Our two papers entitled “Autoencoder Attractors for Uncertainty Estimation” and “Autoencoder for Synthetic to Real Generalization From Simple to More Complex Scenes” have been accepted for publication at IEEE International Conference on Pattern Recognition (ICPR 2022).
We propose a novel view on the recursive application of previously trained autoencoder models for uncertainty estimation. The latter can be interpreted as a dynamical system and combining it with MC dropout yields good uncertainty estimation on a wide range of datasets. Samples close to the training distribution should converge to the same attractor and samples far from the training distribution should converge to different attractors potentially of different classes.
We present a detailed analysis for the transferability from synthetic images to real images. We start off with investigations on visually simpler datasets and extend it to visually more complex scenes. We propose to use an autoencoder, a pre-trained feature extractor and the partially impossible reconstruction loss (PIRL) to improve the model performance on real images.