Question about evaluating the model on roxford dataset about affnet HOT 11 CLOSED

Hi,

Well, that is expected, as your pipeline is completely different from one used in paper:

1. Detection was used with constant threshold 5.33 on Hessian response instead of just 3000k per image
2. Matching was done by BoW scheme + Hamming embedding
3. Top-1000 matches very spatially verifying (similar to what you have done)
4. For verified matched, query expansion was done.

[13] is https://hal.inria.fr/hal-00971267/PDF/tolias_lqe14.pdf

from affnet.

Thanks, but i wonder if just use the simple strategy without using the methods you have mentioned, what the result will be.

from affnet.

Your simple strategy is not scalable, unfortunately.
But if you want to improve results:

1. make sure that you do not use orientation detector, but gravity-vector (do_ori = False)

2. opencv ransac is not good. Try this one https://github.com/ducha-aiki/zeromqransac or this https://github.com/cr333/usac-cmake

from affnet.

@ducha-aiki Thank you very much! I will try recently.

from affnet.

@ducha-aiki When the feature is extracted as what you said(do_ori = False, th=5.33), I have improved 10% of mAP. And i am trying to realize your BoW scheme + Hamming embedding + SV process.

1. But in the experiments, i found the feature with affnet(AffNet=AffNetPix) or without affnet(AffNet=None) just have a little difference. I wonder if i do something wrong. As i checked, the output of LAFs (x, y, a, b, c) and descriptors are different. Need some other operations?
2. If 'do_ori = False', then the keypoints extracted don't include orientation information and OriNet is no need to use, is it right?

from affnet.

1. It is possible for some images: if not viewpoint change, AffNet might produce circular shape (aka no affine)
2. yes

@ym547559398 I have found my old private repo with naive implementationretrieval similar to you described in first post, but with query expansion and SV. It reads the extracted features in specific format and then does benchmark on new oxform5k/paris6k
It gives following results:

oxford5knew mAP E: 82.1, M: 70.26, H: 51.84
oxford5knew mP@k[1, 5, 10, 20, 100] 
E: [94.12 76.47 72.06 61.76 47.06],
M: [95.71 88.57 74.29 55.71 24.29],
H: [90.   65.71 44.29 25.71 14.29]

If you want, I can share with you that private repo. But keep it mind, that it is completely undocumented, no guarantees and probably I would have no time to answer the questions about those code

from affnet.

@ducha-aiki That's great! I will compare your private code with mine carefully. And i just have some questions with your QE and SV process, if you share, i can learn it myself. Also, i have implemented a simple pipeline of bow+he+sv process during these days. Thank you for patient reply and wonderful work again! 'AffNet' is beyond this and look forward to your further research.

from affnet.

@ducha-aiki And you can send to [email protected] or [email protected]. Thank you~

from affnet.

I have sent you github invite. If I remember correctly, you can run test_534h.py file

from affnet.

@ducha-aiki OK，thanks

from affnet.

@ducha-aiki Hi, i have tested the code this week. And I got a final result of 67.77 with qe and sv. The base is similar to my process, although 67.77 is much higher than base with your QE process, it still has a gap to get your result. I have checked the process of extracting feature again, the parameters are same with yours, i wonder if there is something i still ignored. Hope to get your help again!

from affnet.