File size: 6,624 Bytes
33569f9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 | # Functions for 1D NMS, modified from:
# https://github.com/open-mmlab/mmcv/blob/master/mmcv/ops/nms.py
import torch
import nms_1d_cpu
class NMSop(torch.autograd.Function):
@staticmethod
def forward(
ctx, segs, scores, cls_idxs,
iou_threshold, min_score, max_num
):
# vanilla nms will not change the score, so we can filter segs first
is_filtering_by_score = (min_score > 0)
if is_filtering_by_score:
valid_mask = scores > min_score
segs, scores = segs[valid_mask], scores[valid_mask]
cls_idxs = cls_idxs[valid_mask]
valid_inds = torch.nonzero(
valid_mask, as_tuple=False).squeeze(dim=1)
# nms op; return inds that is sorted by descending order
inds = nms_1d_cpu.nms(
segs.contiguous().cpu(),
scores.contiguous().cpu(),
iou_threshold=float(iou_threshold))
# cap by max number
if max_num > 0:
inds = inds[:min(max_num, len(inds))]
# return the sorted segs / scores
sorted_segs = segs[inds]
sorted_scores = scores[inds]
sorted_cls_idxs = cls_idxs[inds]
return sorted_segs.clone(), sorted_scores.clone(), sorted_cls_idxs.clone()
class SoftNMSop(torch.autograd.Function):
@staticmethod
def forward(
ctx, segs, scores, cls_idxs,
iou_threshold, sigma, min_score, method, max_num
):
# pre allocate memory for sorted results
dets = segs.new_empty((segs.size(0), 3), device='cpu')
# softnms op, return dets that stores the sorted segs / scores
inds = nms_1d_cpu.softnms(
segs.cpu(),
scores.cpu(),
dets.cpu(),
iou_threshold=float(iou_threshold),
sigma=float(sigma),
min_score=float(min_score),
method=int(method))
# cap by max number
if max_num > 0:
n_segs = min(len(inds), max_num)
else:
n_segs = len(inds)
sorted_segs = dets[:n_segs, :2]
sorted_scores = dets[:n_segs, 2]
sorted_cls_idxs = cls_idxs[inds]
sorted_cls_idxs = sorted_cls_idxs[:n_segs]
return sorted_segs.clone(), sorted_scores.clone(), sorted_cls_idxs.clone()
def seg_voting(nms_segs, all_segs, all_scores, iou_threshold, score_offset=1.5):
"""
blur localization results by incorporating side segs.
this is known as bounding box voting in object detection literature.
slightly boost the performance around iou_threshold
"""
# *_segs : N_i x 2, all_scores: N,
# apply offset
offset_scores = all_scores + score_offset
# computer overlap between nms and all segs
# construct the distance matrix of # N_nms x # N_all
num_nms_segs, num_all_segs = nms_segs.shape[0], all_segs.shape[0]
ex_nms_segs = nms_segs[:, None].expand(num_nms_segs, num_all_segs, 2)
ex_all_segs = all_segs[None, :].expand(num_nms_segs, num_all_segs, 2)
# compute intersection
left = torch.maximum(ex_nms_segs[:, :, 0], ex_all_segs[:, :, 0])
right = torch.minimum(ex_nms_segs[:, :, 1], ex_all_segs[:, :, 1])
inter = (right-left).clamp(min=0)
# lens of all segments
nms_seg_lens = ex_nms_segs[:, :, 1] - ex_nms_segs[:, :, 0]
all_seg_lens = ex_all_segs[:, :, 1] - ex_all_segs[:, :, 0]
# iou
iou = inter / (nms_seg_lens + all_seg_lens - inter)
# get neighbors (# N_nms x # N_all) / weights
seg_weights = (iou >= iou_threshold).to(all_scores.dtype) * all_scores[None, :] * iou
seg_weights /= torch.sum(seg_weights, dim=1, keepdim=True)
refined_segs = seg_weights @ all_segs
return refined_segs
def batched_nms(
segs,
scores,
cls_idxs,
iou_threshold,
min_score,
max_seg_num,
use_soft_nms=True,
multiclass=True,
sigma=0.5,
voting_thresh=0.75,
):
# Based on Detectron2 implementation,
num_segs = segs.shape[0]
# corner case, no prediction outputs
if num_segs == 0:
return torch.zeros([0, 2]),\
torch.zeros([0,]),\
torch.zeros([0,], dtype=cls_idxs.dtype)
if multiclass:
# multiclass nms: apply nms on each class independently
new_segs, new_scores, new_cls_idxs = [], [], []
for class_id in torch.unique(cls_idxs):
curr_indices = torch.where(cls_idxs == class_id)[0]
# soft_nms vs nms
if use_soft_nms:
sorted_segs, sorted_scores, sorted_cls_idxs = SoftNMSop.apply(
segs[curr_indices],
scores[curr_indices],
cls_idxs[curr_indices],
iou_threshold,
sigma,
min_score,
2,
max_seg_num
)
else:
sorted_segs, sorted_scores, sorted_cls_idxs = NMSop.apply(
segs[curr_indices],
scores[curr_indices],
cls_idxs[curr_indices],
iou_threshold,
min_score,
max_seg_num
)
# disable seg voting for multiclass nms, no sufficient segs
# fill in the class index
new_segs.append(sorted_segs)
new_scores.append(sorted_scores)
new_cls_idxs.append(sorted_cls_idxs)
# cat the results
new_segs = torch.cat(new_segs)
new_scores = torch.cat(new_scores)
new_cls_idxs = torch.cat(new_cls_idxs)
else:
# class agnostic
if use_soft_nms:
new_segs, new_scores, new_cls_idxs = SoftNMSop.apply(
segs, scores, cls_idxs, iou_threshold,
sigma, min_score, 2, max_seg_num
)
else:
new_segs, new_scores, new_cls_idxs = NMSop.apply(
segs, scores, cls_idxs, iou_threshold,
min_score, max_seg_num
)
# seg voting
if voting_thresh > 0:
new_segs = seg_voting(
new_segs,
segs,
scores,
voting_thresh
)
# sort based on scores and return
# truncate the results based on max_seg_num
_, idxs = new_scores.sort(descending=True)
max_seg_num = min(max_seg_num, new_segs.shape[0])
# needed for multiclass NMS
new_segs = new_segs[idxs[:max_seg_num]]
new_scores = new_scores[idxs[:max_seg_num]]
new_cls_idxs = new_cls_idxs[idxs[:max_seg_num]]
return new_segs, new_scores, new_cls_idxs
|