Some questions about you provided retrieval results

[xxlxsyhl]

I was unable to reproduce the retrieval results https://github.com/cvg/Hierarchical-Localization/blob/master/pairs/aachen/pairs-query-netvlad50.txt using the model provided at https://github.com/uzh-rpg/netvlad_tf_open. Are you using your own trained NetVLAD model?

[sarlinpe]

I used this exact same repository with the model trained on pitts30k. I used the following script, with some additional changes to the netvlad_tf lib to support TensorFlow v1.

import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()

import cv2
from pathlib import Path
import h5py
from tqdm import tqdm
import numpy as np

import netvlad_tf.nets as nets

def inference(root, output_file, resize_max=1024):
    paths = []
    globs = ['*.jpg', '*.png', '*.jpeg', '*.JPG', '*.PNG']
    for glob in globs:
        paths += list(Path(root).glob('**/'+glob))
    print(f'Found {len(paths)} images')
    assert len(paths) > 0

    tf.reset_default_graph()

    image_batch = tf.placeholder(
        dtype=tf.float32, shape=[None, None, None, 3])
    net_out = nets.vgg16NetvladPca(image_batch)
    saver = tf.train.Saver()
    gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.5)
    sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
    ckpt = nets.defaultCheckpoint('vd16_pitts30k_conv5_3_vlad_preL2_intra_white')
    saver.restore(sess, ckpt)

    hfile = h5py.File(str(output_file), 'a')

    for path in tqdm(paths):
        img = cv2.imread(str(path))

        h, w = img.shape[:2]
        if max(h, w) > resize_max:
            scale = resize_max / max(h, w)
            h_new, w_new = int(round(h*scale)), int(round(w*scale))
            img = cv2.resize(img, (w_new, h_new), interpolation=cv2.INTER_LINEAR)

        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
        desc = sess.run(net_out, feed_dict={image_batch: img[None]})[0]
        desc = desc.astype(np.float32)

        name = path.relative_to(root)
        grp = hfile.create_group(str(name))
        grp.create_dataset('global_descriptor', data=desc)

    hfile.close()

if __name__ == '__main__':
    root = '/path/to/aachen/images/images_upright'
    output_file = 'aachen-v1.1_tf-netvlad.h5'
    inference(root, output_file)

I plan to port this NetVLAD model to PyTorch and integrate it into hloc, but this will not happen too soon - any contribution is welcome :)

[yopi1838]

Hi. I've been following this issue. Just to confirm my understanding, once we obtained the h5 file from this NetVLAD model, we just have to run the image pairs using the pairs_from_retrieval.py script, right?

[sarlinpe]

That is correct.

[xxlxsyhl]

The problem is solved. Thank you!