gtolias / how Goto Github PK

In my understanding, the paper contains two kinds of image-level similarity, one is measured by the global feature, another is measured by the local feature with ASMK.

In some research, similarity measured by global feature is enough. I guess that the reason you choose local feature similarity is the ASMK similarity can improve the performance of global feature. So I tested how much performance can be improved. The results is following.

HOW INFO: Evaluated roxford5k: mAP E: 65.56, M: 49.66, H: 24.41

HOW INFO: Evaluated rparis6k: mAP E: 80.64, M: 63.07, H: 36.0

The improvement is expected, but the performance of the global feature is too poor. I thought the mAP of roxford5k(M) and rparis6k(M) could be 60-70. This is a strange phenomenon that bad global feature can generate good local feature. I want to know more about it. Does it mean ASMK is a wonderful method?

The paper refers to using a batch size of 5, and indeed a batch size of 5 is requested from the dataloader.
However, the code explicitly iterates and backprops over each sample in the batch separately - effectively a batch size of 1.

The relevant code is here: https://github.com/gtolias/how/blob/master/how/stages/train.py#L91-L97

The ContrastiveLoss function used supports multiple queries - there's no reason to compute the loss per-tuple. Additionally, why does the model iterate over each image in the tuple instead of processing them as a batch?

Hi,
I would like to ask how to prepare a custom dataset to train the descriptor on? I want to finetune the network on my custom SfM dataset. Thanks.

Hello~ I read the codes and found that local features are selected based on the feature maps which have smaller resolution than that of the input images. But in the paper, Fig.3 illustrates local features on the original images. So how does that work? First rescaling the feature map as well as the heatmap and then selecting features?

I find query_ivf takes more time than extracting local features. For example, extracting features takes 14ms while query_ivf takes 66ms when searching in 3000 database images.
So how to speed up the searching process. Is it possible to move this to gpu? I find the local features are moved to cpu before query_ivf. Does ASMK support GPU?

Hi!

I am running HOW on a relatively large dataset and I would like to store the index into a file after it is built (and of course reload it).
is there a way of doing this?

Thanks!