The Augmented Autoencoder (AAE) pipeline estimates a 6D pose of an object purely trained on synthetic data. The resulting pose can be either directly used or further refined by a Iterative Closest Point (ICP) approach. Please note, while the AAE only requires a single RGB image the ICP approach runs on the corresponding depth image.
Install & Run: Full details of how to use the Augmented Autoencoder can be found in the Read Me section on GitHub.
Additional information: In order to estimate the pose of an object a bounding box detector first has to detect and classify the object. The resulting region of interest is then forwarded to the AAE, which predicts the 3D orientation of the corresponding object. To do so, the AAE generates a latent representation of the object appearance, which can be translated to a 3D orientation via a codebook. Based on this rotation information the translation can be computed via the projective distance.
No user data is used or any data saved. - lawful - respecting all applicable laws and regulations: The used software does not have any special regulations neither a dual-use issue. - robust: robustness is tested in several applications
The applied software does not require/use any user data. During training the algorithm, all images and labels are generated via a simulator. The inference phase requires real-world images of objects, however, these are processed on the fly and not saved.