search

DLR-RM / AugmentedAutoencoder

Official Code: Implicit 3D Orientation Learning for 6D Object Detection from RGB Images
MIT License
338 stars 97 forks source link

Evaluation on Custom dataset and occasional fail #47

Closed kuwt closed 4 years ago

kuwt commented 4 years ago

Thanks for the code for evaluation. I have trained it on a custom object: image

I tested it on 3 samples(first row) and the result is (second row): image

The middle object has an incorrect output. What is your idea on this issue?

kuwt commented 4 years ago

I changed the crop image aspect ratio. The result is better. image

MartinSmeyer commented 4 years ago

Yes, I assume you used the aae_image.py script to generate these results? In contrast to the other test scripts, it does not pad the input images before inserting them into the AAE but simply resizes them to a quadratic fixed size. This introduces strong distortions if your original input is not quadratic and therefore the pose estimation fails. So simply padding with black pixels up to a quadratic aspect ratio should prevent that. I will add the padding also to aae_image.py.

kshpv commented 4 years ago

@MartinSmeyer

Also, I have not good results using aae_image.py

Where could it be a problem? Do you have any suggestions?

1_resized 1_res 2_resized 2_res 3_resized 3_res 4_resized 4_res 5_resized 5_res 6_resized 6_res

kshpv commented 4 years ago

Here are my training figures

training_images_9999 training_images_19999 training_images_29999

.