Nick Theisen

Nick Theisen, Robin Bartsch, Dietrich Paulus et al.

Semantic segmentation is an essential step for many vision applications in order to understand a scene and the objects within. Recent progress in hyperspectral imaging technology enables the application in driving scenarios and the hope is that the devices perceptive abilities provide an advantage over RGB-cameras. Even though some datasets exist, there is no standard benchmark available to systematically measure progress on this task and evaluate the benefit of hyperspectral data. In this paper, we work towards closing this gap by providing the HyperSpectral Semantic Segmentation benchmark (HS3-Bench). It combines annotated hyperspectral images from three driving scenario datasets and provides standardized metrics, implementations, and evaluation protocols. We use the benchmark to derive two strong baseline models that surpass the previous state-of-the-art performances with and without pre-training on the individual datasets. Further, our results indicate that the existing learning-based methods benefit more from leveraging additional RGB training data than from leveraging the additional hyperspectral channels. This poses important questions for future research on hyperspectral imaging for semantic segmentation in driving scenarios. Code to run the benchmark and the strong baseline approaches are available under https://github.com/nickstheisen/hyperseg.

Nick Theisen, Peer Neubert

Hyperspectral imaging (HSI) semantic segmentation typically relies on in-domain training, but limited data availability often restricts model performance in real-world applications. Current approaches to leverage foundation models in proximal sensing use cross-modality techniques, bridging RGB and HSI to exploit vision foundation models. However, these methods either discard spectral information or introduce architectural complexity. We propose cross-domain transfer as an alternative, reusing HSI foundation models - originally trained in remote sensing - for proximal sensing applications. By eliminating the need to bridge modality gaps, our approach preserves spectral information while maintaining a simple architecture. Using the HS3-Bench benchmark, we systematically evaluate and compare conventional in-domain, in-modality training, cross-modality transfer and cross-domain transfer strategies. Our results demonstrate that cross-domain transfer achieves large performance improvements over in-domain, in-modality training, reduces the performance gap to cross-modality approaches and maintains strong performance in limited data settings. Thus, this work advances more effective HSI semantic segmentation in diverse applications.

Nick Theisen, Kenny Schlegel, Dietrich Paulus et al.

The classification of pixel spectra of hyperspectral images, i.e. spectral classification, is used in many fields ranging from agricultural, over medical to remote sensing applications and is currently also expanding to areas such as autonomous driving. Even though for full hyperspectral images the best-performing methods exploit spatial-spectral information, performing classification solely on spectral information has its own advantages, e.g. smaller model size and thus less data required for training. Moreover, spectral information is complementary to spatial information and improvements on either part can be used to improve spatial-spectral approaches in the future. Recently, 1D-Justo-LiuNet was proposed as a particularly efficient model with very few parameters, which currently defines the state of the art in spectral classification. However, we show that with limited training data the model performance deteriorates. Therefore, we investigate MiniROCKET and HDC-MiniROCKET for spectral classification to mitigate that problem. The model extracts well-engineered features without trainable parameters in the feature extraction part and is therefore less vulnerable to limited training data. We show that even though MiniROCKET has more parameters it outperforms 1D-Justo-LiuNet in limited data scenarios and is mostly on par with it in the general case

Raphael Memmesheimer, Simon Häring, Nick Theisen et al.

One-shot action recognition allows the recognition of human-performed actions with only a single training example. This can influence human-robot-interaction positively by enabling the robot to react to previously unseen behaviour. We formulate the one-shot action recognition problem as a deep metric learning problem and propose a novel image-based skeleton representation that performs well in a metric learning setting. Therefore, we train a model that projects the image representations into an embedding space. In embedding space the similar actions have a low euclidean distance while dissimilar actions have a higher distance. The one-shot action recognition problem becomes a nearest-neighbor search in a set of activity reference samples. We evaluate the performance of our proposed representation against a variety of other skeleton-based image representations. In addition, we present an ablation study that shows the influence of different embedding vector sizes, losses and augmentation. Our approach lifts the state-of-the-art by 3.3% for the one-shot action recognition protocol on the NTU RGB+D 120 dataset under a comparable training setup. With additional augmentation our result improved over 7.7%.

Raphael Memmesheimer, Nick Theisen, Dietrich Paulus

Recognizing an activity with a single reference sample using metric learning approaches is a promising research field. The majority of few-shot methods focus on object recognition or face-identification. We propose a metric learning approach to reduce the action recognition problem to a nearest neighbor search in embedding space. We encode signals into images and extract features using a deep residual CNN. Using triplet loss, we learn a feature embedding. The resulting encoder transforms features into an embedding space in which closer distances encode similar actions while higher distances encode different actions. Our approach is based on a signal level formulation and remains flexible across a variety of modalities. It further outperforms the baseline on the large scale NTU RGB+D 120 dataset for the One-Shot action recognition protocol by 5.6%. With just 60% of the training data, our approach still outperforms the baseline approach by 3.7%. With 40% of the training data, our approach performs comparably well to the second follow up. Further, we show that our approach generalizes well in experiments on the UTD-MHAD dataset for inertial, skeleton and fused data and the Simitate dataset for motion capturing data. Furthermore, our inter-joint and inter-sensor experiments suggest good capabilities on previously unseen setups.

Raphael Memmesheimer, Nick Theisen, Dietrich Paulus

We present a simple, yet effective and flexible method for action recognition supporting multiple sensor modalities. Multivariate signal sequences are encoded in an image and are then classified using a recently proposed EfficientNet CNN architecture. Our focus was to find an approach that generalizes well across different sensor modalities without specific adaptions while still achieving good results. We apply our method to 4 action recognition datasets containing skeleton sequences, inertial and motion capturing measurements as well as \wifi fingerprints that range up to 120 action classes. Our method defines the current best CNN-based approach on the NTU RGB+D 120 dataset, lifts the state of the art on the ARIL Wi-Fi dataset by +6.78%, improves the UTD-MHAD inertial baseline by +14.4%, the UTD-MHAD skeleton baseline by 1.13% and achieves 96.11% on the Simitate motion capturing data (80/20 split). We further demonstrate experiments on both, modality fusion on a signal level and signal reduction to prevent the representation from overloading.

Raphael Memmesheimer, Ivanna Mykhalchyshyna, Viktor Seib et al.

In this paper we present an approach for learning to imitate human behavior on a semantic level by markerless visual observation. We analyze a set of spatial constraints on human pose data extracted using convolutional pose machines and object informations extracted from 2D image sequences. A scene analysis, based on an ontology of objects and affordances, is combined with continuous human pose estimation and spatial object relations. Using a set of constraints we associate the observed human actions with a set of executable robot commands. We demonstrate our approach in a kitchen task, where the robot learns to prepare a meal.