Hi, Could you tell me how to evaluate the model on PF-willow dataset. Thanks. In your code, I just see the way to evaluate PF-pascal

Hello,

I have a question about the experiments on the Table 3: Average PCK comparison of different loss functions.

In the paper you stated that you fixed the parameters as (λmask = 3, λflow = 16, λsmooth = 0.5).

How did you set the parameters in your ablation studies?

For example, If you train only with the two losses (Lmask, L_flow), did you set the λmask and λflow as (3, 16) or did you find another combination?

Best regards,

Ahyun

What's the reason about 'corr_T2S = corr_feat3 * corr_feat4', and where can found the explanation in the paper?

thanks!

Hello,
I want to execute your code for train-phrase, but I get this error:

FileNotFoundError: [Errno 2] No such file or directory: './data/VOC2012_seg_img.npy'

Where can I get this file?

Thank you :)

Hi,
I can't find the Config file when reproducing your paper: "Bi-directional Contrastive Learning for Domain Adaptive Semantic Segmentation" in github: https://github.com/cvlab-yonsei/DASS
Can you upload the Config file? Thanks

Hello, do you use re-ranking or some tricks in your experiment? I can't reproduce the metric by using the config in your paper. In Market1501 dataset, I just got a 0.75 mAP with GCP and one-vs.-rest module. Expected for your reply, thank you.

I want to know the computation of contrastive feature.
Pcontrast = 1Pavg-1Pmax
or
Pcontrast = 1Pavg-1/6Pmax
or
Pcontrast = 6Pavg-1Pmax

Can you share the complete code?

Hi, could you open your dukemtmc-reid and cuhk dataset used for relation network? Thx :)

Thank you for sharing nice implementations!

I have questions about SFNet.

First, do you have the plan to provide the evaluation code of Caltech101?
Second, how did you make the corresponding map like Fig3 in the paper?

Thanks.

Thanks for the inspiring work and implementation.
I'm having troubles on reproducing the reported results of ISGAN using default configures without making any changes to codes.
I have trained models on Market and Duke, and was able to reproduce the stage1 (backbone) results as the paper reported, however, for stage3 model, I got mAP/rank-1 around 0.859/0.948 on Market and 0.773/0.884 on Duke, I also tested the provided trained final models, which achieved the reported results.
Can you shed some insights on how to reproduce the results? Is this because of the training configures should be different from provided defaults or maybe something else?
I'm currently using one gpu, I'm not sure if I should deploy the model into a multi-gpu setting as I didn't find DataParaller function within codes. Thank you.