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.
Our paper entitled “Autoencoder Based Inter-Vehicle Generalization for In-Cabin Occupant Classification” has been accepted for publication at IEEE Intelligent Vehicles Symposium (IV 2021). We highlight the challenges for generalizing in between different vehicle interiors when a limited amount of variation is available during training (i.e. only a single vehicle and a few class instances). We propose an autoencoder based approach which performs better than classification models trained from scratch. Moreover, the autoencoder approach can transform images from unseen vehicles into the vehicle seen during training.
Our paper entitled “Illumination Normalization by Partially Impossible Encoder-Decoder Cost Function” has been accepted for publication at IEEE Winter Conference on Applications of Computer Vision (WACV 2021). We present an image normalization method based on a new strategy for the cost function formulation of encoder-decoder networks to average out all the unimportant information in the input images, in particular, to remove illumination changes and environmental features. Together with our proposed method, we release a synthetic dataset – an extension of SVIRO – of sceneries from three different passenger compartments where each scenery is rendered under ten different illumination and environmental.
Our paper entitled “SVIRO: Synthetic Vehicle Interior Rear Seat Occupancy Dataset and Benchmark” has been accepted for publication at IEEE Winter Conference on Applications of Computer Vision (WACV 2020). This paper gives a detailed overview about SVIRO, some baseline investigations to highlight the challenging conditions of our dataset and some results on real images to show that insights on SVIRO are transferable to real images. This launches the release of the dataset and start of the public benchmark.
We presented an extended abstract entitled “An Overview of the SVIRO Dataset and Benchmark” at the ACM Computer Science in Cars Symposium (CSCS 2019) held at the DFKI in Kaiserslautern in October. This work gives a short introduction about our dataset and the problem statement we want to investigate. The full paper is accepted at IEEE WACV 2020.