vbalnt / tfeat Goto Github PK

The test accuracy is always change. Did you get the average precision of the test data ?

I mean, If we train 10 epoch, we can get a precision. If we train 20 epoches, we can get another result. How did you solve this problem ?

Hi, thank you for the great work first. May I ask two questions:

1. Basically ratio loss is unsuitable (doesn't converge at all) for keypoint matching according to your paper, so the best choice is ranking with anchor swap. Am I correct?
2. How are the training patches from Lowe's DoG keypoints extracted? Is it just a fixed-size window cropping or based on the keypoints' scale ( possibly orientation as well?) followed by a normalization step?

Thanks in advance.

Dear vbalnt,
Thank you very much for sharing the code.
I met a problem that the link of "patches-dataset" in the web "https://github.com/vbalnt/UBC-Phototour-Patches-Torch" is invalid. Can you share me a new one?
Thank you,
Best regards,
Dandan li.

Could you send the link to the input_video.webm and object_img.png files?

Because the software presents this message:
File "tfeat_demo.py", line 220
print "Not enough matches are found with TFEAT - %d/%d" % (len(good1), MIN_MATCH_COUNT)

Paper says:

. For the optimization the Stochastic Gradient Descend [3] is used, and the training is done in batches
of 128 items, with a learning rate of 0.1 which is temporally annealed, momentum of 0.9
and weight decay of 10−6 . We also reduce the learning rate every epoch

Could you please clarify, which exactly factor is used for reducing learning rate schedule? And which epoch snapshot is the one used for getting results in paper?

I am trying to run tfeat_demo.py 2 days already with no luck.
My configurations are Ubuntu 18.04 Cuda 9.0 Cudnn 7.05 Python 2.7 Torch 7
Please can you show details on all dependencies with versions and steps.
There is also a syntax error when running tfeat_demo.py and get_nets.sh not working.
My goal is to run demo in opencv as shown in youtube demo

Provided tensorflow variant is great, but it doesn`t work with t7 weights :(

When do the "UBC all" trained params will be ready? The benefit of the descriptors like SURF and SIFT is that they are independent of dataset and your descriptors are kinda overfitted over different UBC datasets. I'm getting better results with SIFT/SURF descriptors on most of my images and thus TFeat is not production ready. It seems currently the yosemite pretrained parameters are best, I would like to try "UBC all", meanwhile I will stick with SIFT descriptors.

Hint: For scientific success to win the battle against SIFT I recommend you to train on RGB patches instead of Grayscale, but seems it will be hard to get such dataset? It just requires to change 1 parameter from 1 to 3 in nn.Conv2d layer:

        self.features = nn.Sequential(
            nn.InstanceNorm2d(1, affine=False),
            nn.Conv2d(3, 32, kernel_size=7),
            nn.Tanh(),
       ...

these days,i have tried to train the network with 1281283 images with sgd and i generate 1.28M triplets.
however,it's so difficult to convergence.
can you give me some tips to train~

Hello,

I evaluated the pretrained models that you provide in the repo but the results are not the same as in the paper. Are the pretrained models the same used in the paper?

The strange part is actually I get lower results (1-2 %)

Thank you

@vbalnt
Thank you for providing such wonderful work. You have provided a script to show the correspondence, it shows the TFeat descriptor is obviously better than BRISK. However, if I use SIFT to detect the keypoint, I found the performance is really worse.

Could you give me some advice on how to solve this problem ?

Dear vbalnt,

Thank you very much for open sourcing the code, it is very easy to read.
I would like to point out a certain thing. In the eval section, there needs to be an update of the offset ( offset += FLAGS.batch_size ).

Thank you again,
Best Regards,
ManyIds

Hello, I think the website "http://www.iis.ee.ic.ac.uk/" is unreachable, can you post a new website which can offer the .t7 model for me? Thank you very much!

First of all, thanks for open-sourcing your code. I have two questions/issues regarding tfeat_demo.py:

1. The extracted patches should be normalize to the ranges the network was trained on. In training, I belive one uses [0,1] range and subtracts mean (~0.48) and divides by stddev (~0.18). In testing, opencv works with [0,255] range and L75 just subtracts mean of each patch.

2. ORB may be involuntarily disadvantaged by using improper matching (cv2.NORM_HAMMING is recommended over cv2.NORM_L2, which is useful for tfeat).

Hi, thank you for sharing the code. The download link in get_nets.sh seems not working. Could you please check it or provide another link? Thanks a lot!