def accuracy(data, out):
vgt = data[4]
#
ss_gt = vgt[:, 0:3]
sr1_gt = vgt[:, 3:6]
sr2_gt = vgt[:, 6:9]
sb_gt = sr2box(ss_gt, sr1_gt, sr2_gt)
#
ts_gt = vgt[:, 9:12]
tr1_gt = vgt[:, 15:18]
tr2_gt = vgt[:, 18:21]
tb_gt = sr2box(ts_gt, tr1_gt, tr2_gt)
#
ss = out['ss']
sr1 = out['sr1']
sr2 = out['sr2']
sb = out['sb']
#
ts = out['ts']
tr1 = out['tr1']
tr2 = out['tr2']
tb = out['tb']
#
loss = {}
loss['size'] = 0.5 * torch.sum(
(ss - ss_gt.data)**2 + (ts - ts_gt.data)**2, dim=1)
erot = (sr1 - sr1_gt.data)**2
erot += (sr2 - sr2_gt.data)**2
erot += (tr1 - tr1_gt.data)**2
erot += (tr2 - tr2_gt.data)**2
loss['rot6'] = 0.5 * torch.sum(erot, dim=1)
loss['box'] = 0.5 * (chmf(sb, sb_gt.data) + chmf(tb, tb_gt.data))
loss['overall'] = loss['box']
return loss
def accuracy(data,out):
vgt = data[5];
#
ss_gt = vgt[:,0:3].contiguous();
ss_gt = norm_size(ss_gt);
sr1_gt = vgt[:,3:6].contiguous();
sr2_gt = vgt[:,6:9].contiguous();
sb_gt = sr2box(ss_gt,sr1_gt,sr2_gt);
#
ts_gt = vgt[:,9:12].contiguous();
ts_gt = norm_size(ts_gt);
tr1_gt = vgt[:,15:18].contiguous();
tr2_gt = vgt[:,18:21].contiguous();
tb_gt = sr2box(ts_gt,tr1_gt,tr2_gt);
#
ss = out['ss'];
sr1 = out['sr1'];
sr2 = out['sr2'];
sb = out['sb'];
#
ts = out['ts'];
tr1 = out['tr1'];
tr2 = out['tr2'];
tb = out['tb'];
#
loss = {};
loss['size'] = 0.5*torch.sum( ( ss - ss_gt.data )**2 + ( ts - ts_gt.data )**2, dim = 1 );
erot = ( sr1 - sr1_gt.data )**2;
erot += ( sr2 - sr2_gt.data )**2;
erot += ( tr1 - tr1_gt.data )**2;
erot += ( tr2 - tr2_gt.data )**2;
loss['rot6'] = 0.5*torch.sum( erot, dim=1 );
loss['bcd'] = 0.5*( box_cd(ss,sr1,sr2,ss_gt,sr1_gt,sr2_gt) + box_cd(ts,tr1,tr2,ts_gt,tr1_gt,tr2_gt) );
loss['overall'] = loss['bcd'];
return loss;
def run(**kwargs):
nepoch = kwargs['nepoch']
bs = kwargs['batch_size']
size_gt = torch.from_numpy(
np.random.uniform(0.1, 1, [bs, 3]).astype(np.float32))
r1_gt = torch.from_numpy(
np.random.uniform(-1, 1, [bs, 3]).astype(np.float32))
r2_gt = torch.from_numpy(
np.random.uniform(-1, 1, [bs, 3]).astype(np.float32))
size = torch.zeros([bs, 3], requires_grad=True)
r1 = torch.zeros([bs, 3], requires_grad=True)
r2 = torch.zeros([bs, 3], requires_grad=True)
size.data.uniform_(0.1, 1)
r1.data.normal_(0.0, 1.0)
r2.data.normal_(0.0, 1.0)
opt = torch.optim.Adam([size, r1, r2], lr=0.001)
boxgt = sr2box(size_gt.cuda(), r1_gt.cuda(), r2_gt.cuda())
write_box('./log/boxcd/gt',
boxgt.data.cpu().numpy())
for i in range(nepoch):
opt.zero_grad()
cd = box_cd(size.cuda(), r1.cuda(), r2.cuda(), size_gt.cuda(),
r1_gt.cuda(), r2_gt.cuda())
loss = torch.mean(cd)
loss.backward()
print(loss.data.cpu().numpy())
opt.step()
boxout = sr2box(size.cuda(), r1.cuda(), r2.cuda())
if i % 100 == 0:
write_box('./log/boxcd/out_%d' % i,
boxout.data.cpu().numpy())
return
def accuracy(data, out):
vgt = data[5]
#
ss_gt = vgt[:, 0:3].contiguous()
srs_gt, _ = torch.max(ss_gt, dim=1, keepdim=True)
sr1_gt = vgt[:, 3:6].contiguous()
sr2_gt = vgt[:, 6:9].contiguous()
sb_gt = sr2box(ss_gt, sr1_gt, sr2_gt)
#
ts_gt = vgt[:, 9:12].contiguous()
trs_gt, _ = torch.max(ts_gt, dim=1, keepdim=True)
tr1_gt = vgt[:, 15:18].contiguous()
tr2_gt = vgt[:, 18:21].contiguous()
tb_gt = sr2box(ts_gt, tr1_gt, tr2_gt)
#
ss = out['ss']
sr1 = out['sr1']
sr2 = out['sr2']
sb = out['sb']
#
ts = out['ts']
tr1 = out['tr1']
tr2 = out['tr2']
tb = out['tb']
#
loss = {}
loss['size'] = 0.5 * torch.sum((ss * srs_gt.data - ss_gt.data)**2 +
(ts * trs_gt.data - ts_gt.data)**2,
dim=1)
erot = (sr1 - sr1_gt.data)**2
erot += (sr2 - sr2_gt.data)**2
erot += (tr1 - tr1_gt.data)**2
erot += (tr2 - tr2_gt.data)**2
loss['rot6'] = 0.5 * torch.sum(erot, dim=1)
loss['bcd'] = 0.5 * (
box_cd(ss * srs_gt.data, sr1, sr2, ss_gt, sr1_gt, sr2_gt) +
box_cd(ts * trs_gt.data, tr1, tr2, ts_gt, tr1_gt, tr2_gt))
loss['overall'] = loss['bcd']
return loss
def loss(data, out):
vgt = data[5]
#
ss_gt = vgt[:, 0:3].contiguous()
ss_gt = norm_size(ss_gt)
sr1_gt = vgt[:, 3:6].contiguous()
sr2_gt = vgt[:, 6:9].contiguous()
sb_gt = sr2box(ss_gt, sr1_gt, sr2_gt)
#
ts_gt = vgt[:, 9:12].contiguous()
ts_gt = norm_size(ts_gt)
tr1_gt = vgt[:, 15:18].contiguous()
tr2_gt = vgt[:, 18:21].contiguous()
tb_gt = sr2box(ts_gt, tr1_gt, tr2_gt)
#
ss = out['ss']
sr1 = out['sr1']
sr2 = out['sr2']
sb = out['sb']
#
ts = out['ts']
tr1 = out['tr1']
tr2 = out['tr2']
tb = out['tb']
#
loss = {}
loss['size'] = 0.5 * torch.sum(
(ss - ss_gt.data)**2 + (ts - ts_gt.data)**2, dim=1)
erot = (sr1 - sr1_gt.data)**2
erot += (sr2 - sr2_gt.data)**2
erot += (tr1 - tr1_gt.data)**2
erot += (tr2 - tr2_gt.data)**2
loss['rot6'] = 0.5 * torch.sum(erot, dim=1)
loss['bcd'] = 0.5 * (box_cd(ss, sr1, sr2, ss_gt, sr1_gt, sr2_gt) +
box_cd(ts, tr1, tr2, ts_gt, tr1_gt, tr2_gt))
loss['overall'] = torch.mean(loss['bcd'])
return loss
