crocodl-benchmark/lamar/run.py at main · cvg/crocodl-benchmark · GitHub

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
import argparse
from pathlib import Path
from typing import Optional, Dict
from pprint import pformat

from scantools.capture import Capture

from lamar.tasks import (
    FeatureExtraction, PairSelection, FeatureMatching, Mapping, PoseEstimation, ChunkAlignment)
from lamar.tasks.chunk_alignment import keys_from_chunks
from lamar.utils.capture import (
    read_query_list, build_chunks, avoid_duplicate_keys_in_chunks,
    rig_list_to_image_list, rig_poses_to_image_poses)

from lamar import logger


def run(outputs: Path,
        capture: Capture,
        ref_id: str,
        query_id: str,
        retrieval: str,
        feature: str,
        matcher: str,
        matcher_query: str = None,
        use_radios: bool = False,
        is_rig: bool = False,
        sequence_length_seconds: Optional[int] = None,
        num_pairs_loc: int = 10,
        num_pairs_map: int = 10,
        retrieval_mapping: Optional[str] = None,
        filter_pairs_mapping: Optional[Dict] = None,
        do_rig: bool = True,
        query_filename: str = 'keyframes_original.txt'):

    if matcher_query is None:
        matcher_query = matcher

    session_q = capture.sessions[query_id]
    is_sequential = sequence_length_seconds is not None
    if filter_pairs_mapping is None:
        filter_pairs_mapping = {
            'filter_frustum': {'do': True},
            'filter_pose': {'do': True, 'num_pairs_filter': 250},
        }

    configs = {
        'extraction': FeatureExtraction.methods[feature],
        'pairs_map': {
            'method': PairSelection.methods[retrieval_mapping or retrieval],
            'num_pairs': num_pairs_map,
            **filter_pairs_mapping,
        },
        'matching': FeatureMatching.methods[matcher],
        'mapping': Mapping.methods['triangulation'],
        'pairs_loc': {
            'method': PairSelection.methods[retrieval],
            'num_pairs': num_pairs_loc,
        },
        'poses': PoseEstimation.methods['rig' if is_rig and do_rig else 'single_image'],
        'matching_query': FeatureMatching.methods[matcher_query],
        # for multi-frame localization
        'extra_pairs_reloc': {
            'filter_frustum': {'do': True},
            'filter_pose': {'do': True, 'num_pairs_filter': 100}
        },
        'chunks': ChunkAlignment.methods['rig' if is_rig and do_rig else 'single_image'],
    }
    if use_radios:
        configs['pairs_loc']['filter_radio'] = {
            'do': True, 'window_us': 2_000_000, 'frac_pairs_filter': 0.025}
    if retrieval == 'overlap':  # add pose filtering to speed up the overlap
        configs['pairs_loc'].update({
            'filter_frustum': {'do': True},
            'filter_pose': {'do': True, 'num_pairs_filter': 250},
        })

    logger.info(f"Running query {query_id} with reference {ref_id} using {query_filename}.")

    query_list_path = capture.session_path(query_id) / 'proc' / query_filename
    query_list = image_keys = read_query_list(query_list_path)
    if is_rig and not do_rig:
        rig_query_list = query_list
        query_list = rig_list_to_image_list(rig_query_list, session_q)
    if is_sequential:
        query_list, query_chunks = build_chunks(
            capture, query_id, query_list, sequence_length_seconds)
        image_keys = keys_from_chunks(query_chunks)

    extraction_map = FeatureExtraction(outputs, capture, ref_id, configs['extraction'])
    pairs_map = PairSelection(outputs, capture, ref_id, ref_id, configs['pairs_map'])
    matching_map = FeatureMatching(
        outputs, capture, ref_id, ref_id, configs['matching'], pairs_map, extraction_map)

    mapping = Mapping(
        configs['mapping'], outputs, capture, ref_id, extraction_map, matching_map)

    extraction_query = FeatureExtraction(
        outputs, capture, query_id, configs['extraction'], image_keys)

    if is_sequential:
        query_list, query_chunks = avoid_duplicate_keys_in_chunks(
            session_q, query_list, query_chunks)
        T_c2w_gt = session_q.proc.alignment_trajectories
        chunk_alignment = ChunkAlignment(
            configs, outputs, capture, query_id, extraction_query, mapping, query_chunks,
            sequence_length_seconds)
        if T_c2w_gt:
            results = chunk_alignment.evaluate(T_c2w_gt, query_list)
        else:
            results = str(chunk_alignment.paths.poses)
    else:
        T_c2w_gt = session_q.proc.alignment_trajectories
        if T_c2w_gt and is_rig and not do_rig:
            T_c2w_gt = rig_poses_to_image_poses(rig_query_list, T_c2w_gt, session_q)
        pairs_loc = PairSelection(
            outputs, capture, query_id, ref_id, configs['pairs_loc'], query_list,
            query_poses=T_c2w_gt)
        matching_query = FeatureMatching(
            outputs, capture, query_id, ref_id, configs['matching_query'],
            pairs_loc, extraction_query, extraction_map)
        pose_estimation = PoseEstimation(
            configs['poses'], outputs, capture, query_id,
            extraction_query, matching_query, mapping, query_list)
        if T_c2w_gt:
            results = pose_estimation.evaluate(T_c2w_gt)
        else:
            results = str(pose_estimation.paths.poses)

    return results

if __name__ == '__main__':
    parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
    parser.add_argument('--ref_id', type=str, required=True)
    parser.add_argument('--query_id', type=str, required=True)
    parser.add_argument(
        '--captures', type=Path, default=Path("./data/"), help="Path to captures directory")
    parser.add_argument(
        '--outputs', type=Path, default=Path("./outputs/"), help="Path to outputs directory")
    parser.add_argument(
        '--retrieval', type=str, required=True, choices=list(PairSelection.methods))
    parser.add_argument(
        '--feature', type=str, required=True, choices=list(FeatureExtraction.methods))
    parser.add_argument(
        '--matcher', type=str, required=True, choices=list(FeatureMatching.methods))
    parser.add_argument(
        '--matcher_query', type=str, choices=list(FeatureMatching.methods))
    parser.add_argument('--use_radios', action='store_true')
    parser.add_argument('--sequence_length_seconds', type=int)
    parser.add_argument('--is_rig', action='store_true', help="If the session is a rigs-based one")
    parser.add_argument(
        '--query_filename', type=str,
        choices=['keyframes_original.txt', 'keyframes_pruned.txt', 'keyframes_pruned_subsampled.txt'],
        default='keyframes_original.txt')
    args = parser.parse_args().__dict__
    scene = args.pop("scene")
    args['capture'] = Capture.load(args.pop('captures'))
    results_ = run(**args)

    if isinstance(results_, str):
        logger.info('%s is a test sequence. Submit %s to the benchmark to obtain the results.',
                    args['query_id'], results_)
    else:
        logger.info('Results:\n%s', pformat(results_))