def evaluate_icp(args, test_loader):
dura = []
t_errors, R_errors, degree_errors = [], [], []
for i, (ref_cloud, src_cloud, gtR, gtt) in tqdm(enumerate(test_loader)):
if args.cuda:
ref_cloud, src_cloud, gtR, gtt = ref_cloud.cuda(), src_cloud.cuda(), \
gtR.cuda(), gtt.cuda()
ref_cloud = torch.squeeze(ref_cloud).cpu().numpy()
src_cloud = torch.squeeze(src_cloud).cpu().numpy()
tic = time.time()
R, t, pred_ref_cloud = icp(npy2pcd(src_cloud), npy2pcd(ref_cloud))
toc = time.time()
R = torch.from_numpy(np.expand_dims(R, 0)).to(gtR)
t = torch.from_numpy(np.expand_dims(t, 0)).to(gtt)
dura.append(toc - tic)
cur_t_error = translation_error(t, -gtt)
cur_R_error = rotation_error(R, gtR.permute(0, 2, 1).contiguous())
cur_degree_error = degree_error(R, gtR.permute(0, 2, 1).contiguous())
t_errors.append(cur_t_error.item())
R_errors.append(cur_R_error.item())
degree_errors.append(cur_degree_error.item())
if args.show:
print(cur_t_error.item(), cur_R_error.item(),
cur_degree_error.item())
pcd1 = npy2pcd(ref_cloud, 0)
pcd2 = npy2pcd(src_cloud, 1)
pcd3 = pred_ref_cloud
o3d.visualization.draw_geometries([pcd1, pcd2, pcd3])
return dura, np.mean(t_errors), np.mean(R_errors), np.mean(degree_errors)
def evaluate_benchmark(args, test_loader):
model = IterativeBenchmark(in_dim1=args.in_dim, niters=args.niters)
if args.cuda:
model = model.cuda()
model.load_state_dict(torch.load(args.checkpoint))
else:
model.load_state_dict(
torch.load(args.checkpoint, map_location=torch.device('cpu')))
model.eval()
dura = []
t_errors, R_errors, degree_errors = [], [], []
with torch.no_grad():
for i, (ref_cloud, src_cloud, gtR,
gtt) in tqdm(enumerate(test_loader)):
if args.cuda:
ref_cloud, src_cloud, gtR, gtt = ref_cloud.cuda(), src_cloud.cuda(), \
gtR.cuda(), gtt.cuda()
tic = time.time()
R, t, pred_ref_cloud = model(
src_cloud.permute(0, 2, 1).contiguous(),
ref_cloud.permute(0, 2, 1).contiguous())
toc = time.time()
dura.append(toc - tic)
cur_t_error = translation_error(t, -gtt)
cur_R_error = rotation_error(R, gtR.permute(0, 2, 1).contiguous())
cur_degree_error = degree_error(R,
gtR.permute(0, 2, 1).contiguous())
t_errors.append(cur_t_error.item())
R_errors.append(cur_R_error.item())
degree_errors.append(cur_degree_error.item())
if args.show:
print(cur_t_error.item(), cur_R_error.item(),
cur_degree_error.item())
ref_cloud = torch.squeeze(ref_cloud).cpu().numpy()
src_cloud = torch.squeeze(src_cloud).cpu().numpy()
pred_ref_cloud = torch.squeeze(pred_ref_cloud).cpu().numpy()
pcd1 = npy2pcd(ref_cloud, 0)
pcd2 = npy2pcd(src_cloud, 1)
pcd3 = npy2pcd(pred_ref_cloud, 2)
o3d.visualization.draw_geometries([pcd1, pcd2, pcd3])
return dura, np.mean(t_errors), np.mean(R_errors), np.mean(degree_errors)
def test_one_epoch(test_loader, model, loss_fn):
model.eval()
losses, t_errors, R_errors, degree_errors = [], [], [], []
with torch.no_grad():
for ref_cloud, src_cloud, gtR, gtt in tqdm(test_loader):
ref_cloud, src_cloud, gtR, gtt = ref_cloud.cuda(), src_cloud.cuda(), \
gtR.cuda(), gtt.cuda()
R, t, pred_ref_cloud = model(
src_cloud.permute(0, 2, 1).contiguous(),
ref_cloud.permute(0, 2, 1).contiguous())
loss = loss_fn(ref_cloud, pred_ref_cloud)
cur_t_error = translation_error(t, -gtt)
cur_R_error = rotation_error(R, gtR.permute(0, 2, 1).contiguous())
cur_degree_error = degree_error(R,
gtR.permute(0, 2, 1).contiguous())
losses.append(loss.item())
t_errors.append(cur_t_error.item())
R_errors.append(cur_R_error.item())
degree_errors.append(cur_degree_error.item())
model.train()
return np.mean(losses), np.mean(t_errors), np.mean(R_errors), np.mean(
degree_errors)
def train_one_epoch(train_loader, model, loss_fn, optimizer):
losses, t_errors, R_errors, degree_errors = [], [], [], []
for ref_cloud, src_cloud, gtR, gtt in tqdm(train_loader):
ref_cloud, src_cloud, gtR, gtt = ref_cloud.cuda(), src_cloud.cuda(), \
gtR.cuda(), gtt.cuda()
optimizer.zero_grad()
R, t, pred_ref_cloud = model(
src_cloud.permute(0, 2, 1).contiguous(),
ref_cloud.permute(0, 2, 1).contiguous())
loss = loss_fn(ref_cloud, pred_ref_cloud)
loss.backward()
optimizer.step()
cur_t_error = translation_error(t, -gtt)
cur_R_error = rotation_error(R, gtR.permute(0, 2, 1).contiguous())
cur_degree_error = degree_error(R, gtR.permute(0, 2, 1).contiguous())
losses.append(loss.item())
t_errors.append(cur_t_error.item())
R_errors.append(cur_R_error.item())
degree_errors.append(cur_degree_error.item())
return np.mean(losses), np.mean(t_errors), np.mean(R_errors), np.mean(
degree_errors)
