arah-release
arah-release copied to clipboard
taconite
Segmentation Augmentation while Training???
Is there some augmentation feature in codes??? Because that looks so sharp. When I see logfiles, there are something strange about foreground images. If there is no problem with my datas, does it bother training?
No. There is no segmentation augmentation. Your case is most likely the result of a misalignment between the 2D bounding box (from which we sample camera rays) and the actual 3D region where the person resides. This can happen if you scale up/down the image, but do not change the camera intrinsics correctly. I believe the unmodified code should not produce such results. There are too many possible pitfalls on this issue so I cannot tell the cause with only the two pictures you posted. Can you specify your modifications?
I changed img_size with img_size=(512, 512) if not high_res else imgsize_data in ./im2mesh/config.py
imgsize_data is actual image size and I'm using original intrinsic parameter.
I'm using original easymocap data. But your camera.json isn't same with my camera.json.
I know your image resoultion is 512, so that means your focal length is half of mine. But It wasn't.
This is my focal length for view 01 01: "K":1417.43163872 . And I also used this data for neural body.
and yours : "K": [[1074.28138836669
I am not sure what "original easymocap data" means. Here are the raw camera 1 intrinsics from the official, unmodified ZJU-MoCap dataset, subject 377:
Besides, our code does not modify the original intrinsics until I call this line. I think your camera intrinsics is somehow wrong, how did you obtain them?
Closed due to inactivity. Feel free to reopen it if you have further update.