hi , I have a question about the result

[bsdj2]

   I have run the demo succesfully and I also find a problem about it. 
   when I run the code on the images in h36m-sample folder, it works well. But when I run it on the other images , the result would become very strange and completely wrong. Before run the demo, I also give a correct boundingbox of the test images.
  Have you found this problem? 

[geopavlakos]

I reproduced the procedure for downloading the test images and running the model on them, and I could indeed get the expected results on the full test set of h36m as well. Have you followed the README instructions correctly? If so, then you should be able to use the file 'demoH36M.m' to visualize the predictions (after updating the variables 'datapath' and 'predpath' to point to the corresponding folders of the complete h36m test data).

[bsdj2]

Thanks for your reply. Sorry, maybe my description of the problem is not clear. I actually want to ask is how can I use it get a better performance on the in-the-wild images?Do you have some advice? I have followed the README insructions correctly and got a good result on the test set of h36m. But when I test on in-the-wild images which I also give a BBox input, its performance can not be satisfactory. For example, for a same outdoor-scene image, if I put the background of the body painted white, the predicted result will be better than the result test on original image.