search

jiawei151 / SGIEL_VIReID

Official PyTorch Implementation of "Shape-Erased Feature Learning for Visible-Infrared Person Re-Identification" (CVPR'23)
MIT License
38 stars 3 forks source link

SYSU_MM01_SHAPE #2

Closed Mrkkew closed 1 year ago

Mrkkew commented 1 year ago

I want to ask how to obtain SYSU_MM01_SHAPE and SYSU_MM01_MASK ?

jiawei151 commented 1 year ago

I want to ask how to obtain SYSU_MM01_SHAPE and SYSU_MM01_MASK ?

Hi,@Mrkkew .Please refer to 'Body Shape Data' in README, and only SYSU_MM01_SHAPE is required. We will release this dataset as soon as possible.

Velpro-collab commented 1 year ago

I would like to know how to obtain the human parsing of VCM-HIT. Is there any detailed instructions？

jiawei151 commented 1 year ago

I would like to know how to obtain the human parsing of VCM-HIT. Is there any detailed instructions？

Hi @Outsider-shroud . There is no difference across datasets. We utilized the same preprocessing one-to-one mapping strategy.

Velpro-collab commented 1 year ago

Does infrared image generate human parsing? But the effect of Self-Correction-Human-Parsing is poor.

jiawei151 commented 1 year ago

Does infrared image generate human parsing? But the effect of Self-Correction-Human-Parsing is poor.

I think it can perform better if the pretrained SCHP is tuned on infrared data, which we haven't tried yet. You may have a try if you were interested. Actually, they need not be too good cause they are just as a guidance and not used during testing.

Velpro-collab commented 1 year ago

vcm png Hello, is the baseline you use only AGW? May I ask if there are any adjustments? I used the baseline（AGW） to verify that rank1 was only 53% in the HIT-VCM data set。

jiawei151 commented 1 year ago

vcm png Hello, is the baseline you use only AGW? May I ask if there are any adjustments? I used the baseline（AGW） to verify that rank1 was only 53% in the HIT-VCM data set。

Please refer to implementation details or our code. The baseline for 3 datasets was implemented on a ResNet50 backbone and BNNeck head with additional data augmentations.

Velpro-collab commented 1 year ago

Hi, I can't reproduce the body shape dataset, can you share this part of the code?

Mrkkew commented 1 year ago

Can you share your segment body shape datasets of SYSU-MM01 and RegDB?

jiawei151 commented 1 year ago

Hi @Mrkkew and @Outsider-shroud , we just uploaded a link for the body shape data in README.