Hello, thank you for your excellent work, but I don’t quite understand the reordering network. Why should I choose max pooling and average pooling to refine features? What is its theoretical basis? Can you give some non-experimental explanations of interpretability?

Thank you for your great work! I am very interested in your re-ranking network. It seems to be weighted for retrieval features through preliminary ranking results, without involving more complex network optimization processes. Is my understanding correct? Is β among them learnable?

R

Hey @ShihaoShao-GH ,

First all of thanks for the paper, releasing your code on github and following up on issues. My knowledge of the image retrieval world is brand new, and my stats knowledge is basic, so forgive me if this is naive.

My questions relates to training a CVNET backbone with your modifications in place (GeM+, Rgem, ScaleGem). Is it as simple as plug and play. I have ported your existing code in forward and forward_singlesacle to here, with the same losses proposed in CVNET (CurricularFace, ContrastiveLoss), would this work?

Edit: After deeper consideration, nothing wrong with doing this

resnet.py的 264行有一个F.normalize(x6, p=2, dim=-1)，在原始CVNET没有这个操作

when train model, I have known to use arcface loss, but in the model, how to use the GeM?
whether the p is trainable when training, if yes, how to initialize it? if not, how to set the p for it?
if do not ise gem when training, then where the feature come from?

Thank you very much!

Hello,
I'm a newbie in the field of image retrieval. I'd like to run your code using my own dataset. I've downloaded the pre-trained CVNet weights as you've described and would like to use the CVNet model. My dataset consists of query images and a bunch of surrounding images for each query. Unlike the roxford5k dataset, it's not annotated. The query images are simply in a query folder, and the images that should go into the database are in a separate database folder. I'm not sure how to apply my custom dataset to your code in this situation. Additionally, you mentioned grid search in your paper, which I don't fully understand. If I use my custom dataset, do I need to perform grid search separately? Is there separate code available for this?

Hi,
Thank you for your great work and sharing your code.

I was able to reproduce mAP scores for ROxford and RParis datasets, also with +1M distractors, both with and without re-ranking. Unfortunately, I cannot reproduce mAP@100 scores for GLDv2 retrieval test set. I am around 0.8 below your scores even without re-ranking. The same applies when I try to reproduce original CVNet - I can reproduce ROP but not GLDv2.

I cannot figure out what I am doing wrong. Is there any special pre-processing of the GLDv2 images which is different to ROP images? For example some image resizing? I noticed that if I keep GLDv2 images in their original size, some of them have their smaller side too small and the extraction with your code crashes. This does not happen with ROP as they are generally larger images with less extreme aspect ratios.

Thank you for your assistance.

Thank you for your excellent work, I would like to ask you if you have the code for CVnet training? The official project of CVnet has not been provided

Hello, could you please provide SuperGlobal training scripts and tutorials? Thank you

Hi, do you have any idea to publish the training codes? Thanks.

Hi, I'm really sorry, after I read your paper and your open source code, I want to use the main framework of DELG-pytouch, and use the pooling method in your paper plus some of my own ideas to improve it, but after I download DELG's project and configure the environment according to his readme, I can't run it according to the steps he said. to run it, and then thought that you have run that code, so I would like to ask if the code you ran through is still there, or have you ever encountered a problem with not finding val_list.txt or any other problem that

Dear authors,

I recently came across your paper on SuperGlobal and found it to be particularly insightful. In addition, I have also familiarized myself with the contents of CVNET.

For CVNET, I was able to successfully execute the following command without any issues:
python test.py MODEL.DEPTH 50 TEST.WEIGHTS CVPR2022_CVNet_R50.pyth TEST.DATA_DIR ./revisitop/data/datasets
However, when attempting to replicate the results for SuperGlobal using the following command, I encountered an error suggesting an odd length for the parameter list. I would greatly appreciate your guidance in resolving this matter.
python test.py MODEL.DEPTH 50 TEST.WEIGHTS ./weights/CVPR2022_CVNet_R50.pyth TEST.DATA_DIR ./revisitop/data/datasets SupG.gemp SupG.rgem SupG.sgem SupG.relup SupG.rerank SupG.onemeval False
Below is the traceback for your reference:
Traceback (most recent call last): File "D:\test\github\SuperGlobal\test.py", line 17, in <module> main() File "D:\test\github\SuperGlobal\test.py", line 10, in main config.load_cfg_fom_args("test a CVNet model.") File "D:\test\github\SuperGlobal\config.py", line 145, in load_cfg_fom_args _C.merge_from_list(args.opts) File "C:\Users\u\.conda\envs\test\Lib\site-packages\yacs\config.py", line 223, in merge_from_list _assert_with_logging( File "C:\Users\u\.conda\envs\test\Lib\site-packages\yacs\config.py", line 545, in _assert_with_logging assert cond, msg AssertionError: Override list has odd length: ['MODEL.DEPTH', '50', 'TEST.WEIGHTS', './weights/CVPR2022_CVNet_R50.pyth', 'TEST.DATA_DIR', './revisitop/data/datasets', 'SupG.gemp', 'SupG.rgem', 'SupG.sgem', 'SupG.relup', 'SupG.rerank', 'SupG.onemeval', 'False']; it must be a list of pairs
Thank you in advance for your assistance.

how to train model on my own data

Hello again!
Thanks for your helpful response to my previous question. Before building my own pipeline, I'm studying your code, but some parts are unclear.

1. Regional-GeM
   In the rgem.py file, the code looks like this:

class rgem(nn.Module):
    """ Reranking with maximum descriptors aggregation """
    def __init__(self, pr=2.5, size=5):
        super(rgem, self).__init__()
        self.pr = pr
        self.size = size
        self.lppool = nn.LPPool2d(self.pr, int(self.size), stride=1)
        self.pad = nn.ReflectionPad2d(int((self.size-1)//2))
    def forward(self, x):
        nominator = (self.size ** 2)

I thought nominator was for normalization, but I don't understand the role of **(1./self.pr).

2. Paper and Code Alignment
   In your RerankwMDA.py file, both X1 and X2 appear to be refined. According to the paper, I thought res_rerank would correspond to S2 and be computed with the original global descriptor(g_d), not the refined one(g_dr).

I'm sure I'm missing some details, so I'd appreciate your explanation of these aspects.

Thank you!

请问下，这个Oxford+1M Hard数据集，是根据 https://github.com/filipradenovic/revisitop 这个生成的是吧。好像原生的数据集只有Oxford 5k 和 Paris 6k这两个，其他的数据集都是造出来的。但是这两个好像都不让下载.

还有一个问题想问下，如果我没理解错的话，这个数据集是带有监督标签的, 大家的网络模型都在这个数据集上训练。但是现实场景中，图片数据集通常是没有标签的，只是做一个相似度召回系统，是否有考虑到这种情况。这样的话很多模型结构都不成立了，因为没法训练。

I get +1m score based on resnet50 with a little gap from the paper.
Roxford
Without rerank
Retrieval results: mAP E: 87.98, M: 73.13, H: 50.75
With rerank
Retrieval results: mAP E: 92.11, M: 78.7, H: 62.36

Rparis
Without rerank
Retrieval results: mAP E: 90.6, M: 79.3, H: 62.62
With rerank
Retrieval results: mAP E: 92.71, M: 82.24, H: 68.0

Thank you for the great work you do. It seems that your reranking network does not require training, and there is no corresponding loss. I used the same processing logic. It seems that the reranking network for my extracted global features is not very ideal.