Hi, I have a problem about the generated target parsing results of your pre-trained model for human pose transfer. Using your pre-trained checkpoint, I visualize the generated target parsing results. (i.e., self.parsav in class Painet(BaseModel)). As shown in the figure, however, it seems to exist some problems.

1、 It seems that the ParsingNet can only effectively generate parsing maps of some limited regions(e.g., '3':upper clothes and '5' : lower clothes(pants, shorts)), but cannot tackle other regions(e.g., skin, face, hair etc.). 2、It seems that the generated target parsing result is offset to the left relative to GT, which should be located in the middle of the image. In other words, the generated target parsing result is not aligned with the input target pose(i.e., self.input_BP2) in the spatial position. In fact, using your pre-trained checkpoint, the generated target image result is also offset to the left relative to GT, as shown in the figure.

I’m not sure if it’s the problem with your model？ Please check it. I would be very grateful if you can provide your visualization results！

Thanks！

I want to test only the texture transfer model. It is mentioned that the model needs to be changed. How should I change it? What inputs are needed to test the texture transfer model?

while training we are getting this error eventhough we have downloaded datasets and put it in fashiondata directory and in train. FileNotFoundError: [Errno 2] No such file or directory: './dataset/fashion_data/train/fashionWOMENJackets_Coatsid0000658203_3back.jpg' Here Women and jackets are directories but it is showing as single name , and also 03 should be in folder of id_00006582

Hi,

The correspondence loss of image generator in paper section 3.3 is not used in your code.

self.loss_names = [ 'app_gen','content_gen', 'style_gen', #'reg_gen', 'ad_gen', 'dis_img_gen', 'par', 'par1']

Does the correspondence loss not impact the final generated image? Thanks for your reply.

Hi, nice job!

I was sorry to bother you, when I used the test instruction， I can get the eval_results. I was a little confused to get the results as your paper. May I use other calculate the metric program?

Thank you very much!

Great work！I have some questions about the data size. 1、In my opinion, the loadsize of input image, pose map, and parsing map is all 256x256 in your method. However, the key points annotations are obtained from the cropped images with the resolution of 176x256, which means the oldsize should be 176x256. However, why do you set parser.set_defaults(old_size=(256, 256)) in fashion_dataset.py ? 2、Your parsing maps are obtained from the cropped images with the resolution of 176x256, and then padding to 256X256. Is it right? 3、The original images of DeepFashion dataset(256x256) have the backgrounds with inconsistent colors. Will it have a bad effect when used directly for training? Need I crop them to 176x256, and then padding them to 256X256？ Thanks very much!

Hi! Could you tell me the accurate correspondences between different parsing labels and indexes in your provided parsing data? The parsing labels mentioned in your paper are 'Hair', 'Upper Clothes', 'Dress', 'Pants', 'Face', 'Upper skin', 'leg' and 'background', which seems differs from that in your provided parsing data. For example, in your provided parsing data, you set the 'Shoes' to a separate category，meanwhile combine the arm and leg skins into the same category.

Besides, when I visualize your provided parsing data, the area of region with index == 1 seems always equal to 0. Please check it. PS: I don't solve my problem from the similar issue like #2

Thanks!

Hi. Thank you for providing your excellent paper and its code to the world. It's so exciting to touch your paper for me.

Now, I use your code and Pre-trained checkpoint of human pose transfer from here and its result is not looks good than your result reported in paper.

Is there something I have to do for getting more good result? Do i have to retrain model from pre-trained one?

Below is results from pre-trained model.

There are also good ones.

Hi, for region editing, gievn a certain parsing map, how do you edit it to obtain the desired parsing maps？Can you share the scripts or tools you used? Thanks!

Hi, I have some problems about the difference between the pose transfer model and the texture model. 1、As you said, we can uncomment line 162-176 to train a texture model. In my opinion, it mainly changes the predicted par2 from Float to Int by the torch.argmax operation. What's the advantage or motivation of you doing it rather than use the predicted par2 directly? 2、Since argmax operation is non-differentiable, if we uncomment line 162-176 to train a texture model, the image generator can not provide the gradients for the parsing generator. Thus, the pre-trained parsing generator will not be updated during the training. Will it affect the quality of final generated images? Besides, since the parsing generator are disconnected from the image generator, need we to calculate the parsing loss like loss_par and loss_par1 when tarining the image generator?

Thanks!

Hi, In your paper, you "concatenate the source image Is, the source parsing map Ss, the generated parsing map Sg and the target pose Pt in depth (channel) dimension and extract its feature Fp ", as shown in the Fig.

However, in my opinion, Fp should aims to provide the target pose information. Why do you additionally use the source image Is and the source parsing map Ss as input? Do you try to only use Sg and Pt to extract Fp ? Thanks!

Hello,

First of all, thank you for your work. I read in your paper that you first train the parsing generator and image generator respectively, then perform an end-to-end training. However, I was not able to locate in your code the parts that handle the switch of training strategy. The reason for this question is that I would like to train only the parsing generator first and would not like to change everything in the code if the option is already there.

Thank you for your help

Hi, I have read your paper and tried to implement it. But in order for it to be used for virtual try onns I tried to take this project to next level by adding a person identification of generated image in new pose with respect to the real image of the same person in order to calculate the final accuracy of the model?? So can you give me any idea on how to approach this?

Hello,

Thanks for sharing the great work. I'm trying to generate the samples using the pretrained model. But unfortunately my results are so dim, something like this: Mine: Previously available by the authors:

Any help is greatly appreciated by the entire community.