the full image results

[minushuang]

the outputs now are generated from the cropped and transformed images, which can not be attached to the original images directly. what should do if I want to the results aligned on the original images? replaced the cropped images with original ? it seems not a good idea. could you give me some instructions please. many thanks.

[chaneyddtt]

Hi @minushuang, I do not think that replace the cropped image with the original would be a good idea. It will makes the training harder because the scale and location will varies a lot if directly using the original image. May i know why you need to align with the original image?

[minushuang]

Hi @chaneyddtt , thanks for your prompt reply. for video inference scene, and more reasonable visualization results, I want to recovery the full image results from current 224x224 cropped results, with synth_rgb_refine on it.

I have digged into the code, and found that the input image of the model is not a simple bbox region of dog, it may have different scales and transforms, which makes the recovery processes some complicated.

is there any way that I can use to reach my scene?

[minushuang]

I will upload some figures tomorrow for better illustration and understanding.

[minushuang]

as the figure show below, they are original image, current model cropped result and the result I want(ignore the different synth_silhouette on it, and you can imagine a normal result with correct synth_rgb_refine output on it) respectively from left to right.

[minushuang]

i can resize the crop result and attach it to original image according to the coordinates returned from function mutils.crop. it's ok but not the best I want in my scenario. my goal is to generate special video effects on the dog with the output 3d mesh in a realtime, does this project can do this ?

[chaneyddtt]

Yeah, you can transform back to the original image space in this way, and currently our model have not achieved real time inference yet.