[question] Training a thin model and the test result is not correct

DLR-RM / AugmentedAutoencoder

Official Code: Implicit 3D Orientation Learning for 6D Object Detection from RGB Images

MIT License

338 stars 97 forks source link

[question] Training a thin model and the test result is not correct #32

Closed delnasay closed 5 years ago

delnasay commented 5 years ago

Hi!

I trained the AAE with a cad leaf model, which I set VERTEX_SCALE: 80, AUXILIARY_MASK: True, BOOTSTRAP_RATIO: 10. The training image is showed below:

But when I test in images with aae_image.py, the results are not correct. For example, this is one of the images I used: origin image

The predict image is pred image

It is obviously that the predict result is not correct. Could you give me some suggestion for this? Thanks a lot.

MartinSmeyer commented 5 years ago

Hi! So one problem is that the image gets resized to a square image and then looks quite distorted. This shape is approximated by a slight rotation of the original model.

You can use the fucntion extract_square_patch() from dataset.py to create an undistorted image crop from a tight bounding box. This should improve the result. Also you could map the texture on the model to improve results.

For this thin object, maybe even a segmentation is sufficient?