else:
refined_models = torch.zeros((M, 9))
refined_inliers = torch.zeros((M, Y))
for mi in range(M):
inliers = all_best_inliers[best_outer_hypo, mi]
inlier_indices = torch.nonzero(inliers)
if opt.em:
new_model = refined_models[mi]
new_model *= torch.sign(new_model[-1])
else:
new_model = best_models[mi]
refined_models[mi] = new_model
new_distances = sampling.homography_consistency_measure(new_model, data[0], device)
new_inliers = sampling.soft_inlier_fun(new_distances, 5. / (opt.threshold2), opt.threshold2)
refined_inliers[mi] = new_inliers
last_inlier_count = 0
selected_instances = 0
joint_inliers = torch.zeros((Y,))
for mi in range(M):
joint_inliers = torch.max(joint_inliers, refined_inliers[mi, :])
inlier_count = torch.sum(joint_inliers, dim=-1)
new_inliers = inlier_count - last_inlier_count
last_inlier_count = inlier_count
print("instance %d: %.2f additional inliers" % (mi, new_inliers))
if new_inliers < opt.inlier_cutoff:
break
selected_instances += 1
estm_models = []
data = data.to(device)
log_probs = model(data_and_state)
probs = torch.softmax(log_probs[bi, 0, 0:num_data[bi],
0],
dim=0)
probs /= torch.max(probs)
all_probs[oh, mi, :num_data[bi]] = probs
for hi in range(opt.hyps):
estm_line, choice_vec, _, inlier_count, distances = \
sampling.sample_model(
data[bi], num_data[bi], inlier_fun, 2, probs, sampling.vp_from_lines,
sampling.vp_consistency_measure_angle, device=device)
inliers = sampling.soft_inlier_fun(
distances, 5. / opt.threshold, opt.threshold)
all_choices[oh, mi, hi] = choice_vec
all_inliers[oh, mi, hi, 0:num_data[bi]] = inliers
all_inlier_counts[oh, mi, hi] = inlier_count
all_models[oh, mi, hi] = estm_line
best_hypo = torch.argmax(all_inlier_counts[oh, mi])
all_best_hypos[oh, mi] = best_hypo
if not opt.unconditional:
data_and_state[bi, 0:num_data[bi],
ddim - 1] = torch.max(
all_inliers[oh, mi, best_hypo,
0:num_data[bi]],
data_and_state[bi, 0:num_data[bi],
losses = []
inlier_counts = []
cur_probs = torch.softmax(
log_probs[bi, :, 0:num_data[bi]].squeeze(),
dim=-1)
models, inliers, choices, distances = \
sampling.sample_model_pool(data[bi], num_data[bi], opt.hyps, minimal_set_size,
inlier_fun, sampling.homography_from_points,
sampling.homography_consistency_measure, cur_probs,
device=device, model_size=model_dim)
all_grads[oh, :, mi, bi] = choices
inliers = sampling.soft_inlier_fun(
distances, 5. / opt.threshold, opt.threshold)
all_models[oh, :, mi, bi, :] = models
if mi > 0:
all_joint_inliers[oh, :, mi, bi] = torch.max(
inliers,
all_best_inliers[oh, mi - 1,
bi].unsqueeze(0).expand(
opt.hyps, -1))
else:
all_joint_inliers[oh, :, mi, bi] = inliers
inlier_counts = torch.sum(inliers, dim=-1)
cumulative_inlier_counts = torch.sum(
all_joint_inliers[oh, :, mi, bi], dim=-1)
if opt.problem == "homography":
for mi in range(M):
inliers = all_best_inliers[best_outer_hypo, mi]
inlier_indices = torch.nonzero(inliers)
if opt.em:
new_model = refined_models[mi]
new_model *= torch.sign(new_model[-1])
else:
new_model = best_models[mi]
refined_models[mi] = new_model
new_distances = sampling.homography_consistency_measure(
new_model, data[0], device)
new_inliers = sampling.soft_inlier_fun(new_distances,
5. / (opt.threshold2),
opt.threshold2)
refined_inliers[mi] = new_inliers
last_inlier_count = 0
selected_instances = 0
joint_inliers = torch.zeros((Y, ))
for mi in range(M):
joint_inliers = torch.max(joint_inliers, refined_inliers[mi, :])
inlier_count = torch.sum(joint_inliers, dim=-1)
new_inliers = inlier_count - last_inlier_count
last_inlier_count = inlier_count
print("instance %d: %.2f additional inliers" % (mi, new_inliers))
if new_inliers < opt.inlier_cutoff:
break
selected_instances += 1
