=1).squeeze()
pts_y_mean = pts_y.mean(dim=1).squeeze()
bboxes_points = torch.stack((pts_x_mean, pts_y_mean), dim=1)
distances = (bboxes_points[:, None, :] -
gt_points[None, :, :]).pow(2).sum(-1).sqrt()
# Selecting candidates based on the center distance
candidate_idxs = []
start_idx = 0
for level, bboxes_per_level in enumerate(num_level_bboxes):
# on each pyramid level, for each gt,
# select k bbox whose center are closest to the gt center
end_idx = start_idx + bboxes_per_level
distances_per_level = distances[start_idx:end_idx, :]
_, topk_idxs_per_level = distances_per_level.topk(
self.topk, dim=0, largest=False)
candidate_idxs.append(topk_idxs_per_level + start_idx)
start_idx = end_idx
candidate_idxs = torch.cat(candidate_idxs, dim=0)
# max_overlaps, argmax_overlaps = overlaps.max(dim=0)
# gt_max_overlaps, gt_argmax_overlaps = overlaps.max(dim=1)
# get corresponding iou for the these candidates, and compute the
# mean and std, set mean + std as the iou threshold
# add
gt_bboxes_ratios = self.AspectRatio(gt_bboxes)
gt_bboxes_ratios_per_gt = gt_bboxes_ratios.mean(0)
candidate_overlaps = overlaps[candidate_idxs, torch.arange(num_gt)]
overlaps_mean_per_gt = candidate_overlaps.mean(0)
overlaps_std_per_gt = candidate_overlaps.std(0)
overlaps_thr_per_gt = overlaps_mean_per_gt + overlaps_std_per_gt
if self.use_reassign:
iou_thr_weight = torch.exp((-1 / 4) * gt_bboxes_ratios_per_gt)
overlaps_thr_per_gt = overlaps_thr_per_gt * iou_thr_weight
is_pos = candidate_overlaps >= overlaps_thr_per_gt[None, :]
# inside_flag = torch.full([num_bboxes, num_gt],
# 0.).to(gt_bboxes.device).float()
# pointsJf(bboxes_points, gt_bboxes, inside_flag)
inside_flag = points_in_polygons(bboxes_points, gt_bboxes)
is_in_gts = inside_flag[candidate_idxs,
torch.arange(num_gt)].to(is_pos.dtype)
is_pos = is_pos & is_in_gts
for gt_idx in range(num_gt):
candidate_idxs[:, gt_idx] += gt_idx * num_bboxes
candidate_idxs = candidate_idxs.view(-1)
# if an anchor box is assigned to multiple gts,
# the one with the highest IoU will be selected.
overlaps_inf = torch.full_like(overlaps,
-INF).t().contiguous().view(-1)
index = candidate_idxs.view(-1)[is_pos.view(-1)]
overlaps_inf[index] = overlaps.t().contiguous().view(-1)[index]
overlaps_inf = overlaps_inf.view(num_gt, -1).t()
max_overlaps, argmax_overlaps = overlaps_inf.max(dim=1)
assigned_gt_inds[
max_overlaps != -INF] = argmax_overlaps[max_overlaps != -INF] + 1
if gt_labels is not None:
assigned_labels = assigned_gt_inds.new_full((num_bboxes, ), -1)
pos_inds = torch.nonzero(
assigned_gt_inds > 0, as_tuple=False).squeeze()
if pos_inds.numel() > 0:
assigned_labels[pos_inds] = gt_labels[
assigned_gt_inds[pos_inds] - 1]
else:
assigned_labels = None
return AssignResult(
num_gt, assigned_gt_inds, max_overlaps, labels=assigned_labels)
[文档] def kld_overlaps(self, gt_rbboxes, points, eps=1e-6):
"""Compute overlaps between polygons and points by Kullback-Leibler
Divergence loss.
Args:
gt_rbboxes (torch.Tensor): Ground truth polygons, shape (k, 8).
points (torch.Tensor): Points to be assigned, shape(n, 18).
eps (float, optional): Defaults to 1e-6.
Returns:
Tensor: Kullback-Leibler Divergence loss.
"""
points = points.reshape(-1, 9, 2)
gt_rbboxes = gt_rbboxes.reshape(-1, 4, 2)
gmm = GaussianMixture(n_components=1)
gmm.fit(points)
kld = self.kld_mixture2single(gmm, gt2gaussian(gt_rbboxes))
kl_agg = kld.clamp(min=eps)
overlaps = 1 / (2 + kl_agg)
return overlaps
[文档] def kld_mixture2single(self, g1, g2):
"""Compute Kullback-Leibler Divergence between two Gaussian
distribution.
Args:
g1 (dict[str, torch.Tensor]): Gaussian distribution 1.
g2 (torch.Tensor): Gaussian distribution 2.
Returns:
torch.Tensor: Kullback-Leibler Divergence.
"""
if g2[0].shape[0] == 0:
return g2[0].new_zeros((0, 1))
p_mu = g1.mu
p_var = g1.var
assert p_mu.dim() == 3 and p_mu.size()[1] == 1
assert p_var.dim() == 4 and p_var.size()[1] == 1
N, _, d = p_mu.shape
p_mu = p_mu.reshape(1, N, d)
p_var = p_var.reshape(1, N, d, d)
t_mu, t_var = g2
K = t_mu.shape[0]
t_mu = t_mu.unsqueeze(1)
t_var = t_var.unsqueeze(1)
delta = (p_mu - t_mu).unsqueeze(-1)
t_inv = torch.inverse(t_var)
term1 = delta.transpose(-1, -2).matmul(t_inv).matmul(delta)\
.reshape(K, N)
term2 = torch.diagonal(t_inv.matmul(p_var), dim1=-2, dim2=-1)\
.sum(dim=-1) + torch.log(torch.det(t_var) / torch.det(p_var))
return 0.5 * (term1 + term2) - 1
[文档] def get_horizontal_bboxes(self, gt_rbboxes):
"""get_horizontal_bboxes from polygons.
Args:
gt_rbboxes (torch.Tensor): Groundtruth polygons, shape (k, 8).
Returns:
gt_rect_bboxes (torch.Tensor): The horizontal bboxes, shape (k, 4).
"""
gt_xs, gt_ys = gt_rbboxes[:, 0::2], gt_rbboxes[:, 1::2]
gt_xmin, _ = gt_xs.min(1)
gt_ymin, _ = gt_ys.min(1)
gt_xmax, _ = gt_xs.max(1)
gt_ymax, _ = gt_ys.max(1)
gt_rect_bboxes = torch.cat([
gt_xmin[:, None], gt_ymin[:, None], gt_xmax[:, None], gt_ymax[:,
None]
],
dim=1)
return gt_rect_bboxes
[文档] def AspectRatio(self, gt_rbboxes):
"""compute the aspect ratio of all gts.
Args:
gt_rbboxes (torch.Tensor): Groundtruth polygons, shape (k, 8).
Returns:
ratios (torch.Tensor): The aspect ratio of gt_rbboxes,
shape (k, 1).
"""
pt1, pt2, pt3, pt4 = gt_rbboxes[..., :8].chunk(4, 1)
edge1 = torch.sqrt(
torch.pow(pt1[..., 0] - pt2[..., 0], 2) +
torch.pow(pt1[..., 1] - pt2[..., 1], 2))
edge2 = torch.sqrt(
torch.pow(pt2[..., 0] - pt3[..., 0], 2) +
torch.pow(pt2[..., 1] - pt3[..., 1], 2))
edges = torch.stack([edge1, edge2], dim=1)
width, _ = torch.max(edges, 1)
height, _ = torch.min(edges, 1)
ratios = width / height
return ratios