File size: 25,347 Bytes
63f3cf2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
# -*- coding: UTF-8 -*-
'''=================================================
@Project -> File   pram -> map3d
@IDE    PyCharm
@Author fx221@cam.ac.uk
@Date   04/03/2024 10:25
=================================================='''
import numpy as np
from collections import defaultdict
import os.path as osp
import pycolmap
import logging
import time

import torch

from localization.refframe import RefFrame
from localization.frame import Frame
from localization.point3d import Point3D
from colmap_utils.read_write_model import qvec2rotmat, read_model, read_compressed_model
from localization.utils import read_gt_pose


class SingleMap3D:
    def __init__(self, config, matcher, with_compress=False, start_sid: int = 0):
        self.config = config
        self.matcher = matcher
        self.image_path_prefix = self.config['image_path_prefix']
        self.start_sid = start_sid  # for a dataset with multiple scenes
        if not with_compress:
            cameras, images, p3ds = read_model(
                path=osp.join(config['landmark_path'], 'model'), ext='.bin')
            p3d_descs = np.load(osp.join(config['landmark_path'], 'point3D_desc.npy'),
                                allow_pickle=True)[()]
        else:
            cameras, images, p3ds = read_compressed_model(
                path=osp.join(config['landmark_path'], 'compress_model_{:s}'.format(config['cluster_method'])),
                ext='.bin')
            p3d_descs = np.load(osp.join(config['landmark_path'], 'compress_model_{:s}/point3D_desc.npy'.format(
                config['cluster_method'])), allow_pickle=True)[()]

        print('Load {} cameras {} images {} 3D points'.format(len(cameras), len(images), len(p3d_descs)))

        seg_data = np.load(
            osp.join(config['landmark_path'], 'point3D_cluster_n{:d}_{:s}_{:s}.npy'.format(config['n_cluster'],
                                                                                           config['cluster_mode'],
                                                                                           config['cluster_method'])),
            allow_pickle=True)[()]

        p3d_id = seg_data['id']
        seg_id = seg_data['label']
        p3d_seg = {p3d_id[i]: seg_id[i] for i in range(p3d_id.shape[0])}
        seg_p3d = {}
        for k in p3d_seg.keys():
            sid = p3d_seg[k]
            if sid in seg_p3d.keys():
                seg_p3d[sid].append(k)
            else:
                seg_p3d[sid] = [k]

        print('Load {} segments and {} 3d points'.format(len(seg_p3d.keys()), len(p3d_seg.keys())))
        seg_vrf = np.load(
            osp.join(config['landmark_path'], 'point3D_vrf_n{:d}_{:s}_{:s}.npy'.format(config['n_cluster'],
                                                                                       config['cluster_mode'],
                                                                                       config['cluster_method'])),
            allow_pickle=True)[()]

        # construct 3D map
        self.initialize_point3Ds(p3ds=p3ds, p3d_descs=p3d_descs, p3d_seg=p3d_seg)
        self.initialize_ref_frames(cameras=cameras, images=images)

        all_vrf_frame_ids = []
        self.seg_ref_frame_ids = {}
        for sid in seg_vrf.keys():
            self.seg_ref_frame_ids[sid] = []
            for vi in seg_vrf[sid].keys():
                vrf_frame_id = seg_vrf[sid][vi]['image_id']
                self.seg_ref_frame_ids[sid].append(vrf_frame_id)
                if with_compress and vrf_frame_id in self.reference_frames.keys():
                    self.reference_frames[vrf_frame_id].point3D_ids = seg_vrf[sid][vi]['original_points3d']

            all_vrf_frame_ids.extend(self.seg_ref_frame_ids[sid])

        if with_compress:
            all_ref_ids = list(self.reference_frames.keys())
            for fid in all_ref_ids:
                valid = self.reference_frames[fid].associate_keypoints_with_point3Ds(point3Ds=self.point3Ds)
                if not valid:
                    del self.reference_frames[fid]

        all_vrf_frame_ids = np.unique(all_vrf_frame_ids)
        all_vrf_frame_ids = [v for v in all_vrf_frame_ids if v in self.reference_frames.keys()]
        self.build_covisibility_graph(frame_ids=all_vrf_frame_ids, n_frame=config['localization'][
            'covisibility_frame'])  # build covisible frames for vrf frames only

        logging.info(
            f'Construct {len(self.reference_frames.keys())} ref frames and {len(self.point3Ds.keys())} 3d points')

        self.gt_poses = {}
        if config['gt_pose_path'] is not None:
            gt_pose_path = osp.join(config['dataset_path'], config['gt_pose_path'])
            self.read_gt_pose(path=gt_pose_path)

    def read_gt_pose(self, path, prefix=''):
        self.gt_poses = read_gt_pose(path=path)
        print('Load {} gt poses'.format(len(self.gt_poses.keys())))

    def initialize_point3Ds(self, p3ds, p3d_descs, p3d_seg):
        self.point3Ds = {}
        for id in p3ds.keys():
            if id not in p3d_seg.keys():
                continue
            self.point3Ds[id] = Point3D(id=id, xyz=p3ds[id].xyz, error=p3ds[id].error,
                                        refframe_id=-1, rgb=p3ds[id].rgb,
                                        descriptor=p3d_descs[id], seg_id=p3d_seg[id],
                                        frame_ids=p3ds[id].image_ids)

    def initialize_ref_frames(self, cameras, images):
        self.reference_frames = {}
        for id in images.keys():
            im = images[id]
            cam = cameras[im.camera_id]
            self.reference_frames[id] = RefFrame(camera=cam, id=id, qvec=im.qvec, tvec=im.tvec,
                                                 point3D_ids=im.point3D_ids,
                                                 keypoints=im.xys, name=im.name)

    def localize_with_ref_frame(self, q_frame: Frame, q_kpt_ids: np.ndarray, sid, semantic_matching=False):
        ref_frame_id = self.seg_ref_frame_ids[sid][0]
        ref_frame = self.reference_frames[ref_frame_id]
        if semantic_matching and sid > 0:
            ref_data = ref_frame.get_keypoints_by_sid(sid=sid)
        else:
            ref_data = ref_frame.get_keypoints()

        q_descs = q_frame.descriptors[q_kpt_ids]
        q_kpts = q_frame.keypoints[q_kpt_ids, :2]
        q_scores = q_frame.keypoints[q_kpt_ids, 2]

        xyzs = ref_data['xyzs']
        point3D_ids = ref_data['point3D_ids']
        ref_sids = np.array([self.point3Ds[v].seg_id for v in point3D_ids])
        with torch.no_grad():
            indices0 = self.matcher({
                'descriptors0': torch.from_numpy(q_descs)[None].cuda().float(),
                'keypoints0': torch.from_numpy(q_kpts)[None].cuda().float(),
                'scores0': torch.from_numpy(q_scores)[None].cuda().float(),
                'image_shape0': (1, 3, q_frame.camera.width, q_frame.camera.height),

                'descriptors1': torch.from_numpy(ref_data['descriptors'])[None].cuda().float(),
                'keypoints1': torch.from_numpy(ref_data['keypoints'])[None].cuda().float(),
                'scores1': torch.from_numpy(ref_data['scores'])[None].cuda().float(),
                'image_shape1': (1, 3, ref_frame.camera.width, ref_frame.camera.height),
            }
            )['matches0'][0].cpu().numpy()

        valid = indices0 >= 0
        mkpts = q_kpts[valid]
        mkpt_ids = q_kpt_ids[valid]
        mxyzs = xyzs[indices0[valid]]
        mpoint3D_ids = point3D_ids[indices0[valid]]
        matched_sids = ref_sids[indices0[valid]]
        matched_ref_keypoints = ref_data['keypoints'][indices0[valid]]

        # print('mkpts: ', mkpts.shape, mxyzs.shape, np.sum(indices0 >= 0))
        # cfg = q_frame.camera._asdict()
        # q_cam = pycolmap.Camera(model=q_frame.camera.model, )
        # config = {"estimation": {"ransac": {"max_error": ransac_thresh}}, **(config or {})}
        ret = pycolmap.absolute_pose_estimation(mkpts + 0.5,
                                                mxyzs,
                                                q_frame.camera,
                                                estimation_options={
                                                    "ransac": {"max_error": self.config['localization']['threshold']}},
                                                refinement_options={},
                                                # max_error_px=self.config['localization']['threshold']
                                                )
        if ret is None:
            ret = {'success': False, }
        else:
            ret['success'] = True
            ret['qvec'] = ret['cam_from_world'].rotation.quat[[3, 0, 1, 2]]
            ret['tvec'] = ret['cam_from_world'].translation
        ret['matched_keypoints'] = mkpts
        ret['matched_keypoint_ids'] = mkpt_ids
        ret['matched_xyzs'] = mxyzs
        ret['reference_frame_id'] = ref_frame_id
        ret['matched_point3D_ids'] = mpoint3D_ids
        ret['matched_sids'] = matched_sids
        ret['matched_ref_keypoints'] = matched_ref_keypoints

        if not ret['success']:
            ret['num_inliers'] = 0
            ret['inliers'] = np.zeros(shape=(mkpts.shape[0],), dtype=bool)
        return ret

    def match(self, query_data, ref_data):
        q_descs = query_data['descriptors']
        q_kpts = query_data['keypoints']
        q_scores = query_data['scores']
        xyzs = ref_data['xyzs']
        points3D_ids = ref_data['point3D_ids']
        with torch.no_grad():
            indices0 = self.matcher({
                'descriptors0': torch.from_numpy(q_descs)[None].cuda().float(),
                'keypoints0': torch.from_numpy(q_kpts)[None].cuda().float(),
                'scores0': torch.from_numpy(q_scores)[None].cuda().float(),
                'image_shape0': (1, 3, query_data['camera'].width, query_data['camera'].height),

                'descriptors1': torch.from_numpy(ref_data['descriptors'])[None].cuda().float(),
                'keypoints1': torch.from_numpy(ref_data['keypoints'])[None].cuda().float(),
                'scores1': torch.from_numpy(ref_data['scores'])[None].cuda().float(),
                'image_shape1': (1, 3, ref_data['camera'].width, ref_data['camera'].height),
            }
            )['matches0'][0].cpu().numpy()

        valid = indices0 >= 0
        mkpts = q_kpts[valid]
        mkpt_ids = np.where(valid)[0]
        mxyzs = xyzs[indices0[valid]]
        mpoints3D_ids = points3D_ids[indices0[valid]]

        return {
            'matched_keypoints': mkpts,
            'matched_xyzs': mxyzs,
            'matched_point3D_ids': mpoints3D_ids,
            'matched_keypoint_ids': mkpt_ids,
        }

    def build_covisibility_graph(self, frame_ids: list = None, n_frame: int = 20):
        def find_covisible_frames(frame_id):
            observed = self.reference_frames[frame_id].point3D_ids
            covis = defaultdict(int)
            for pid in observed:
                if pid == -1:
                    continue
                if pid not in self.point3Ds.keys():
                    continue
                for img_id in self.point3Ds[pid].frame_ids:
                    covis[img_id] += 1

            covis_ids = np.array(list(covis.keys()))
            covis_num = np.array([covis[i] for i in covis_ids])

            if len(covis_ids) <= n_frame:
                sel_covis_ids = covis_ids[np.argsort(-covis_num)]
            else:
                ind_top = np.argpartition(covis_num, -n_frame)
                ind_top = ind_top[-n_frame:]  # unsorted top k
                ind_top = ind_top[np.argsort(-covis_num[ind_top])]
                sel_covis_ids = [covis_ids[i] for i in ind_top]

            return sel_covis_ids

        if frame_ids is None:
            frame_ids = list(self.referece_frames.keys())

        self.covisible_graph = defaultdict()
        for frame_id in frame_ids:
            self.covisible_graph[frame_id] = find_covisible_frames(frame_id=frame_id)

    def refine_pose(self, q_frame: Frame, refinement_method='matching'):
        if refinement_method == 'matching':
            return self.refine_pose_by_matching(q_frame=q_frame)
        elif refinement_method == 'projection':
            return self.refine_pose_by_projection(q_frame=q_frame)
        else:
            raise NotImplementedError

    def refine_pose_by_matching(self, q_frame):
        ref_frame_id = q_frame.reference_frame_id
        db_ids = self.covisible_graph[ref_frame_id]
        print('Find {} covisible frames'.format(len(db_ids)))
        loc_success = q_frame.tracking_status
        if loc_success and ref_frame_id in db_ids:
            init_kpts = q_frame.matched_keypoints
            init_kpt_ids = q_frame.matched_keypoint_ids
            init_point3D_ids = q_frame.matched_point3D_ids
            init_xyzs = np.array([self.point3Ds[v].xyz for v in init_point3D_ids]).reshape(-1, 3)
            list(db_ids).remove(ref_frame_id)
        else:
            init_kpts = None
            init_xyzs = None
            init_point3D_ids = None

        matched_xyzs = []
        matched_kpts = []
        matched_point3D_ids = []
        matched_kpt_ids = []
        for idx, frame_id in enumerate(db_ids):
            ref_data = self.reference_frames[frame_id].get_keypoints()
            match_out = self.match(query_data={
                'keypoints': q_frame.keypoints[:, :2],
                'scores': q_frame.keypoints[:, 2],
                'descriptors': q_frame.descriptors,
                'camera': q_frame.camera, },
                ref_data=ref_data)
            if match_out['matched_keypoints'].shape[0] > 0:
                matched_kpts.append(match_out['matched_keypoints'])
                matched_xyzs.append(match_out['matched_xyzs'])
                matched_point3D_ids.append(match_out['matched_point3D_ids'])
                matched_kpt_ids.append(match_out['matched_keypoint_ids'])
        if len(matched_kpts) > 1:
            matched_kpts = np.vstack(matched_kpts)
            matched_xyzs = np.vstack(matched_xyzs).reshape(-1, 3)
            matched_point3D_ids = np.hstack(matched_point3D_ids)
            matched_kpt_ids = np.hstack(matched_kpt_ids)
        else:
            matched_kpts = matched_kpts[0]
            matched_xyzs = matched_xyzs[0]
            matched_point3D_ids = matched_point3D_ids[0]
            matched_kpt_ids = matched_kpt_ids[0]
        if init_kpts is not None and init_kpts.shape[0] > 0:
            matched_kpts = np.vstack([matched_kpts, init_kpts])
            matched_xyzs = np.vstack([matched_xyzs, init_xyzs])
            matched_point3D_ids = np.hstack([matched_point3D_ids, init_point3D_ids])
            matched_kpt_ids = np.hstack([matched_kpt_ids, init_kpt_ids])

        matched_sids = np.array([self.point3Ds[v].seg_id for v in matched_point3D_ids])

        print_text = 'Refinement by matching. Get {:d} covisible frames with {:d} matches for optimization'.format(
            len(db_ids), matched_xyzs.shape[0])
        print(print_text)

        t_start = time.time()
        ret = pycolmap.absolute_pose_estimation(matched_kpts + 0.5,
                                                matched_xyzs,
                                                q_frame.camera,
                                                estimation_options={
                                                    'ransac': {
                                                        'max_error': self.config['localization']['threshold'],
                                                        'min_num_trials': 1000,
                                                        'max_num_trials': 10000,
                                                        'confidence': 0.995,
                                                    }},
                                                refinement_options={},
                                                # max_error_px=self.config['localization']['threshold'],
                                                # min_num_trials=1000, max_num_trials=10000, confidence=0.995)
                                                )
        print('Time of RANSAC: {:.2f}s'.format(time.time() - t_start))

        if ret is None:
            ret = {'success': False, }
        else:
            ret['success'] = True
            ret['qvec'] = ret['cam_from_world'].rotation.quat[[3, 0, 1, 2]]
            ret['tvec'] = ret['cam_from_world'].translation

        ret['matched_keypoints'] = matched_kpts
        ret['matched_keypoint_ids'] = matched_kpt_ids
        ret['matched_xyzs'] = matched_xyzs
        ret['matched_point3D_ids'] = matched_point3D_ids
        ret['matched_sids'] = matched_sids

        if ret['success']:
            inlier_mask = np.array(ret['inliers'])
            best_reference_frame_ids = self.find_reference_frames(matched_point3D_ids=matched_point3D_ids[inlier_mask],
                                                                  candidate_frame_ids=self.covisible_graph.keys())
        else:
            best_reference_frame_ids = self.find_reference_frames(matched_point3D_ids=matched_point3D_ids,
                                                                  candidate_frame_ids=self.covisible_graph.keys())

        ret['refinement_reference_frame_ids'] = best_reference_frame_ids[:self.config['localization'][
            'covisibility_frame']]
        ret['reference_frame_id'] = best_reference_frame_ids[0]

        return ret

    @torch.no_grad()
    def refine_pose_by_projection(self, q_frame):
        q_Rcw = qvec2rotmat(q_frame.qvec)
        q_tcw = q_frame.tvec
        q_Tcw = np.eye(4, dtype=float)  # [4 4]
        q_Tcw[:3, :3] = q_Rcw
        q_Tcw[:3, 3] = q_tcw
        cam = q_frame.camera
        imw = cam.width
        imh = cam.height
        K = q_frame.get_intrinsics()  # [3, 3]
        reference_frame_id = q_frame.reference_frame_id
        covis_frame_ids = self.covisible_graph[reference_frame_id]
        if reference_frame_id not in covis_frame_ids:
            covis_frame_ids.append(reference_frame_id)
        all_point3D_ids = []

        for frame_id in covis_frame_ids:
            all_point3D_ids.extend(list(self.reference_frames[frame_id].point3D_ids))

        all_point3D_ids = np.unique(all_point3D_ids)
        all_xyzs = []
        all_descs = []
        all_sids = []
        for pid in all_point3D_ids:
            all_xyzs.append(self.point3Ds[pid].xyz)
            all_descs.append(self.point3Ds[pid].descriptor)
            all_sids.append(self.point3Ds[pid].seg_id)

        all_xyzs = np.array(all_xyzs)  # [N 3]
        all_descs = np.array(all_descs)  # [N 3]
        all_point3D_ids = np.array(all_point3D_ids)
        all_sids = np.array(all_sids)

        # move to gpu (distortion is not included)
        # proj_uv = pycolmap.camera.img_from_cam(
        #     np.array([1, 1, 1]).reshape(1, 3),
        # )
        all_xyzs_cuda = torch.from_numpy(all_xyzs).cuda()
        ones = torch.ones(size=(all_xyzs_cuda.shape[0], 1), dtype=all_xyzs_cuda.dtype).cuda()
        all_xyzs_cuda_homo = torch.cat([all_xyzs_cuda, ones], dim=1)  # [N 4]
        K_cuda = torch.from_numpy(K).cuda()
        proj_uvs = K_cuda @ (torch.from_numpy(q_Tcw).cuda() @ all_xyzs_cuda_homo.t())[:3, :]  # [3, N]
        proj_uvs[0] /= proj_uvs[2]
        proj_uvs[1] /= proj_uvs[2]
        mask = (proj_uvs[2] > 0) * (proj_uvs[2] < 100) * (proj_uvs[0] >= 0) * (proj_uvs[0] < imw) * (
                proj_uvs[1] >= 0) * (proj_uvs[1] < imh)

        proj_uvs = proj_uvs[:, mask]

        print('Projection: out of range {:d}/{:d}'.format(all_xyzs_cuda.shape[0], proj_uvs.shape[1]))

        mxyzs = all_xyzs[mask.cpu().numpy()]
        mpoint3D_ids = all_point3D_ids[mask.cpu().numpy()]
        msids = all_sids[mask.cpu().numpy()]

        q_kpts_cuda = torch.from_numpy(q_frame.keypoints[:, :2]).cuda()
        proj_error = q_kpts_cuda[..., None] - proj_uvs[:2][None]
        proj_error = torch.sqrt(torch.sum(proj_error ** 2, dim=1))  # [M N]
        out_of_range_mask = (proj_error >= 2 * self.config['localization']['threshold'])

        q_descs_cuda = torch.from_numpy(q_frame.descriptors).cuda().float()  # [M D]
        all_descs_cuda = torch.from_numpy(all_descs).cuda().float()[mask]  # [N D]
        desc_dist = torch.sqrt(2 - 2 * q_descs_cuda @ all_descs_cuda.t() + 1e-6)
        desc_dist[out_of_range_mask] = desc_dist[out_of_range_mask] + 100
        dists, ids = torch.topk(desc_dist, k=2, largest=False, dim=1)
        # apply nn ratio
        ratios = dists[:, 0] / dists[:, 1]  # smaller, better
        ratio_mask = (ratios <= 0.995) * (dists[:, 0] < 100)
        ratio_mask = ratio_mask.cpu().numpy()
        ids = ids.cpu().numpy()[ratio_mask, 0]

        ratio_num = torch.sum(ratios <= 0.995)
        proj_num = torch.sum(dists[:, 0] < 100)

        print('Projection: after ratio {:d}/{:d}, ratio {:d}, proj {:d}'.format(q_kpts_cuda.shape[0],
                                                                                np.sum(ratio_mask),
                                                                                ratio_num, proj_num))

        mkpts = q_frame.keypoints[ratio_mask]
        mkpt_ids = np.where(ratio_mask)[0]
        mxyzs = mxyzs[ids]
        mpoint3D_ids = mpoint3D_ids[ids]
        msids = msids[ids]
        print('projection: ', mkpts.shape, mkpt_ids.shape, mxyzs.shape, mpoint3D_ids.shape, msids.shape)

        t_start = time.time()
        ret = pycolmap.absolute_pose_estimation(mkpts[:, :2] + 0.5, mxyzs, q_frame.camera,
                                                estimation_options={
                                                    "ransac": {"max_error": self.config['localization']['threshold']}},
                                                refinement_options={},
                                                # max_error_px=self.config['localization']['threshold']
                                                )
        if ret is None:
            ret = {'success': False, }
        else:
            ret['success'] = True
            ret['qvec'] = ret['cam_from_world'].rotation.quat[[3, 0, 1, 2]]
            ret['tvec'] = ret['cam_from_world'].translation
        # inlier_mask = np.ones(shape=(mkpts.shape[0],), dtype=bool).tolist()
        # ret = pycolmap.pose_refinement(q_frame.tvec, q_frame.qvec, mkpts[:, :2] + 0.5, mxyzs, inlier_mask, cfg)
        # ret['num_inliers'] = np.sum(inlier_mask).astype(int)
        # ret['inliers'] = np.array(inlier_mask)

        print_text = 'Refinement by projection. Get {:d} inliers of {:d} matches for optimization'.format(
            ret['num_inliers'], mxyzs.shape[0])
        print(print_text)
        print('Time of RANSAC: {:.2f}s'.format(time.time() - t_start))

        ret['matched_keypoints'] = mkpts
        ret['matched_xyzs'] = mxyzs
        ret['matched_point3D_ids'] = mpoint3D_ids
        ret['matched_sids'] = msids
        ret['matched_keypoint_ids'] = mkpt_ids

        if ret['success']:
            inlier_mask = np.array(ret['inliers'])
            best_reference_frame_ids = self.find_reference_frames(matched_point3D_ids=mpoint3D_ids[inlier_mask],
                                                                  candidate_frame_ids=self.covisible_graph.keys())
        else:
            best_reference_frame_ids = self.find_reference_frames(matched_point3D_ids=mpoint3D_ids,
                                                                  candidate_frame_ids=self.covisible_graph.keys())

        ret['refinement_reference_frame_ids'] = best_reference_frame_ids[:self.config['localization'][
            'covisibility_frame']]
        ret['reference_frame_id'] = best_reference_frame_ids[0]

        if not ret['success']:
            ret['num_inliers'] = 0
            ret['inliers'] = np.zeros(shape=(mkpts.shape[0],), dtype=bool)

        return ret

    def find_reference_frames(self, matched_point3D_ids, candidate_frame_ids=None):
        covis_frames = defaultdict(int)
        for pid in matched_point3D_ids:
            for im_id in self.point3Ds[pid].frame_ids:
                if candidate_frame_ids is not None and im_id in candidate_frame_ids:
                    covis_frames[im_id] += 1

        covis_ids = np.array(list(covis_frames.keys()))
        covis_num = np.array([covis_frames[i] for i in covis_ids])
        sorted_idxes = np.argsort(covis_num)[::-1]  # larger to small
        sorted_frame_ids = covis_ids[sorted_idxes]
        return sorted_frame_ids

    def check_semantic_consistency(self, q_frame: Frame, sid, overlap_ratio=0.5):
        ref_frame_id = self.seg_ref_frame_ids[sid][0]
        ref_frame = self.reference_frames[ref_frame_id]

        q_sids = q_frame.seg_ids
        ref_sids = np.array([self.point3Ds[v].seg_id for v in ref_frame.point3D_ids]) + self.start_sid
        overlap_sids = np.intersect1d(q_sids, ref_sids)

        overlap_num1 = 0
        overlap_num2 = 0
        for sid in overlap_sids:
            overlap_num1 += np.sum(q_sids == sid)
            overlap_num2 += np.sum(ref_sids == sid)

        ratio1 = overlap_num1 / q_sids.shape[0]
        ratio2 = overlap_num2 / ref_sids.shape[0]

        # print('semantic_check: ', overlap_sids, overlap_num1, ratio1, overlap_num2, ratio2)

        return min(ratio1, ratio2) >= overlap_ratio