0] #zero-indexing creates and removes a dummy batch dimension
probs = torch.exp(log_probs).cpu()
out_model = torch.zeros((3, 3)).float() # estimated model
out_inliers = torch.zeros(log_probs.size()) # inlier mask of estimated model
out_gradients = torch.zeros(
log_probs.size()) # gradient tensor (only used during training)
rand_seed = 0 # random seed to by used in C++
# run NG-RANSAC
if opt.fmat:
# === CASE FUNDAMENTAL MATRIX =========================================
# undo normalization of x and y image coordinates
util.denormalize_pts(correspondences[0:2], img1.shape)
util.denormalize_pts(correspondences[2:4], img2.shape)
incount = ngransac.find_fundamental_mat(correspondences, probs, rand_seed,
opt.hyps, opt.threshold,
opt.refine, out_model, out_inliers,
out_gradients)
else:
# === CASE ESSENTIAL MATRIX =========================================
incount = ngransac.find_essential_mat(correspondences, probs, rand_seed,
opt.hyps, opt.threshold, out_model,
out_inliers, out_gradients)
print("\n=== Model found by NG-RANSAC: =======\n")
for b in range(correspondences.size(0)):
gradients = torch.zeros(
probs[b].size()
) # not used in test mode, indicates which correspondence have been sampled
inliers = torch.zeros(
probs[b].size()) # inlier mask of winning model
rand_seed = random.randint(
0, 10000) # provide a random seed for C++
if opt.fmat:
# === CASE FUNDAMENTAL MATRIX =========================================
# restore pixel coordinates
util.denormalize_pts(correspondences[b, 0:2], im_size1[b])
util.denormalize_pts(correspondences[b, 2:4], im_size2[b])
F = torch.zeros((3, 3))
#run NG-RANSAC
start_time = time.time()
incount = ngransac.find_fundamental_mat(
correspondences[b], probs[b], rand_seed, opt.hyps,
opt.threshold, opt.refine, F, inliers, gradients)
ransac_time += time.time() - start_time
# essential matrix from fundamental matrix (for evaluation via relative pose)
E = K2[b].transpose(0, 1).mm(F.mm(K1[b]))
pts1 = correspondences[b, 0:2].numpy()
