HF-Net（三）基于NetVLAD及superpoint构建的sfm场景模型的查询图像位姿解算

参考：HF-Net git地址

0.数据准备

0.1NetVLAD预训练权重，存放在编译hfnet后生成的$DATA_PATH/weights/路径下

0.2aachen数据下载，存放在编译hfnet后生成的$DATA_PATH/路径下

0.3基于superpoint生成的aachen三维点云场景

0.4aachen数据对应的每张图像相机内参文件

1.基于NetVLAD生成aachen所有图像的global descriptors

参考HF-Net（一）基于NetVLAD的global descriptor的特征提取

2.基于superpoint的local descriptors生成

2.1执行代码

#对底库图像的superpoint特征提取
python3 hfnet/export_predictions.py hfnet/configs/superpoint_export_aachen_db.yaml superpoint/aachen --keys keypoints,scores,local_descriptor_map
#对查询图像的superpoint特征提取
python3 hfnet/export_predictions.py hfnet/configs/superpoint_export_aachen_queries.yaml superpoint/aachen --keys keypoints,scores,local_descriptor_map

2.2结果生成

3.基于NetVLAD及superpoint构建的sfm场景模型的查询图像位姿解算

3.1核心代码

import logging
from pathlib import Path
import argparse
from pprint import pformat
import yaml
import numpy as np
from pyquaternion import Quaternion

from hfnet.evaluation.localization import Localization, evaluate
from hfnet.evaluation.loaders import export_loader
from hfnet.settings import EXPER_PATH


configs_global = {
    'netvlad': {
        'db_name': 'globaldb_netvlad.pkl',
        'experiment': 'netvlad/aachen',
        'predictor': export_loader,
        'has_keypoints': False,
        'has_descriptors': False,
        'pca_dim': 1024,
        'num_prior': 10,
    },
    'hfnet': {
        'db_name': 'globaldb_hfnet.pkl',
        'experiment':  'hfnet/aachen',
        'predictor': export_loader,
        'has_keypoints': False,
        'has_descriptors': False,
        'pca_dim': 1024,
        'num_prior': 10,
    },
}

configs_local = {
    'superpoint': {
        'db_name': 'localdb_superpoint.pkl',
        'experiment': 'superpoint/aachen',
        'predictor': export_loader,
        'has_keypoints': True,
        'has_descriptors': True,
        'binarize': False,
        # 'do_nms': True,
        # 'nms_thresh': 4,
        'num_features': 2000,
        'ratio_thresh': 0.9,
    },
    'hfnet': {
        'db_name': 'localdb_hfnet.pkl',
        'experiment':  'hfnet/aachen',
        'predictor': export_loader,
        'has_keypoints': True,
        'has_descriptors': True,
        # 'do_nms': True,
        # 'nms_thresh': 4,
        'num_features': 2000,
        'ratio_thresh': 0.9,
    },
    'sift': {
        'db_name': 'localdb_sift.pkl',
        'colmap_db': 'aachen.db',
        'root': False,
        'ratio_thresh': 0.7,
        'fast_matching': False,
    },
    'doap': {
        'db_name': 'localdb_doap_kpts-sp-nms4.pkl',
        'experiment': 'doap/aachen_resize-960',
        'predictor': export_loader,
        'keypoint_predictor': export_loader,
        'keypoint_config': {
            'experiment': 'superpoint/aachen',
            # 'do_nms': True,
            # 'nms_thresh': 4,
        },
        'num_features': 2000,
        'ratio_thresh': 0.9,
    },
    'netvlad': {
        'db_name': 'localdb_netvlad_kpts-sp-nms4.pkl',
        'experiment': 'netvlad/aachen',
        'predictor': export_loader,
        'keypoint_predictor': export_loader,
        'keypoint_config': {
            'experiment': 'superpoint/aachen',
            # 'do_nms': True,
            # 'nms_thresh': 4,
        },
        'num_features': 1000,
        'ratio_thresh': 0.7,
    }
}

config_pose = {
    'reproj_error': 10,
    'min_inliers': 12,
}

config_aachen = {
    'resize_max': 960,
}


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('model', type=str)
    parser.add_argument('eval_name', type=str)
    parser.add_argument('--local_method', type=str)
    parser.add_argument('--global_method', type=str)
    parser.add_argument('--build_db', action='store_true')
    parser.add_argument('--queries', type=str, default='day_time')
    parser.add_argument('--max_iter', type=int)
    parser.add_argument('--export_poses', action='store_true')
    parser.add_argument('--cpp_backend', action='store_true')
    args = parser.parse_args()

    config = {
        'global': configs_global[args.global_method],
        'local': configs_local[args.local_method],
        'aachen': config_aachen,
        'pose': config_pose,
        'model': args.model,
        'max_iter': args.max_iter,
        'queries': args.queries,
        'use_cpp': args.cpp_backend,
    }
    logging.info(f'Configuration: \n'+pformat(config))

    logging.info('Evaluating Aachen with configuration: ')
    loc = Localization('aachen', args.model, config, build_db=args.build_db)

    query_file = f'{args.queries}_queries_with_intrinsics.txt'
    queries, query_dataset = loc.init_queries(query_file, config_aachen)

    logging.info('Starting evaluation')
    metrics, results = evaluate(
        loc, queries, query_dataset, max_iter=args.max_iter)
    logging.info('Evaluation metrics: \n'+pformat(metrics))

    output = {'config': config, 'metrics': metrics}
    output_dir = Path(EXPER_PATH, 'eval/aachen')
    output_dir.mkdir(exist_ok=True, parents=True)
    eval_path = Path(output_dir, f'{args.eval_name}.yaml')
    with open(eval_path, 'w') as f:
        yaml.dump(output, f, default_flow_style=False)

    if args.export_poses:
        poses_path = Path(output_dir, f'{args.eval_name}_poses.txt')
        with open(poses_path, 'w') as f:
            for query, result in zip(queries, results):
                query_T_w = np.linalg.inv(result.T)
                qvec_nvm = list(Quaternion(matrix=query_T_w))
                pos_nvm = query_T_w[:3, 3].tolist()
                line = f'{Path(query.name).name}'
                line += ' ' + ' '.join(map(str, qvec_nvm+pos_nvm))
                f.write(line+'\n')

3.2执行代码

python3 hfnet/evaluate_aachen.py /path/superpoint_sfm /DATA_PATH/aachen/images_upright/query/night/milestone --local_method superpoint --global_method netvlad --build_db --queries night_time --export_poses 

3.3执行结果
求解的位姿