The technology increases the accuracy of state-of-the-art solutions for depth-from-images estimation based on conventional or machine learning. It can be applied to pre-trained algorithms, as well as to train a new algorithm to obtain improvements. It is also suited for non learning-based methods.
| Patent title | Depth determination method based on images, and relative system |
|---|---|
| Thematic area | Industry, Digital and Security |
| Ownership | ALMA MATER STUDIORUM - UNIVERSITA' DI BOLOGNA |
| Inventors | Fabio Tosi, Davide Pallotti, Stefano Mattoccia, Matteo Poggi |
| Protection | Italy, Europe, USA, China |
| Licensing status | Available for development agreements, option, license and other exploitation agreements |
| Keywords | Depth estimation, 3D reconstruction, Machine learning, Deep learning, Active sensors, Depth sensors fusion. |
| Filed on | 17 May 2019 |
State-of-the-art depth estimation systems from images mostly rely on stereo matching techniques. In particular, deep learning algorithms achieves the most accurate results. Previous works addressed the fusion between stereo technology and active sensors (for instance, ToF). Nevertheless, these techniques are not suited to modern deep learning models, that are the current state-of-the-art in this field.
The proposed technology acts on the abstract representation of the observed scene inside the deep learning algorithm as well as a conventional method. In particular, measurements obtained from active sensors (or non-learned stereo algorithms) are used to enhance features inside the deep learning algorithm in correspondence of values consistent with depth measured by the sensor or algorithm. Such formulation is flexible and adaptable to other fusion configurations, such as active sensor together with non-learned stereo algorithm.
The proposed technique greatly improve accuracy of state-of-the-art stereo algorithms based on deep learning as well as conventional methods. In particular, former methods perform poorly mainly when working in environments that are greatly different from those observed during the training procedure, while the proposed technique allows to partially remove this affect, vouching much more accurate results on challenging conditions caused by poor illuminations or bad weather conditions. Moreover, the methodology is also suited for conventional (i.e., non learning-based) methods.