File size: 8,677 Bytes
404d2af |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from .geom import rnd_sample, interpolate
class MultiSampler (nn.Module):
""" Similar to NghSampler, but doesnt warp the 2nd image.
Distance to GT => 0 ... pos_d ... neg_d ... ngh
Pixel label => + + + + + + 0 0 - - - - - - -
Subsample on query side: if > 0, regular grid
< 0, random points
In both cases, the number of query points is = W*H/subq**2
"""
def __init__(self, ngh, subq=1, subd=1, pos_d=0, neg_d=2, border=None,
maxpool_pos=True, subd_neg=0):
nn.Module.__init__(self)
assert 0 <= pos_d < neg_d <= (ngh if ngh else 99)
self.ngh = ngh
self.pos_d = pos_d
self.neg_d = neg_d
assert subd <= ngh or ngh == 0
assert subq != 0
self.sub_q = subq
self.sub_d = subd
self.sub_d_neg = subd_neg
if border is None: border = ngh
assert border >= ngh, 'border has to be larger than ngh'
self.border = border
self.maxpool_pos = maxpool_pos
self.precompute_offsets()
def precompute_offsets(self):
pos_d2 = self.pos_d**2
neg_d2 = self.neg_d**2
rad2 = self.ngh**2
rad = (self.ngh//self.sub_d) * self.ngh # make an integer multiple
pos = []
neg = []
for j in range(-rad, rad+1, self.sub_d):
for i in range(-rad, rad+1, self.sub_d):
d2 = i*i + j*j
if d2 <= pos_d2:
pos.append( (i,j) )
elif neg_d2 <= d2 <= rad2:
neg.append( (i,j) )
self.register_buffer('pos_offsets', torch.LongTensor(pos).view(-1,2).t())
self.register_buffer('neg_offsets', torch.LongTensor(neg).view(-1,2).t())
def forward(self, feat0, feat1, noise_feat0, noise_feat1, conf0, conf1, noise_conf0, noise_conf1, pos0, pos1, B, H, W, N=2500):
pscores_ls, nscores_ls, distractors_ls = [], [], []
valid_feat0_ls = []
noise_pscores_ls, noise_nscores_ls, noise_distractors_ls = [], [], []
valid_noise_feat0_ls = []
valid_pos1_ls, valid_pos2_ls = [], []
qconf_ls = []
noise_qconf_ls = []
mask_ls = []
for i in range(B):
tmp_mask = (pos0[i][:, 1] >= self.border) * (pos0[i][:, 1] < W-self.border) \
* (pos0[i][:, 0] >= self.border) * (pos0[i][:, 0] < H-self.border)
selected_pos0 = pos0[i][tmp_mask]
selected_pos1 = pos1[i][tmp_mask]
valid_pos0, valid_pos1 = rnd_sample([selected_pos0, selected_pos1], N)
# sample features from first image
valid_feat0 = interpolate(valid_pos0 / 4, feat0[i]) # [N, 128]
valid_feat0 = F.normalize(valid_feat0, p=2, dim=-1) # [N, 128]
qconf = interpolate(valid_pos0 / 4, conf0[i])
valid_noise_feat0 = interpolate(valid_pos0 / 4, noise_feat0[i]) # [N, 128]
valid_noise_feat0 = F.normalize(valid_noise_feat0, p=2, dim=-1) # [N, 128]
noise_qconf = interpolate(valid_pos0 / 4, noise_conf0[i])
# sample GT from second image
mask = (valid_pos1[:, 1] >= 0) * (valid_pos1[:, 1] < W) \
* (valid_pos1[:, 0] >= 0) * (valid_pos1[:, 0] < H)
def clamp(xy):
xy = xy
torch.clamp(xy[0], 0, H-1, out=xy[0])
torch.clamp(xy[1], 0, W-1, out=xy[1])
return xy
# compute positive scores
valid_pos1p = clamp(valid_pos1.t()[:,None,:] + self.pos_offsets[:,:,None].to(valid_pos1.device)) # [2, 29, N]
valid_pos1p = valid_pos1p.permute(1, 2, 0).reshape(-1, 2) # [29, N, 2] -> [29*N, 2]
valid_feat1p = interpolate(valid_pos1p / 4, feat1[i]).reshape(self.pos_offsets.shape[-1], -1, 128) # [29, N, 128]
valid_feat1p = F.normalize(valid_feat1p, p=2, dim=-1) # [29, N, 128]
valid_noise_feat1p = interpolate(valid_pos1p / 4, feat1[i]).reshape(self.pos_offsets.shape[-1], -1, 128) # [29, N, 128]
valid_noise_feat1p = F.normalize(valid_noise_feat1p, p=2, dim=-1) # [29, N, 128]
pscores = (valid_feat0[None,:,:] * valid_feat1p).sum(dim=-1).t() # [N, 29]
pscores, pos = pscores.max(dim=1, keepdim=True)
sel = clamp(valid_pos1.t() + self.pos_offsets[:,pos.view(-1)].to(valid_pos1.device))
qconf = (qconf + interpolate(sel.t() / 4, conf1[i]))/2
noise_pscores = (valid_noise_feat0[None,:,:] * valid_noise_feat1p).sum(dim=-1).t() # [N, 29]
noise_pscores, noise_pos = noise_pscores.max(dim=1, keepdim=True)
noise_sel = clamp(valid_pos1.t() + self.pos_offsets[:,noise_pos.view(-1)].to(valid_pos1.device))
noise_qconf = (noise_qconf + interpolate(noise_sel.t() / 4, noise_conf1[i]))/2
# compute negative scores
valid_pos1n = clamp(valid_pos1.t()[:,None,:] + self.neg_offsets[:,:,None].to(valid_pos1.device)) # [2, 29, N]
valid_pos1n = valid_pos1n.permute(1, 2, 0).reshape(-1, 2) # [29, N, 2] -> [29*N, 2]
valid_feat1n = interpolate(valid_pos1n / 4, feat1[i]).reshape(self.neg_offsets.shape[-1], -1, 128) # [29, N, 128]
valid_feat1n = F.normalize(valid_feat1n, p=2, dim=-1) # [29, N, 128]
nscores = (valid_feat0[None,:,:] * valid_feat1n).sum(dim=-1).t() # [N, 29]
valid_noise_feat1n = interpolate(valid_pos1n / 4, noise_feat1[i]).reshape(self.neg_offsets.shape[-1], -1, 128) # [29, N, 128]
valid_noise_feat1n = F.normalize(valid_noise_feat1n, p=2, dim=-1) # [29, N, 128]
noise_nscores = (valid_noise_feat0[None,:,:] * valid_noise_feat1n).sum(dim=-1).t() # [N, 29]
if self.sub_d_neg:
valid_pos2 = rnd_sample([selected_pos1], N)[0]
distractors = interpolate(valid_pos2 / 4, feat1[i])
distractors = F.normalize(distractors, p=2, dim=-1)
noise_distractors = interpolate(valid_pos2 / 4, noise_feat1[i])
noise_distractors = F.normalize(noise_distractors, p=2, dim=-1)
pscores_ls.append(pscores)
nscores_ls.append(nscores)
distractors_ls.append(distractors)
valid_feat0_ls.append(valid_feat0)
noise_pscores_ls.append(noise_pscores)
noise_nscores_ls.append(noise_nscores)
noise_distractors_ls.append(noise_distractors)
valid_noise_feat0_ls.append(valid_noise_feat0)
valid_pos1_ls.append(valid_pos1)
valid_pos2_ls.append(valid_pos2)
qconf_ls.append(qconf)
noise_qconf_ls.append(noise_qconf)
mask_ls.append(mask)
N = np.min([len(i) for i in qconf_ls])
# merge batches
qconf = torch.stack([i[:N] for i in qconf_ls], dim=0).squeeze(-1)
mask = torch.stack([i[:N] for i in mask_ls], dim=0)
pscores = torch.cat([i[:N] for i in pscores_ls], dim=0)
nscores = torch.cat([i[:N] for i in nscores_ls], dim=0)
distractors = torch.cat([i[:N] for i in distractors_ls], dim=0)
valid_feat0 = torch.cat([i[:N] for i in valid_feat0_ls], dim=0)
valid_pos1 = torch.cat([i[:N] for i in valid_pos1_ls], dim=0)
valid_pos2 = torch.cat([i[:N] for i in valid_pos2_ls], dim=0)
noise_qconf = torch.stack([i[:N] for i in noise_qconf_ls], dim=0).squeeze(-1)
noise_pscores = torch.cat([i[:N] for i in noise_pscores_ls], dim=0)
noise_nscores = torch.cat([i[:N] for i in noise_nscores_ls], dim=0)
noise_distractors = torch.cat([i[:N] for i in noise_distractors_ls], dim=0)
valid_noise_feat0 = torch.cat([i[:N] for i in valid_noise_feat0_ls], dim=0)
# remove scores that corresponds to positives or nulls
dscores = torch.matmul(valid_feat0, distractors.t())
noise_dscores = torch.matmul(valid_noise_feat0, noise_distractors.t())
dis2 = (valid_pos2[:, 1] - valid_pos1[:, 1][:,None])**2 + (valid_pos2[:, 0] - valid_pos1[:, 0][:,None])**2
b = torch.arange(B, device=dscores.device)[:,None].expand(B, N).reshape(-1)
dis2 += (b != b[:,None]).long() * self.neg_d**2
dscores[dis2 < self.neg_d**2] = 0
noise_dscores[dis2 < self.neg_d**2] = 0
scores = torch.cat((pscores, nscores, dscores), dim=1)
noise_scores = torch.cat((noise_pscores, noise_nscores, noise_dscores), dim=1)
gt = scores.new_zeros(scores.shape, dtype=torch.uint8)
gt[:, :pscores.shape[1]] = 1
return scores, noise_scores, gt, mask, qconf, noise_qconf
|