Hi, I met a gap between val and test in Medadepth dataset. I recently use 4 gpus to train the model for 15 epochs. The map(@20) is 0.766 on val_list and 0.720 on test_list. I want to know if it is normal?

Hi,

First I'd like to sincerely thank the author for sharing this amazing work. When I tried the provided colab with default setting (indoor), the resulting plot was quite poor.

Similar problem occurred when I cloned the repo and tested on both indoor and outdoor pairs on indoor_ds.ckpt and outdoor_ds.ckpt

Can anyone advise on this? Thank you very much!

Thank you for your amazing work. I have some question that the expec_f_gt after spvs_fine() function are all very large number. And this case cause the correct_mask in computing fine loss is zero.

In your paper, the evaluation comparison demonstrated includes the accuracy of the prediction. But I opened loftr_ pred_ eval. NPY found that the only content in it was' pair name '', identifier '', mconf '', mkpts1f '', mkpts0f '', EPI '', R '', tT'', inliers'。 How to show the prediction accurac？ y

I have tried matching 2 (2000*2000) images. It took around 80 gb memory for finding 17k matched keypoints. Attached the screenshot.. was that normal?

Hi, thanks for your great work. I tried to reproduce the training code, and then I met some problems:

1. I tried to set match threshold to 0.2 without loading pretrained model, and I found there will be 0 matches from coarse mathcing layer.
2. If I load the pretrained model to train, the Loss_coarse would go up while Loss_fine going down, and as training time goes on, the num of points from coarse matching layer reduces to 0.
3. Should I use gradient clipping?

Thanks for amazing work! About the reproduction for the visual localization work with hloc, I have some confusions. As for my understanding, loftr will give a match result include local keypoints0, local keypoint1 and their matching confidence although the matching process is "local features" free. So the local key point locations are concerned with the image pairs. Different image pairs may produce different keypoints. Like pair (image a, image b) and pair (image a, image c) may result totally different keypoints distribution on image a. During the sfm procedure in hloc,, triangluation process requires long keypoints tracks to ensure the 3D points' quality. So how to handle this situation while different keyframe pairs produce variance keypoints for one keyframe.
Is there any strategy or principle to combine the keypoints which are close distributed?

What exactly would I need to provide to be able to train LoFTR on a custom dataset? It was pointed out in some of the previous issues that all one would need would be depth map, extrinsics and intrinsics for every image and the ground truth is generated on the fly.

The npz files seem to have fields - 'image_paths', 'depth_paths', 'intrinsics', 'poses', 'pair_infos'. Assuming 'poses' is just the extrinsics (R | t), how exactly is 'pair_infos' generated and how are the covisibility scores calculated?

Hi, thanks for the fabulous work!

I was looking for works leveraging dense feature matching operations and ended up in this repository. I am new to the feature matching task, so please be patient with me :)

1. What is the purpose of bordering the matching mask? (The mask_border() function)

2. In the function, does indexing [..-b:0..] apply borders on any portion of tensors? If it is intended, I don't get the purpose of it.

Thanks!

Hi, I have tried to use np.savez_compressed() to save the npz and read data from the npz I created, it still is NpzFile object, not dict object. this issue

Hello, Thank you for the great works.

I have tried to use your method for outdoor images. However, after bug fixing, it seems that the weight "outdoor_ds.ckpt" is not working as good as it used to be... Therefore, could you please provide the update for the "outdoor_ds.ckpt" as well?

There seems to be some error when downloading the weights:

FileNotFoundError                         Traceback (most recent call last)
[<ipython-input-29-011ecbd77aeb>](https://localhost:8080/#) in <module>
      6   matcher.load_state_dict(torch.load("weights/indoor_ds.ckpt")['state_dict'])
      7 elif image_type == 'outdoor':
----> 8   matcher.load_state_dict(torch.load("weights/outdoor_ds.ckpt")['state_dict'])
      9 else:
     10   raise ValueError("Wrong image_type is given.")

2 frames
[/usr/local/lib/python3.8/dist-packages/torch/serialization.py](https://localhost:8080/#) in __init__(self, name, mode)
    249 class _open_file(_opener):
    250     def __init__(self, name, mode):
--> 251         super(_open_file, self).__init__(open(name, mode))
    252 
    253     def __exit__(self, *args):

FileNotFoundError: [Errno 2] No such file or directory: 'weights/outdoor_ds.ckpt'

I want to extract the key points and corresponding feature vector using the demo. My question is the following, how can you do it using the notebook demo. I know perhaps keypoints are returned in the dictionary ( mkpts0 = batch['mkpts0_f'].cpu().numpy() mkpts1 = batch['mkpts1_f'].cpu().numpy()).

How can I get the correponding feature vector. Perhaps in loftr.py these (feat_f0_unfold, feat_f1_unfold) are corresponding feature vectors ? could you please guide me.

I am someone who is new to the field, so excuse me if the questions are super obvious.

Greetings there!

I want to calculate a similarity score between two images (i.e. as we calculate the cosine similarity). So that an image can be said x percentage similar to another image. A score with a fixed range i.e. may be -1 to 1 or 0 to 1. Is there any way to calculate such a score?

Hello,

I am trying to train the LoFTR model and run into KeyError: 'spv_b_ids'. On other Github issues, I saw that LoFTR to be used in eval() mode, however, I want to train the LoFTR model. What's the appropriate solution?

Thank you.