Avian Terrestrial Locomotion: Evolution, Dynamics, and Computer Vision
Bioinformatics and Theoretical Biology
Mechanics
Final Report Abstract
In the second funding period a main work point was the knowledge transfer between different Active Appearance Models trained on different animal locomotion sequences. For that a instance-weighted transfer learning method was developed. Furthermore, in the second funding period we focused datasets with cyclic locomotion which were not exploitable with methods of the first funding period caused by temporal disappearance of body parts and severe self-occlusions. With an extension of the existing probabilistic framework with a new prior knowledge in form of a tracking-by-detection approach we overcome these problems of occlusion. A new detection method was used for localizing body parts and the associated landmarks which needs only one single representative example of the object of interest. The resulting detection hypotheses afterwards are used in a new designed two-staged graph-based tracking algorithm. With this new local tracking prior we analyzed all datasets with temporal disappearance of body parts and severe self-occlusions. Another main focus of the second funding period was on the analysis of non-cyclic locomotion, in which birds had to overcome obstacles when running. It has been noticed here that the methods developed of the first funding period are not applicable to the short non-cyclic locomotion video sequences. Only a small number of frames inside the sequence contain all trackable landmarks, which is one fundamental prerequisite for the probabilistic framework. Additionally, this framework is adapted to the cyclicity of the steps. Therefore, in the second funding period a new automatic landmark tracking approach was proposed which can handle tracking landmarks in sequences where not the hole bird is in the scene. Hence, individual landmarks can be tracked as they enter the scene until they leave it. The independence of the anatomical knowledge is another advantage of the new tracking approach. For our project partners were all the advantages very useful facts for future work. Thus our partners can evaluate not only birds, but also other animal species. With the evaluation of 38 datasets with 36348 frames we have shown that our new tracking method outperforms standard methods and provides reasonable results for all landmark types.
Publications
- Comparative large-scale evaluation of human and active appearance model based tracking performance of anatomical landmarks in x-ray locomotion sequences. Pattern Recognition and Image Analysis. Advances in Mathematical Theory and Applications (PRIA), pages 86–92, 2014
Daniel Haase, John A. Nyakatura, and Joachim Denzler
(See online at https://doi.org/10.1134/S1054661814010222) - Instance-weighted transfer learning of active appearance models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1426–1433, 2014
Daniel Haase, Erik Rodner, and Joachim Denzler
(See online at https://doi.org/10.1109/CVPR.2014.185) - Robust pictorial structures for x-ray animal skeleton tracking. In International Conference on Computer Vision Theory and Applications (VISAPP), pages 351–359, 2014
Manuel Amthor, Daniel Haase, and Joachim Denzler
(See online at https://doi.org/10.5220/0004693403510359) - Mixed gaits in small avian terrestrial locomotion. Scientific Reports, 2015
Emanuel Andrada, Daniel Haase, Yefta Sutedja, John A. Nyakatura, Brandon M. Kilbourne, Joachim Denzler, Martin S. Fischer, and Reinhard Blickhan
(See online at https://doi.org/10.1038/srep13636) - Robust Data- and Model-Driven Anatomical Landmark Localization in Biomedical Scenarios. Verlag Dr. Hut, 2015. ISBN 9783843921954
D. Haase
- Anatomical landmark tracking by one-shot learned priors for augmented active appearance models. In International Conference on Computer Vision Theory and Applications (VISAPP), pages 246–254, 2017
Oliver Mothes and Joachim Denzler
(See online at https://doi.org/10.5220/0006133302460254)
