JDAI-CV / Down-to-the-Last-Detail-Virtual-Try-on-with-Detail-Carving

Virtural try-on under arbitrary poses
MIT License
216 stars 64 forks source link

Bad Results on different pose and background. #20

Open ayush-angelium opened 3 years ago

ayush-angelium commented 3 years ago

Hello @wzhang34 @wyvernbai @L1aoXingyu @sandcu225 @Chen94yue @AIprogrammer Model gives bad results when we used different pose and background on custom images. Here is some issue that I faced :

  1. Source image : Different background with simple human pose Target Image : Woman cloth type(sleeveless top) Result Image : This cloth is looks like cloth over the t-shirt
  2. Source image : Different background with simple human pose Target Image : Woman cloth type(Jacket) Result Image : No doubt Jacket is fit on that girl but bad result on lower body cloth(it seems like pants)
  3. Source image : Different background with complex human pose Target Image : Woman cloth type(sleeveless top) Result Image : Model didn't give good result on pose keypoints, geometric matching module(gmm), semantic parsing and final result.
  4. Source image : Different background with complex human pose Target Image : Woman cloth type(sleeveless top) Result Image : Model didn't give good result on pose keypoints, geometric matching module(gmm), semantic parsing and final result.

Please give me a suggestion what should i do ? Screenshot from 2020-07-21 11-41-59

AIprogrammer commented 3 years ago

Hi, maybe the main reason of unreasonable results is the domain difference between the testing images and training dataset(MPV), where the background of person images are more simple(gray or white). Besides, some kinds of cloth types, such as the long women coat, jacket, are seldom appears in the training set we used. Perhaps recollecting more diverse data following the MPV structure, and retraining the model could hleps. For the complex background, processing the foreground(person) and background separately maybe a method if you want the generated image shares the same background with the source image.

ayush-angelium commented 3 years ago

@AIprogrammer when we run this model, our source faces is different or not clear on results as you can see in your original repo, so can you please let me know how do we set same faces. Thanks

AIprogrammer commented 3 years ago

If you test on the custom dataset, maybe retraining or fine tuning our model on your own dataset would be a choice.

spencerxhani commented 3 years ago

Hi @AIprogrammer @ayush-angelium , I encounter the similar problem but with simple T-shirt. the try-on clothes seem not fit on the person, do you guys have any suggestions about it ? thank you !

virtual try on pic