videos_to_colmap.py 17.4 KB
Newer Older
Clement Pinard's avatar
Clement Pinard committed
1
2
3
4
5
6
from colmap_util import read_model as rm, database as db
import anafi_metadata as am
from wrappers import FFMpeg, PDraw
from sklearn.cluster import KMeans
from sklearn.metrics import pairwise_distances_argmin_min
from argparse import ArgumentParser, ArgumentDefaultsHelpFormatter
nicolas's avatar
nicolas committed
7
from edit_exif import set_gps_location
Clement Pinard's avatar
Clement Pinard committed
8
9
10
11
12
from path import Path
import pandas as pd
import numpy as np
from pyproj import Proj
from tqdm import tqdm
13
import tempfile
Clement Pinard's avatar
Clement Pinard committed
14
15
16
17
18
19
20
21
22

parser = ArgumentParser(description='Take all the drone videos of a folder and put the frame '
                                    'location in a COLMAP file for vizualisation',
                        formatter_class=ArgumentDefaultsHelpFormatter)

parser.add_argument('--video_folder', metavar='DIR',
                    help='path to videos', type=Path)
parser.add_argument('--system', default='epsg:2154')
parser.add_argument('--centroid_path', default=None)
Clément Pinard's avatar
Clément Pinard committed
23
parser.add_argument('--colmap_img_root', metavar='DIR', type=Path)
Clement Pinard's avatar
Clement Pinard committed
24
25
26
27
28
29
30
parser.add_argument('--output_format', metavar='EXT', default="bin")
parser.add_argument('--vid_ext', nargs='+', default=[".mp4", ".MP4"])
parser.add_argument('--pic_ext', nargs='+', default=[".jpg", ".JPG", ".png", ".PNG"])
parser.add_argument('--nw', default='',
                    help="native-wrapper.sh file location")
parser.add_argument('--fps', default=1, type=int,
                    help="framerate at which videos will be scanned WITH reconstruction")
31
parser.add_argument('--total_frames', default=200, type=int)
Clement Pinard's avatar
Clement Pinard committed
32
33
34
parser.add_argument('--orientation_weight', default=1, type=float)
parser.add_argument('--resolution_weight', default=1, type=float)
parser.add_argument('--save_space', action="store_true")
Clément Pinard's avatar
Clément Pinard committed
35
36
parser.add_argument('--thorough_db', type=Path)
parser.add_argument('-v', '--verbose', action="count", default=0)
Clement Pinard's avatar
Clement Pinard committed
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55


def world_coord_from_frame(frame_qvec, frame_tvec):
    '''
    frame_qvec is written in the NED system (north east down)
    frame_tvec is already is the world system (east norht up)
    '''
    world2NED = np.float32([[0, 1, 0],
                            [1, 0, 0],
                            [0, 0, -1]])
    NED2cam = np.float32([[0, 1, 0],
                          [0, 0, 1],
                          [1, 0, 0]])
    world2cam = NED2cam @ rm.qvec2rotmat(frame_qvec).T @ world2NED
    cam_tvec = - world2cam  @ frame_tvec
    cam_qvec = rm.rotmat2qvec(world2cam)
    return cam_qvec, cam_tvec


nicolas's avatar
nicolas committed
56
57
58
59
60
61
def set_gps(frames_list, metadata, image_path):
    for frame in frames_list:
        relative = str(frame.relpath(image_path))
        row = metadata[metadata["image_path"] == relative]
        if len(row) > 0:
            row = row.iloc[0]
62
63
64
65
66
            if row["location_valid"]:
                set_gps_location(frame,
                                 lat=row["location_latitude"],
                                 lng=row["location_longitude"],
                                 altitude=row["location_altitude"])
nicolas's avatar
nicolas committed
67
68
69
70
71
72
73
74
75
76
77
78
79


def get_georef(metadata):
    relevant_data = metadata[["location_valid", "image_path", "x", "y", "z"]]
    path_list = []
    georef_list = []
    for _, (gps, path, x, y, alt) in relevant_data.iterrows():
        path_list.append(path)
        if gps == 1:
            georef_list.append("{} {} {} {}\n".format(path, x, y, alt))
    return georef_list, path_list


Clement Pinard's avatar
Clement Pinard committed
80
def optimal_sample(metadata, num_frames, orientation_weight, resolution_weight):
81
82
83
84
85
86
87
    valid_metadata = metadata[~metadata["sampled"]].dropna()
    XYZ = valid_metadata[["x", "y", "z"]].values
    axis_angle = valid_metadata[["frame_quat_x", "frame_quat_y", "frame_quat_z"]].values
    if True in valid_metadata["indoor"].unique():
        # We have indoor videos, without absolute positions. We assume each video is very far
        # from the other ones. As such we will have an optimal subsampling of each video
        # It won't leverage video proximity from each other but it's better than nothing
Clement Pinard's avatar
Clement Pinard committed
88
        diameter = (XYZ.max(axis=0) - XYZ.min(axis=0))
89
90
91
92
        indoor_videos = valid_metadata.loc[valid_metadata["indoor"]]["video"].unique()
        new_centroids = 2 * diameter * np.linspace(0, 10, len(indoor_videos)).reshape(-1, 1)
        for centroid, v in zip(new_centroids, indoor_videos):
            video_index = (valid_metadata["video"] == v).values
Clement Pinard's avatar
Clement Pinard committed
93
94
            XYZ[video_index] += centroid

95
    frame_size = valid_metadata["video_quality"].values
Clement Pinard's avatar
Clement Pinard committed
96
97
98
99
100
101
102
103
    weighted_point_cloud = np.concatenate([XYZ, orientation_weight * axis_angle], axis=1)

    if resolution_weight == 0:
        weights = None
    else:
        weights = frame_size ** resolution_weight
    km = KMeans(n_clusters=num_frames).fit(weighted_point_cloud, sample_weight=weights)
    closest, _ = pairwise_distances_argmin_min(km.cluster_centers_, weighted_point_cloud)
104
    metadata.at[valid_metadata.index[closest], "sampled"] = True
Clement Pinard's avatar
Clement Pinard committed
105
106
107
    return metadata


108
def register_new_cameras(cameras_dataframe, database, camera_dict):
Clement Pinard's avatar
Clement Pinard committed
109
    camera_ids = []
Clément Pinard's avatar
Clément Pinard committed
110
    for _, (w, h, f, hfov, vfov, camera_model, *_) in cameras_dataframe.iterrows():
111
112
113
114
115
116
117
118
119
        if not (np.isnan(hfov) or np.isnan(vfov)):
            fx = w / (2 * np.tan(hfov * np.pi/360))
            fy = h / (2 * np.tan(vfov * np.pi/360))
        else:
            fx = w  # This is just a placeholder meant to be optimized
            fy = w
        model_id = rm.CAMERA_MODEL_NAMES[camera_model].model_id
        num_params = rm.CAMERA_MODEL_NAMES[camera_model].num_params
        params = np.array([fx, fy, w/2, h/2] + [0] * (num_params - 4))
Clement Pinard's avatar
Clement Pinard committed
120
121
122
        db_id = database.add_camera(model_id, w, h, params, prior_focal_length=True)
        camera_ids.append(db_id)
        camera_dict[db_id] = rm.Camera(id=db_id,
123
                                       model=camera_model,
Clement Pinard's avatar
Clement Pinard committed
124
125
126
127
128
129
130
131
                                       width=int(w),
                                       height=int(h),
                                       params=params)
    ids_series = pd.Series(camera_ids)
    return cameras_dataframe.set_index(ids_series)


def process_video_folder(videos_list, existing_pictures, output_video_folder, image_path, system, centroid,
Clément Pinard's avatar
Clément Pinard committed
132
                         thorough_db, fps=1, total_frames=500, orientation_weight=1, resolution_weight=1,
133
134
                         output_colmap_format="bin", save_space=False, include_lowfps_thorough=False,
                         max_sequence_length=1000, **env):
Clement Pinard's avatar
Clement Pinard committed
135
136
137
138
139
    proj = Proj(system)
    final_metadata = []
    video_output_folders = {}
    images = {}
    colmap_cameras = {}
140
    tempfile_database = Path(tempfile.NamedTemporaryFile().name)
Clément Pinard's avatar
Clément Pinard committed
141
142
    if thorough_db.isfile():
        thorough_db.copy(thorough_db.stripext() + "_backup.db")
Clement Pinard's avatar
Clement Pinard committed
143
    path_lists_output = {}
144
    database = db.COLMAPDatabase.connect(thorough_db)
Clement Pinard's avatar
Clement Pinard committed
145
146
147
    database.create_tables()

    print("extracting metadata for {} videos...".format(len(videos_list)))
148
    videos_summary = {"anafi": {"indoor": 0, "outdoor": 0}, "generic": 0}
Clement Pinard's avatar
Clement Pinard committed
149
    for v in tqdm(videos_list):
150
        width, height, framerate, num_frames = env["ffmpeg"].get_size_and_framerate(v)
Clement Pinard's avatar
Clement Pinard committed
151
152
153
154
        video_output_folder = output_video_folder / "{}x{}".format(width, height) / v.namebase
        video_output_folder.makedirs_p()
        video_output_folders[v] = video_output_folder

155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
        try:
            metadata = am.extract_metadata(v.parent, v, env["pdraw"], proj,
                                           width, height, framerate)
            metadata["model"] = "anafi"
            metadata["camera_model"] = "PINHOLE"
            if metadata["indoor"].iloc[0]:
                videos_summary["anafi"]["indoor"] += 1
            else:
                videos_summary["anafi"]["outdoor"] += 1
                raw_positions = metadata[["x", "y", "z"]]
                if centroid is None:
                    '''No centroid (possibly because there was no georeferenced lidar model in the first place)
                    set it as the first valid GPS position of the first outdoor video'''
                    centroid = raw_positions[metadata["location_valid"] == 1].iloc[0].values
                zero_centered_positions = raw_positions.values - centroid
                metadata["x"], metadata["y"], metadata["z"] = zero_centered_positions.transpose()
        except Exception:
            # No metadata found, construct a simpler dataframe without location
            metadata = pd.DataFrame({"video": [v] * num_frames})
            metadata["height"] = height
            metadata["width"] = width
            metadata["framerate"] = framerate
            metadata["video_quality"] = height * width / framerate
            metadata['frame'] = metadata.index + 1
            # timestemp is in microseconds
            metadata['time'] = 1e6 * metadata.index / framerate
            metadata['indoor'] = True
            metadata['location_valid'] = 0
            metadata["model"] = "generic"
            metadata["camera_model"] = "OPENCV"
            metadata["picture_hfov"] = height
            metadata["picture_vfov"] = height
            videos_summary["generic"] += 1
        if include_lowfps_thorough:
            by_time = metadata.set_index(pd.to_datetime(metadata["time"], unit="us"))
            by_time_lowfps = by_time.resample("{:.3f}S".format(1/fps)).first()
            metadata["sampled"] = by_time["time"].isin(by_time_lowfps["time"]).values
Clément Pinard's avatar
Clément Pinard committed
192
        else:
193
194
            metadata["sampled"] = False
        final_metadata.append(metadata)
Clement Pinard's avatar
Clement Pinard committed
195
    final_metadata = pd.concat(final_metadata, ignore_index=True)
196
197
198
199
    print(final_metadata["sampled"])
    print("{} outdoor anafi videos".format(videos_summary["anafi"]["outdoor"]))
    print("{} indoor anafi videos".format(videos_summary["anafi"]["indoor"]))
    print("{} generic videos".format(videos_summary["generic"]))
Clement Pinard's avatar
Clement Pinard committed
200
201
202

    print("{} frames in total".format(len(final_metadata)))

203
204
205
    cam_fields = ["width", "height", "framerate", "picture_hfov", "picture_vfov", "camera_model"]
    cameras_dataframe = final_metadata[final_metadata["model"] == "anafi"][cam_fields].drop_duplicates()
    cameras_dataframe = register_new_cameras(cameras_dataframe, database, colmap_cameras)
Clement Pinard's avatar
Clement Pinard committed
206
207
208
    final_metadata["camera_id"] = 0
    for cam_id, row in cameras_dataframe.iterrows():
        final_metadata.loc[(final_metadata[cam_fields] == row).all(axis=1), "camera_id"] = cam_id
209
210
211
    if any(final_metadata["model"] == "generic"):
        print("Undefined remaining cameras, assigning generic models to them")
        generic_frames = final_metadata[final_metadata["model"] == "generic"]
Clément Pinard's avatar
Clément Pinard committed
212
213
        generic_cam_fields = cam_fields + ["video"]
        generic_cameras_dataframe = generic_frames[generic_cam_fields]
214
215
216
217
218
219
        fixed_camera = True
        if fixed_camera:
            generic_cameras_dataframe = generic_cameras_dataframe.drop_duplicates()
        generic_cameras_dataframe = register_new_cameras(generic_cameras_dataframe, database, colmap_cameras)
        if fixed_camera:
            for cam_id, row in generic_cameras_dataframe.iterrows():
Clément Pinard's avatar
Clément Pinard committed
220
                final_metadata.loc[(final_metadata[generic_cam_fields] == row).all(axis=1), "camera_id"] = cam_id
221
222
223
224
225
226
227
228
229
230
        else:
            final_metadata.loc[generic_frames.index, "camera_id"] = generic_cameras_dataframe.index
        cameras_dataframe = cameras_dataframe.append(generic_cameras_dataframe)
    print("Cameras : ")
    print(cameras_dataframe)

    to_extract = total_frames - len(existing_pictures) - sum(final_metadata["sampled"])
    print(to_extract, total_frames, len(existing_pictures), sum(final_metadata["sampled"]))
    print(final_metadata["sampled"])
    print(final_metadata["sampled"].unique())
Clement Pinard's avatar
Clement Pinard committed
231

Clément Pinard's avatar
Clément Pinard committed
232
    if to_extract <= 0:
233
        pass
Clément Pinard's avatar
Clément Pinard committed
234
    elif to_extract < len(final_metadata):
Clement Pinard's avatar
Clement Pinard committed
235
236
237
238
239
240
        print("subsampling based on K-Means, to get {}"
              " frames from videos, for a total of {} frames".format(to_extract, total_frames))
        final_metadata = optimal_sample(final_metadata, total_frames - len(existing_pictures),
                                        orientation_weight,
                                        resolution_weight)
        print("Done.")
Clément Pinard's avatar
Clément Pinard committed
241
242
    else:
        final_metadata["sampled"] = True
Clement Pinard's avatar
Clement Pinard committed
243

244
    print("Constructing COLMAP model with {:,} frames".format(sum(final_metadata["sampled"])))
Clement Pinard's avatar
Clement Pinard committed
245

246
247
248
249
    database.commit()
    thorough_db.copy(tempfile_database)
    temp_database = db.COLMAPDatabase.connect(tempfile_database)

Clement Pinard's avatar
Clement Pinard committed
250
    final_metadata["image_path"] = ""
251
252
    final_metadata["db_id"] = -1
    for current_id, row in tqdm(final_metadata.iterrows(), total=len(final_metadata)):
Clement Pinard's avatar
Clement Pinard committed
253
254
255
256
257
        video = row["video"]
        frame = row["frame"]
        camera_id = row["camera_id"]
        current_image_path = video_output_folders[video].relpath(image_path) / video.namebase + "_{:05d}.jpg".format(frame)

258
259
260
        final_metadata.at[current_id, "image_path"] = current_image_path
        db_image_id = temp_database.add_image(current_image_path, int(camera_id))
        final_metadata.at[current_id, "db_id"] = db_image_id
Clement Pinard's avatar
Clement Pinard committed
261
262
263
264
265
266

        if row["sampled"]:
            frame_qvec = row[["frame_quat_w",
                              "frame_quat_x",
                              "frame_quat_y",
                              "frame_quat_z"]].values
267
268
            if True in pd.isnull(frame_qvec):
                frame_qvec = np.array([1, 0, 0, 0])
nicolas's avatar
nicolas committed
269
270
            x, y, z = row[["x", "y", "z"]]
            frame_tvec = np.array([x, y, z])
Clement Pinard's avatar
Clement Pinard committed
271
272
273
274
275
276
            if row["location_valid"]:
                frame_gps = row[["location_longitude", "location_latitude", "location_altitude"]]
            else:
                frame_gps = np.full(3, np.NaN)

            world_qvec, world_tvec = world_coord_from_frame(frame_qvec, frame_tvec)
277
278
279
280
            database.add_image(current_image_path, int(camera_id), prior_t=frame_gps, image_id=db_image_id)
            images[db_image_id] = rm.Image(id=db_image_id, qvec=world_qvec, tvec=world_tvec,
                                           camera_id=camera_id, name=current_image_path,
                                           xys=[], point3D_ids=[])
Clement Pinard's avatar
Clement Pinard committed
281
282
283

    database.commit()
    database.close()
284
285
    temp_database.commit()
    temp_database.close()
Clement Pinard's avatar
Clement Pinard committed
286
287
288
    rm.write_model(colmap_cameras, images, {}, output_video_folder, "." + output_colmap_format)
    print("COLMAP model created")

nicolas's avatar
nicolas committed
289
290
291
292
    thorough_georef, thorough_paths = get_georef(final_metadata[final_metadata["sampled"]])
    path_lists_output["thorough"] = {}
    path_lists_output["thorough"]["frames"] = thorough_paths
    path_lists_output["thorough"]["georef"] = thorough_georef
Clement Pinard's avatar
Clement Pinard committed
293
294
295
296
297

    print("Extracting frames from videos")

    for v in tqdm(videos_list):
        video_metadata = final_metadata[final_metadata["video"] == v]
nicolas's avatar
nicolas committed
298
        by_time = video_metadata.set_index(pd.to_datetime(video_metadata["time"], unit="us"))
Clement Pinard's avatar
Clement Pinard committed
299
300
        video_folder = video_output_folders[v]
        video_metadata.to_csv(video_folder/"metadata.csv")
nicolas's avatar
nicolas committed
301
302
303
304
305
        path_lists_output[v] = {}
        video_metadata_1fps = by_time.resample("{:.3f}S".format(1/fps)).first()
        georef, frame_paths = get_georef(video_metadata_1fps)
        path_lists_output[v]["frames_lowfps"] = frame_paths
        path_lists_output[v]["georef_lowfps"] = georef
Clément Pinard's avatar
Clément Pinard committed
306
        num_chunks = len(video_metadata) // max_sequence_length + 1
307
308
309
310
311
312
313
        chunks = [list(frames) for frames in np.array_split(video_metadata["image_path"],
                                                            num_chunks)]
        # Add some overlap between chunks, in order to ease the model merging afterwards
        for chunk, next_chunk in zip(chunks, chunks[1:]):
            chunk.extend(next_chunk[:10])
        path_lists_output[v]["frames_full"] = chunks

Clement Pinard's avatar
Clement Pinard committed
314
        if save_space:
Clément Pinard's avatar
Clément Pinard committed
315
316
            frame_ids = set(video_metadata[video_metadata["sampled"]]["frame"].values) | \
                set(video_metadata_1fps["frame"].values)
317
            frame_ids = sorted(list(frame_ids))
Clément Pinard's avatar
Clément Pinard committed
318
            if len(frame_ids) > 0:
nicolas's avatar
nicolas committed
319
                extracted_frames = env["ffmpeg"].extract_specific_frames(v, video_folder, frame_ids)
Clement Pinard's avatar
Clement Pinard committed
320
        else:
nicolas's avatar
nicolas committed
321
322
            extracted_frames = env["ffmpeg"].extract_images(v, video_folder)
        set_gps(extracted_frames, video_metadata, image_path)
Clement Pinard's avatar
Clement Pinard committed
323
324
325
326
327
328
329
330

    return path_lists_output, video_output_folders


if __name__ == '__main__':
    args = parser.parse_args()
    env = vars(args)
    env["videos_list"] = sum((list(args.video_folder.walkfiles('*{}'.format(ext))) for ext in args.vid_ext), [])
Clément Pinard's avatar
Clément Pinard committed
331
    output_video_folder = args.colmap_img_root / "Videos"
Clement Pinard's avatar
Clement Pinard committed
332
    output_video_folder.makedirs_p()
Clément Pinard's avatar
Clément Pinard committed
333
    env["image_path"] = args.colmap_img_root
Clement Pinard's avatar
Clement Pinard committed
334
    env["output_video_folder"] = output_video_folder
Clément Pinard's avatar
Clément Pinard committed
335
    env["existing_pictures"] = sum((list(args.colmap_img_root.walkfiles('*{}'.format(ext))) for ext in args.pic_ext), [])
Clément Pinard's avatar
Clément Pinard committed
336
337
338
    env["pdraw"] = PDraw(args.nw, verbose=args.verbose)
    env["ffmpeg"] = FFMpeg(verbose=args.verbose)
    env["output_colmap_format"] = args.output_format
Clement Pinard's avatar
Clement Pinard committed
339
340
341
342
343
344
345
346
347

    if args.centroid_path is not None:
        centroid = np.loadtxt(args.centroid_path)
    else:
        centroid = np.zeros(3)
    env["centroid"] = centroid
    lists, extracted_video_folders = process_video_folder(**env)

    if lists is not None:
Clément Pinard's avatar
Clément Pinard committed
348
349
350
        with open(args.colmap_img_root/"video_frames_for_thorough_scan.txt", "w") as f:
            f.write("\n".join(lists["thorough"]["frames"]) + "\n")
        with open(args.colmap_img_root/"georef.txt", "w") as f:
Clément Pinard's avatar
Clément Pinard committed
351
            f.write("\n".join(lists["thorough"]["georef"]))
Clement Pinard's avatar
Clement Pinard committed
352
        for v in env["videos_list"]:
Clément Pinard's avatar
Clément Pinard committed
353
354
355
356
357
358
359
360
361
            video_folder = extracted_video_folders[v]
            with open(video_folder / "lowfps.txt", "w") as f:
                f.write("\n".join(lists[v]["frames_lowfps"]) + "\n")
            with open(video_folder / "georef.txt", "w") as f:
                f.write("\n".join(lists["thorough"]["georef"]) + "\n")
                f.write("\n".join(lists[v]["georef_lowfps"]) + "\n")
            for j, l in enumerate(lists[v]["frames_full"]):
                with open(video_folder / "full_chunk_{}.txt".format(j), "w") as f:
                    f.write("\n".join(l) + "\n")