Hi! I would like to ask for guidance on reproducing Aachen for SIFT and SGMNet+SIFT?

At a starter, I have tried to run MNN+SIFT for Aachen day-night...but I get a much worse result for 8196 keypoints... 23.5 / 33.7 / 41.8 I am guessing that maybe because I am using SIFT instead of RootSIFT... But I am not sure about other settings that could be different...

So, as I saw the result of SGMNet+SIFT as well as MNN+SIFT on both papers and on https://www.visuallocalization.net/details/17655/, I am wondering how to reproduce them (especially Table 4. in the paper)...

Could you please share the setting? I would like to learn how to get a similar result...

So I would like to scope down my question as follows:

• Firstly, it is RootSIFT not SIFT...where self.root == True (in https://github.com/vdvchen/SGMNet/blob/main/components/extractors.py#L43 ) ? Is that correct? Do you have to further normalize the features (to keep the norm of feature dimension to 1) ?
• Did you restrict image size? For example, in D2-Net, the maximum size was restricted to 1600 ? Did you do something like that too?
• For MNN+SIFT, is there any thresholding applied for MNN matching?
• Is the setting for SGMNet+SIFT for Aachen Day-Night similar to the following? Also, as I read the name ...rootsift8k_upright_512_0.2_SGMNet... What do you mean by 512 and 0.2?

matcher:
  name: SGM
  model_dir: ../weights/sgm/root
  seed_top_k: [256,256]
  seed_radius_coe: 0.01
  net_channels: 128
  layer_num: 9
  head: 4
  seedlayer: [0,6]
  use_mc_seeding: True
  use_score_encoding: False
  conf_bar: [1.11,0.1] #set to [1,0.1] for sp
  sink_iter: [10,100]
  detach_iter: 1000000
  p_th: 0.2

• Also, is the setting for SG+SIFT for Aachen Day-Night similar to ?

matcher:
  name: SG
  model_dir: ../weights/sg/root
  net_channels: 128
  layer_num: 9
  head: 4
  use_score_encoding: True
  sink_iter: [100]
  p_th: 0.2

Hi! Thank you so much for releasing the code. Your paper is very impressive and contains so many interesting findings.

Here, I would like to kindly ask for the following setting in using SGMNet.

1. SGMNet+SP
2. SGMNet+SP-10 sink

I have tried running your work with HPatch it seems pretty good with the setting of SIFT (but switch the dimension to 256). However, I don't know what is the proper setting.

@vdvchen 你好， 当运行SGMNet+SP时出现了一个错误，应该是因为superpoint是256维特征，但是SMGNet的权重是128维的，请问方便发送我邮箱[email protected]一个基于superpoint的SGMNet权重么？

aug_desc1, aug_desc2 = x1_pos_embedding + desc1, x2_pos_embedding + desc2 RuntimeError: The size of tensor a (128) must match the size of tensor b (256) at non-singleton dimension 1

Why can't I find the hdf5 file？Does the code not generate hdF5 files in the folder?

Traceback (most recent call last): File "dump.py", line 27, in dataset.format_dump_data() File "/data5/huZhao/code/SGMNet/datadump/dumper/gl3d_train.py", line 244, in format_dump_data pool.map(self.format_seq,indices) File "/data5/huZhao/anaconda3/envs/sgmnet/lib/python3.7/multiprocessing/pool.py", line 268, in map return self._map_async(func, iterable, mapstar, chunksize).get() File "/data5/huZhao/anaconda3/envs/sgmnet/lib/python3.7/multiprocessing/pool.py", line 657, in get raise self._value FileNotFoundError: [Errno 2] Unable to open file (unable to open file: name = '/data5/huZhao/code/GL3D-2/dump_desc_dir/000000000000000000000009/00000010.jpg_root_1000.hdf5', errno = 2, error message = 'No such file or directory', flags = 0, o_flags = 0)

Thanks for your kind support last time and thank you very much for sharing the training script... It is quite interesting for me.
Here, I would like to kindly ask about the data for training. As I have tried to follow the instruction to download the data from https://github.com/lzx551402/GL3D

• I have a question on which of these three datasets (1) gl3d_imgs, (2) gl3d_raw_imgs (3) gl3d_blended_images from https://github.com/lzx551402/GL3D#downloads ...... are to be downloaded ... or all of them ?

• I have downloaded gl3d_raw_imgs... However, I received the error (below).... Does this mean that I did not download correctly? Or that I have downloaded the wrong dataset?

• My setting for gl3d.yaml file is as follows. Should rawdata_dir be the cloned directory of https://github.com/lzx551402/GL3D ? I am very sorry as this is not what you wrote in the instruction. The reason that I thought that this maybe the GL3D cloned directory is because dump.py also looks for GL3D/data/list/comb/imageset_train.txt... :

data_name: gl3d_train
rawdata_dir: /mnt/HDD4TB2/GL3D   
feature_dump_dir: /mnt/HDD4TB3/SGMNet/gl3d_desc_dir
dataset_dump_dir: /mnt/HDD4TB3/SGMNet/gl3d_dataset_dir

The error:

python dump.py --config_path configs/gl3d.yaml
dump.py:20: YAMLLoadWarning: calling yaml.load() without Loader=... is deprecated, as the default Loader is unsafe. Please read https://msg.pyyaml.org/load for full details.
  config = yaml.load(f)
Formatting data...
  0%|                                                                                                                                                     | 0/109 [00:00<?, ?it/s]
multiprocessing.pool.RemoteTraceback: 
"""
Traceback (most recent call last):
  File "/home/gabby-suwichaya/anaconda3/envs/sgmnet/lib/python3.7/multiprocessing/pool.py", line 121, in worker
    result = (True, func(*args, **kwds))
  File "/home/gabby-suwichaya/anaconda3/envs/sgmnet/lib/python3.7/multiprocessing/pool.py", line 44, in mapstar
    return list(map(*args))
  File "/mnt/HDD4TB3/SGMNet/datadump/dumper/gl3d_train.py", line 147, in format_seq
    pair_list=np.loadtxt(os.path.join(seq_dir,'geolabel','common_track.txt'),dtype=float)[:,:2].astype(int)
  File "/home/gabby-suwichaya/anaconda3/envs/sgmnet/lib/python3.7/site-packages/numpy/lib/npyio.py", line 1067, in loadtxt
    fh = np.lib._datasource.open(fname, 'rt', encoding=encoding)
  File "/home/gabby-suwichaya/anaconda3/envs/sgmnet/lib/python3.7/site-packages/numpy/lib/_datasource.py", line 193, in open
    return ds.open(path, mode, encoding=encoding, newline=newline)
  File "/home/gabby-suwichaya/anaconda3/envs/sgmnet/lib/python3.7/site-packages/numpy/lib/_datasource.py", line 533, in open
    raise IOError("%s not found." % path)
OSError: /mnt/HDD4TB2/GL3D/data/586326ad712e276146904571/geolabel/common_track.txt not found.
"""

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "dump.py", line 27, in <module>
    dataset.format_dump_data()
  File "/mnt/HDD4TB3/SGMNet/datadump/dumper/gl3d_train.py", line 244, in format_dump_data
    pool.map(self.format_seq,indices)
  File "/home/gabby-suwichaya/anaconda3/envs/sgmnet/lib/python3.7/multiprocessing/pool.py", line 268, in map
    return self._map_async(func, iterable, mapstar, chunksize).get()
  File "/home/gabby-suwichaya/anaconda3/envs/sgmnet/lib/python3.7/multiprocessing/pool.py", line 657, in get
    raise self._value
OSError: /mnt/HDD4TB2/GL3D/data/586326ad712e276146904571/geolabel/common_track.txt not found.

When I run python dump.py --config_path configs/gl3d.yaml

I encounter the following issue, seems like it can not find hdf5 file how to solve it?

Traceback (most recent call last): File "/usr/lib/python3.6/multiprocessing/pool.py", line 119, in worker result = (True, func(*args, **kwds)) File "/usr/lib/python3.6/multiprocessing/pool.py", line 44, in mapstar return list(map(*args)) File "/home/sshuang/SGMNet/datadump/dumper/gl3d_train.py", line 192, in format_seq with h5py.File(os.path.join(self.config['feature_dump_dir'],fea_path1),'r') as fea1,
File "/usr/local/lib/python3.6/dist-packages/h5py/_hl/files.py", line 394, in init swmr=swmr) File "/usr/local/lib/python3.6/dist-packages/h5py/_hl/files.py", line 170, in make_fid fid = h5f.open(name, flags, fapl=fapl) File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper File "h5py/h5f.pyx", line 85, in h5py.h5f.open OSError: Unable to open file (unable to open file: name = '/home/sshuang/SGMNet/datadump/dump_desc/000000000000000000000009/00000007.jpg_sp_500.hdf5', errno = 2, error message = 'No such file or directory', flags = 0, o_flags = 0)"""

The above exception was the direct cause of the following exception:

Traceback (most recent call last): File "dump.py", line 27, in dataset.format_dump_data() File "/home/sshuang/SGMNet/datadump/dumper/gl3d_train.py", line 244, in format_dump_data pool.map(self.format_seq,indices) File "/usr/lib/python3.6/multiprocessing/pool.py", line 266, in map return self._map_async(func, iterable, mapstar, chunksize).get() File "/usr/lib/python3.6/multiprocessing/pool.py", line 644, in get raise self._value OSError: Unable to open file (unable to open file: name = '/home/sshuang/SGMNet/datadump/dump_desc/000000000000000000000009/00000007.jpg_sp_500.hdf5', errno = 2, error message = 'No such file or directory', flags = 0, o_flags = 0)

Hello, I read your paper, there is a combination of SIFT+Superglue in your comparative experiment, how do you solve the problem that the descriptor and matching network dimensions are inconsistent

Hi! Can you answer a couple of questions

1. How many epochs did you use for training the SGMNet? And where to set this parameter in the code?
2. When checkpoints are made? Because I didn't understand from the code.