Make fragments

The first step of the scene reconstruction system is to create fragments from short RGBD sequences.

Input arguments

The script runs with python run_system.py [config] --make. In [config], ["path_dataset"] should have subfolders image and depth to store the color images and depth images respectively. We assume the color images and the depth images are synchronized and registered. In [config], the optional argument ["path_intrinsic"] specifies the path to a json file that stores the camera intrinsic matrix (See /tutorial/pipelines/rgbd_odometry.ipynb#read-camera-intrinsic for details). If it is not given, the PrimeSense factory setting is used instead.

Register RGBD image pairs

46# examples/python/reconstruction_system/make_fragments.py
47
48def register_one_rgbd_pair(s, t, color_files, depth_files, intrinsic,
49                           with_opencv, config):
50    source_rgbd_image = read_rgbd_image(color_files[s], depth_files[s], True,
51                                        config)
52    target_rgbd_image = read_rgbd_image(color_files[t], depth_files[t], True,
53                                        config)
54
55    option = o3d.pipelines.odometry.OdometryOption()
56    option.depth_diff_max = config["depth_diff_max"]
57    if abs(s - t) != 1:
58        if with_opencv:
59            success_5pt, odo_init = pose_estimation(source_rgbd_image,
60                                                    target_rgbd_image,
61                                                    intrinsic, False)
62            if success_5pt:
63                [success, trans, info
64                ] = o3d.pipelines.odometry.compute_rgbd_odometry(
65                    source_rgbd_image, target_rgbd_image, intrinsic, odo_init,
66                    o3d.pipelines.odometry.RGBDOdometryJacobianFromHybridTerm(),
67                    option)
68                return [success, trans, info]
69        return [False, np.identity(4), np.identity(6)]
70    else:
71        odo_init = np.identity(4)
72        [success, trans, info] = o3d.pipelines.odometry.compute_rgbd_odometry(
73            source_rgbd_image, target_rgbd_image, intrinsic, odo_init,
74            o3d.pipelines.odometry.RGBDOdometryJacobianFromHybridTerm(), option)
75        return [success, trans, info]
76

The function reads a pair of RGBD images and registers the source_rgbd_image to the target_rgbd_image. The Open3D function compute_rgbd_odometry is called to align the RGBD images. For adjacent RGBD images, an identity matrix is used as the initialization. For non-adjacent RGBD images, wide baseline matching is used as the initialization. In particular, the function pose_estimation computes OpenCV ORB feature to match sparse features over wide baseline images, then performs 5-point RANSAC to estimate a rough alignment, which is used as the initialization of compute_rgbd_odometry.

Multiway registration

 76# examples/python/reconstruction_system/make_fragments.py
 77
 78def make_posegraph_for_fragment(path_dataset, sid, eid, color_files,
 79                                depth_files, fragment_id, n_fragments,
 80                                intrinsic, with_opencv, config):
 81    o3d.utility.set_verbosity_level(o3d.utility.VerbosityLevel.Error)
 82    pose_graph = o3d.pipelines.registration.PoseGraph()
 83    trans_odometry = np.identity(4)
 84    pose_graph.nodes.append(
 85        o3d.pipelines.registration.PoseGraphNode(trans_odometry))
 86    for s in range(sid, eid):
 87        for t in range(s + 1, eid):
 88            # odometry
 89            if t == s + 1:
 90                print(
 91                    "Fragment %03d / %03d :: RGBD matching between frame : %d and %d"
 92                    % (fragment_id, n_fragments - 1, s, t))
 93                [success, trans,
 94                 info] = register_one_rgbd_pair(s, t, color_files, depth_files,
 95                                                intrinsic, with_opencv, config)
 96                trans_odometry = np.dot(trans, trans_odometry)
 97                trans_odometry_inv = np.linalg.inv(trans_odometry)
 98                pose_graph.nodes.append(
 99                    o3d.pipelines.registration.PoseGraphNode(
100                        trans_odometry_inv))
101                pose_graph.edges.append(
102                    o3d.pipelines.registration.PoseGraphEdge(s - sid,
103                                                             t - sid,
104                                                             trans,
105                                                             info,
106                                                             uncertain=False))
107
108            # keyframe loop closure
109            if s % config['n_keyframes_per_n_frame'] == 0 \
110                    and t % config['n_keyframes_per_n_frame'] == 0:
111                print(
112                    "Fragment %03d / %03d :: RGBD matching between frame : %d and %d"
113                    % (fragment_id, n_fragments - 1, s, t))
114                [success, trans,
115                 info] = register_one_rgbd_pair(s, t, color_files, depth_files,
116                                                intrinsic, with_opencv, config)
117                if success:
118                    pose_graph.edges.append(
119                        o3d.pipelines.registration.PoseGraphEdge(
120                            s - sid, t - sid, trans, info, uncertain=True))
121    o3d.io.write_pose_graph(
122        join(path_dataset, config["template_fragment_posegraph"] % fragment_id),
123        pose_graph)

This script uses the technique demonstrated in /tutorial/pipelines/multiway_registration.ipynb. The function make_posegraph_for_fragment builds a pose graph for multiway registration of all RGBD images in this sequence. Each graph node represents an RGBD image and its pose which transforms the geometry to the global fragment space. For efficiency, only key frames are used.

Once a pose graph is created, multiway registration is performed by calling the function optimize_posegraph_for_fragment.

51# examples/python/reconstruction_system/optimize_posegraph.py
52
53def optimize_posegraph_for_fragment(path_dataset, fragment_id, config):
54    pose_graph_name = join(path_dataset,
55                           config["template_fragment_posegraph"] % fragment_id)
56    pose_graph_optimized_name = join(
57        path_dataset,
58        config["template_fragment_posegraph_optimized"] % fragment_id)
59    run_posegraph_optimization(pose_graph_name, pose_graph_optimized_name,
60            max_correspondence_distance = config["depth_diff_max"],
61            preference_loop_closure = \
62            config["preference_loop_closure_odometry"])
63

This function calls global_optimization to estimate poses of the RGBD images.

Make a fragment

124# examples/python/reconstruction_system/make_fragments.py
125
126def integrate_rgb_frames_for_fragment(color_files, depth_files, fragment_id,
127                                      n_fragments, pose_graph_name, intrinsic,
128                                      config):
129    pose_graph = o3d.io.read_pose_graph(pose_graph_name)
130    volume = o3d.pipelines.integration.ScalableTSDFVolume(
131        voxel_length=config["tsdf_cubic_size"] / 512.0,
132        sdf_trunc=0.04,
133        color_type=o3d.pipelines.integration.TSDFVolumeColorType.RGB8)
134    for i in range(len(pose_graph.nodes)):
135        i_abs = fragment_id * config['n_frames_per_fragment'] + i
136        print(
137            "Fragment %03d / %03d :: integrate rgbd frame %d (%d of %d)." %
138            (fragment_id, n_fragments - 1, i_abs, i + 1, len(pose_graph.nodes)))
139        rgbd = read_rgbd_image(color_files[i_abs], depth_files[i_abs], False,
140                               config)
141        pose = pose_graph.nodes[i].pose
142        volume.integrate(rgbd, intrinsic, np.linalg.inv(pose))
143    mesh = volume.extract_triangle_mesh()
144    mesh.compute_vertex_normals()
145    return mesh
146

Once the poses are estimates, /tutorial/pipelines/rgbd_integration.ipynb is used to reconstruct a colored fragment from each RGBD sequence.

Batch processing

181# examples/python/reconstruction_system/make_fragments.py
182
183def run(config):
184
185    print("making fragments from RGBD sequence.")
186    make_clean_folder(join(config["path_dataset"], config["folder_fragment"]))
187
188    [color_files, depth_files] = get_rgbd_file_lists(config["path_dataset"])
189    n_files = len(color_files)
190    n_fragments = int(
191        math.ceil(float(n_files) / config['n_frames_per_fragment']))
192
193    if config["python_multi_threading"] is True:
194        from joblib import Parallel, delayed
195        import multiprocessing
196        import subprocess
197        MAX_THREAD = min(multiprocessing.cpu_count(), n_fragments)
198        Parallel(n_jobs=MAX_THREAD)(delayed(process_single_fragment)(
199            fragment_id, color_files, depth_files, n_files, n_fragments, config)
200                                    for fragment_id in range(n_fragments))
201    else:
202        for fragment_id in range(n_fragments):
203            process_single_fragment(fragment_id, color_files, depth_files,
204                                    n_files, n_fragments, config)

The main function calls each individual function explained above.

Results

Fragment 000 / 013 :: RGBD matching between frame : 0 and 1
Fragment 000 / 013 :: RGBD matching between frame : 0 and 5
Fragment 000 / 013 :: RGBD matching between frame : 0 and 10
Fragment 000 / 013 :: RGBD matching between frame : 0 and 15
Fragment 000 / 013 :: RGBD matching between frame : 0 and 20
:
Fragment 000 / 013 :: RGBD matching between frame : 95 and 96
Fragment 000 / 013 :: RGBD matching between frame : 96 and 97
Fragment 000 / 013 :: RGBD matching between frame : 97 and 98
Fragment 000 / 013 :: RGBD matching between frame : 98 and 99

The following is a log from optimize_posegraph_for_fragment.

[GlobalOptimizationLM] Optimizing PoseGraph having 100 nodes and 195 edges.
Line process weight : 389.309502
[Initial     ] residual : 3.223357e+05, lambda : 1.771814e+02
[Iteration 00] residual : 1.721845e+04, valid edges : 157, time : 0.022 sec.
[Iteration 01] residual : 1.350251e+04, valid edges : 168, time : 0.017 sec.
:
[Iteration 32] residual : 9.779118e+03, valid edges : 179, time : 0.013 sec.
Current_residual - new_residual < 1.000000e-06 * current_residual
[GlobalOptimizationLM] total time : 0.519 sec.
[GlobalOptimizationLM] Optimizing PoseGraph having 100 nodes and 179 edges.
Line process weight : 398.292104
[Initial     ] residual : 5.120047e+03, lambda : 2.565362e+02
[Iteration 00] residual : 5.064539e+03, valid edges : 179, time : 0.014 sec.
[Iteration 01] residual : 5.037665e+03, valid edges : 178, time : 0.015 sec.
:
[Iteration 11] residual : 5.017307e+03, valid edges : 177, time : 0.013 sec.
Current_residual - new_residual < 1.000000e-06 * current_residual
[GlobalOptimizationLM] total time : 0.197 sec.
CompensateReferencePoseGraphNode : reference : 0

The following is a log from integrate_rgb_frames_for_fragment.

Fragment 000 / 013 :: integrate rgbd frame 0 (1 of 100).
Fragment 000 / 013 :: integrate rgbd frame 1 (2 of 100).
Fragment 000 / 013 :: integrate rgbd frame 2 (3 of 100).
:
Fragment 000 / 013 :: integrate rgbd frame 97 (98 of 100).
Fragment 000 / 013 :: integrate rgbd frame 98 (99 of 100).
Fragment 000 / 013 :: integrate rgbd frame 99 (100 of 100).

The following images show some of the fragments made by this script.

../../_images/fragment_0.png ../../_images/fragment_1.png ../../_images/fragment_2.png ../../_images/fragment_3.png