You can consider supporting LISRD and LoFTR, I personally think SuperGlue and SuperPoint are not very robust

[xmba15]

LoFTR can be better than (SuperPoint+SuperGlue), but I don't think it is true for LISRD and (SuperPoint+SuperGlue). I also intended to support LoFTR at some point.

[yeluoo]

Looking forward to seeing you support LoFTR soon

[ChmarsLuo]

@xmba15 me too，I also think that the effect of LoFTR is better, and I strongly urge the author to integrate it @xmba15

[sanersbug]

@xmba15 Me too, i also find that LoFTR and LISRD are more stable and better effect!

[xmba15]

Please stay tuned. WIP.

[xmba15]

First attempt convert kornia LoFTR's torch weights to onnx did not succeed due to

feat_f0_unfold = F.unfold(feat_f0, kernel_size=(W, W), stride=stride, padding=W // 2)
raise TypeError("iteration over a 0-d tensor")

This probably relates to disscussion in here: https://github.com/kornia/kornia/issues/1504

and the PR in here: https://github.com/kornia/kornia/pull/1758

[sanersbug]

Please stay tuned. WIP.

That sounds exciting！hope you can make it.

[sanersbug]

First attempt convert kornia LoFTR's torch weights to onnx did not succeed due to

feat_f0_unfold = F.unfold(feat_f0, kernel_size=(W, W), stride=stride, padding=W // 2)
raise TypeError("iteration over a 0-d tensor")

This probably relates to disscussion in here: kornia/kornia#1504

and the PR in here: kornia/kornia#1758

@xmba15 have you solved the problem? I meet the same error and i can't deal with it

[yeluoo]

@xmba15 did you solve it？

[xmba15]

Here we go. https://github.com/xmba15/onnx_runtime_cpp/pull/39

[yeluoo]

Hello, I found that LoFTR is not as effective as LISRD in matching large and small images, because the premise of LoFTR input is the same large image, such as [2, 1, 480, 640]. Can you support LISRD? In addition, you can set up to pay with Alipay

[yeluoo]

This problem occurred when converting the .onnx file. onnxruntime.capi.onnxruntime_pybind11_state.InvalidGraph: [ONNXRuntimeError] : 10 : INVALID_GRAPH : Load model from /data01/pot/onnx_runtime_cpp/loftr.onnx failed:This is an invalid model. Type Error: Type 'tensor(bool)' of input parameter (onnx::Mul_2684) of operator (Mul) in node (Mul_2221) is invalid. File "scripts/loftr/convert_to_onnx.py", line 54, in main sess = onnxruntime.InferenceSession("/data01/pot/onnx_runtime_cpp/loftr.onnx", providers=['CPUExecutionProvider']) File "/data01/software/anaconda3/envs/pytorch1.12/lib/python3.8/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 347, in __init__ self._create_inference_session(providers, provider_options, disabled_optimizers) File "/data01/software/anaconda3/envs/pytorch1.12/lib/python3.8/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 384, in _create_inference_session sess = C.InferenceSession(session_options, self._model_path, True, self._read_config_from_model)

[xmba15]

[2, 1, 480, 640]

LoFTR can accept any shapes as long as they are divisible by 8.

Type 'tensor(bool)' of input parameter (onnx::Mul_2684) of operator (Mul) in node (Mul_2221) is invalid.

You need to use the loftr repository that I provided as submodule here.

[submodule "scripts/loftr/LoFTR"]
    path = scripts/loftr/LoFTR
    url = https://github.com/xmba15/LoFTR

There are so many tricks related to this LoFTR so basically you need to follow all the instructions I have added in README.

[xmba15]

Maybe more people will still struggle so I will post another easy way to convert loftr weights using docker:

cd onnx_runtime_cpp
git submodule update --init --recursive
docker pull xmba15/onnx_runtime_cpp:v1.10.0-ubuntu20.04
docker run --rm -it -v `pwd`:/workspace xmba15/onnx_runtime_cpp:v1.10.0-ubuntu20.04
python3 -m pip install -r scripts/loftr/requirements.txt
python3 scripts/loftr/convert_to_onnx.py --model_path /path/to/indoor_ds_new.ckpt

Running without docker is fine. But if you have trouble building your own environment, running with the above commands from docker may save you some time.

[xmba15]

Or download the onnx weights from here. https://github.com/xmba15/onnx_runtime_cpp/releases/tag/0.0.3

[yeluoo]

Thanks, I'll try it.

[sanersbug]

Here we go. #39

Wow,that's great, can you provide a Alipay payment code.

[xmba15]

More about LOFTR's trained indoor weights performance. https://github.com/xmba15/onnx_runtime_cpp/issues/40