extract_descriptors_l2net.py

import torch
import numpy as np
import argparse
import h5py
from tqdm import tqdm
import os
import sys
import shutil
import json
import scipy.io as sio

import torchvision.transforms as transforms
from utils import cv2_greyscale, str2bool, save_h5

def get_transforms():

    transform = transforms.Compose([
        transforms.Lambda(cv2_greyscale),
        transforms.Lambda(cv2_scale),
        transforms.Lambda(np_reshape)

    ])

    return transform

def remove_option(parser, arg):
    for action in parser._actions:
        if (vars(action)['option_strings']
            and vars(action)['option_strings'][0] == arg) \
                or vars(action)['dest'] == arg:
            parser._remove_action(action)

    for action in parser._action_groups:
        vars_action = vars(action)
        var_group_actions = vars_action['_group_actions']
        for x in var_group_actions:
            if x.dest == arg:
                var_group_actions.remove(x)
                return


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--dataset_path",
        default=os.path.join('..', 'benchmark-patches-8k'),
        type=str,
        help='Path to the pre-generated patches')
    parser.add_argument(
        "--mrSize", default=12.0, type=float,
        help=' patch size in image is mrSize * pt.size. Default mrSize is 12' )
    parser.add_argument(
        "--save_path",
        default=os.path.join('..', 'benchmark-features'),
        type=str,
        help='Path to store the features')
    parser.add_argument(
        "--method_name", default='sift8k_8000_l2net', type=str)
    parser.add_argument(
        "--weights_path",
        default=os.path.join('third_party', 'l2net-config', 'l2net_ported_weights_lib+.pth'),
        type=str,
        help='Path to the model weights')
    parser.add_argument(
        "--matlab_weights_path",
        default=os.path.join('third_party', 'l2net', 'matlab', 'L2Net-LIB+.mat'),
        type=str,
        help='Path to the model weights')
    parser.add_argument(
        "--subset",
        default='both',
        type=str,
        help='Options: "val", "test", "both", "spc-fix"')
    parser.add_argument(
        "--clahe-mode",
        default='None',
        type=str,
        help='can be None, detector, descriptor, both')

    args = parser.parse_args()

    if args.subset not in ['val', 'test', 'both', 'spc-fix']:
        raise ValueError('Unknown value for --subset')
    seqs = []
    if args.subset == 'spc-fix':
        seqs += ['st_pauls_cathedral']
    else:
        if args.subset in ['val', 'both']:
            with open(os.path.join('data', 'val.json')) as f:
                seqs += json.load(f)
        if args.subset in ['test', 'both']:
            with open(os.path.join('data', 'test.json')) as f:
                seqs += json.load(f)
    print('Processing the following scenes: {}'.format(seqs))


    # Hacky work-around: reset argv for the HardNet argparse
    sys.path.append(os.path.join('misc', 'l2net'))
    sys.argv = [sys.argv[0]]
    from misc.l2net.l2net_model import L2Net
    from third_party.hardnet.code.Utils import cv2_scale, np_reshape
    import torch


    try:
        if torch.cuda.is_available():
            device = torch.device('cuda:0')
        else:
            device = torch.device('cpu')
    except:
        device = torch.device('cpu')

    suffix = ""
    if args.clahe_mode.lower() == 'detector':
        suffix = "_clahe_det"
    elif args.clahe_mode.lower() == 'descriptor':
        suffix = "_clahe_desc"
    elif args.clahe_mode.lower() == 'both':
        suffix = "_clahe_det_desc"
    elif args.clahe_mode.lower() == 'none':
        pass
    else:
        raise ValueError("unknown CLAHE mode. Try detector, descriptor or both")

    if abs(args.mrSize - 12.) > 0.1:
        suffix+= '_mrSize{:.1f}'.format(args.mrSize)
    args.method_name+=suffix
    print('Saving descriptors to folder: {}'.format(args.method_name))

    # get pre-trained image mean
    l2net_weights = sio.loadmat(args.matlab_weights_path)
    imgMean =  l2net_weights['pixMean']

    transforms = get_transforms()

    model = L2Net()
    model.load_state_dict(torch.load(args.weights_path,map_location=device))
    print('Loaded weights: {}'.format(args.weights_path))

    model = model.to(device)
    model.eval()

    for idx, seq_name in enumerate(seqs):
        print('Processing "{}"'.format(seq_name))

        seq_descriptors = {}
        patches_h5py_file = os.path.join(args.dataset_path, seq_name,
                                         'patches{}.h5'.format(suffix))

        with h5py.File(patches_h5py_file, 'r') as patches_h5py:
            for key, patches in tqdm(patches_h5py.items()):
                patches = patches.value

                bs = 128
                descriptors = np.zeros((len(patches), 128))

                for i in range(0, len(patches), bs):
                    data_a = patches[i:i + bs, :, :, :]
                    data_a = torch.stack([torch.from_numpy(transforms(patch)).squeeze() for patch in data_a]).\
                        unsqueeze(1).float().to(device)
                    # compute output
                    data_a = data_a - torch.from_numpy(imgMean).to(device)
                    with torch.no_grad():
                        out_a = model(data_a)
                        descriptors[i:i + bs] = out_a.cpu().detach().numpy()

                seq_descriptors[key] = descriptors.astype(np.float32)
        print('Processed {} images: {} descriptors/image'.format(
            len(seq_descriptors),
            np.array([s.shape[0] for s in seq_descriptors.values()]).mean()))

        cur_path = os.path.join(args.save_path, args.method_name, seq_name)
        if not os.path.exists(cur_path):
            os.makedirs(cur_path)
        save_h5(seq_descriptors, os.path.join(cur_path, 'descriptors.h5'))

        sub_files_in = ['keypoints{}.h5'.format(suffix), 'scales{}.h5'.format(suffix), 'angles{}.h5'.format(suffix), 'scores{}.h5'.format(suffix)]
        sub_files_out = ['keypoints.h5', 'scales.h5', 'angles.h5', 'scores.h5']

        for sub_file_in, sub_file_out in zip(sub_files_in, sub_files_out):
            shutil.copyfile(
                os.path.join(args.dataset_path, seq_name, sub_file_in),
                os.path.join(cur_path, sub_file_out))

        print('Done sequence: {}'.format(seq_name))