The hesaff-pytorch from ducha-aiki

hesaff-pytorch's Introduction

This is implementation of Hessian-Affine local feature detector. It is heavily based on Michal Perdoch C++ implementation https://github.com/perdoch/hesaff

pytaff - current implementation hesamp - Michal Perdoch C++ one.

There are several differences:

No SIFT description, the output is image patches. If one needs to, patches could be feed into PyTorchSIFT
Subpixel precision is done via "center-of-responce-mass" inspired by LIFT paper, instead of original iterative quadratic fitting
Instead of setting threshold to control number of detection, this implementation simply outputs top-K local extreme points.

You also might be interested in HesAffNet, which gives significantly better results because of learned affine shape estimation procedure.

If you use this code for academic purposes, please cite the following paper:

@article{AffNet2017,
 author = {Dmytro Mishkin, Filip Radenovic, Jiri Matas},
    title = "{Learning Discriminative Affine Regions via Discriminability}",
     year = 2017,
    month = nov}

hesaff-pytorch's People

Contributors

Stargazers

Watchers

hesaff-pytorch's Issues

running error

Hi,

I have tried to run the command: python extract_features_oxaff.py graf1.ppm out.txt 2000. I got the following error:
Traceback (most recent call last):
File "extract_features_oxaff.py", line 36, in
LAFs, resp = HA(var_image)
File "/home/ryan/.local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 541, in call
result = self.forward(*input, **kwargs)
File "/home/ryan/hessian_affine_local_feature_detector/hesaff-pytorch-master/SparseImgRepresenter.py", line 200, in forward
responses, LAFs, final_pyr_idxs, final_level_idxs = self.multiScaleDetector(x,num_features_prefilter)
File "/home/ryan/hessian_affine_local_feature_detector/hesaff-pytorch-master/SparseImgRepresenter.py", line 84, in multiScaleDetector
scales = sigmas_oct[level_idx - 1:level_idx + 2])
File "/home/ryan/.local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 541, in call
result = self.forward(*input, **kwargs)
File "/home/ryan/hessian_affine_local_feature_detector/hesaff-pytorch-master/HandCraftedModules.py", line 253, in forward
num_of_nonzero_responces = (nmsed_resp > 0).sum().data[0]
IndexError: invalid index of a 0-dim tensor. Use tensor.item() to convert a 0-dim tensor to a Python number

Do you know the reasons for this error?

Thanks

Descriptor values?

I ran feature_extractor.py and got results like below:

1.0
906
411.1280822754 216.7289428711 0.0000048985 0.0000001165 0.0000061334
398.0391540527 505.0095214844 0.0103844600 0.0009393070 0.0077745551
342.8026428223 414.6062316895 0.0001460895 0.0000266796 0.0000589753
409.6105346680 654.5926513672 0.0000042293 0.0000002372 0.0000030523
439.0032958984 316.9766540527 0.0062913978 -0.0001653244 0.0018022966
...

I think this only shows affine regions. Can I get descriptor values for each region (n-d features) like hesamp code?

EDIT:

I read readme and I did:

got patches with "patches = extract_patches(x, LAFs, PS = self.PS)"
processed "forward" of SIFTNet for each patch

I think I solved it, but did I do it correctly?

Recommend Projects

ducha-aiki / hesaff-pytorch Goto Github PK

hesaff-pytorch's Introduction

hesaff-pytorch's People

Contributors

Stargazers

Watchers

Forkers

hesaff-pytorch's Issues

running error

Descriptor values?

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent