wzzheng / idml Goto Github PK

Hi Chengkun, your work is very novelty, I have a question that how the model learn uncertainty because there is no ground truth uncertainty label provided.

Hi, thank you for your interesting research.

I have some questions regarding the data augmentation:

1. In the paper, you stated as follows: "The training images were first resized to 256 × 256 and then augmented with random cropping to 224×224 as well as random horizontal flipping with the probability of 50%." But when reading the code, I see that you use the default dataset augmentation from the ProxyAnchor code, which does not contain resized image to 256 x 256 for is_train=True. Can you clarify it?
2. There is no mention, is the test data augmentation used the same one from the ProxyAnchor code.

Hi,
I think your work is promising and interesting. May I kindly ask what the difference is between your work and the Proxy Synthesis[1]? It seems you all try to search for a linear interpretation between different classes.

[1] Gu et al. "Proxy synthesis: Learning with synthetic classes for deep metric learning." AAAI 2021
github.com/navervision/proxy-synthesis

Thank you so much.

When running image_retrieval/code/train.py it shows the error
An error is reported in the opt.step() line of code, showing
local variable 'beta1' referenced before assignment

Thank you for your interesting research.

I would be very happy if you can answer some of my questions:

1. You have this paragraph in your paper:

"In addition, we fixed τ = 5 and set γ to 0, 1, 2, 3, 4 for training. The experimental results vary on the two datasets.
Specifically, our framework achieves the best performance when γ = 0 on the CUB-200-2011 dataset while γ = 3 on the Cars196 dataset. This indicates that the metric is more discreet when comparing images on the Cars196 dataset."

But on the left of figures 4. a and 4. b, we can see that the highest recall@1 corresponds to gamma = 2 for both CUB-200-2011 and Cars196 datasets. What is this different?

2. In your ProxyAnchor loss, you had set the gamma = 4, and tau = 5. Is this the set of hyperparameters you found has the highest performance on recall@1?