kolkir / coarse_loftr_trt Goto Github PK

I tried webcam demo script, and found that only coarse match results are displayed, all points seems to lie on predefined grid, even with the original pretrained weights, predefined grid seem to be denser, matched points are not as good as the original LoFTR, do we miss any refine step in the eval script ?

HI
Can u provide end-to-end example for image1<->image2 matching
thx

When I use the trt method to run webcam.py, I encounter the above error, how can I solve it

Hi author,
Your repo is veryhelpfull for my work, I have a question about ONNX model in your repo.
I find input of data in onnx model 2 image have shape 1,480,640
i dont understand output of model onnx(2 tensor with shape 1,1200,1200 ).
what output meanning. it is output of what step in LOFTR origin repo?
Thank you so much

when i train the code ，the program feedback such information：
0%| | 0/1250 [00:05<?, ?it/s]
Traceback (most recent call last):
File "C:\Users\aust\Desktop\Coarse_LoFTR_TRT-main\train\trainer.py", line 154, in train_loop
self.scaler.scale(loss).backward()
File "C:\Users\aust\anaconda3\lib\site-packages\torch_tensor.py", line 255, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
File "C:\Users\aust\anaconda3\lib\site-packages\torch\autograd_init_.py", line 147, in backward
Variable._execution_engine.run_backward(
RuntimeError: Function 'KlDivBackward' returned nan values in its 0th output.
python-BaseException

it looks like the loss function has problem?

Hi, does anyone now what the accuracy tradeoff is between the original loftr and the coarse loftr?

Hello, I would like to ask how do I create my own dataset, what should be the size of the training samples, and how do I get the depth map of the image?

hi @Kolkir , nice work!
Could you explain why you didn't keep the fine level matching in this work? Is it difficult to achieve?

Hello,thank for your work!
I ran export_onnx.py, and the model(LoFTR_teacher.onnx ) size is 184MB, but the model size you provide is 1.76MB.
Could you fix it? Thanks a lot!

hi,professor:
a new image matching algorithm called ASpanFormer, ASpanFormer: Detector-Free Image Matching with Adaptive Span Transformer
github: https://github.com/apple/ml-aspanformer
this is amazing, please deploy this on nvidia jetson platform, the pipeline like: pytorch -> onnx -> tensorrt engine , then c++ deploy,
PLEASE!!!