Descriptellation: Deep Learned Constellation Descriptors
This work addresses global localization challenges for robotics or autonomous systems, offering an incremental improvement by replacing handcrafted topological descriptors with a learned approach.
The paper tackles the problem of global localization under viewpoint and appearance changes by introducing Descriptellation, a deep learned descriptor based on object constellations, which outperforms state-of-the-art and handcrafted methods on real-world datasets and shows robustness to noise.
Current descriptors for global localization often struggle under vast viewpoint or appearance changes. One possible improvement is the addition of topological information on semantic objects. However, handcrafted topological descriptors are hard to tune and not robust to environmental noise, drastic perspective changes, object occlusion or misdetections. To solve this problem, we formulate a learning-based approach by modelling semantically meaningful object constellations as graphs and using Deep Graph Convolution Networks to map a constellation to a descriptor. We demonstrate the effectiveness of our Deep Learned Constellation Descriptor (Descriptellation) on two real-world datasets. Although Descriptellation is trained on randomly generated simulation datasets, it shows good generalization abilities on real-world datasets. Descriptellation also outperforms state-of-the-art and handcrafted constellation descriptors for global localization, and is robust to different types of noise. The code is publicly available at https://github.com/ethz-asl/Descriptellation.