Abstract
With the recent development in ConvNet-based detectors, a successful solution for vessel detection becomes possible. However, it is essential to access a comprehensive annotated training set from different maritime environments. Creating such a dataset is expensive and time consuming. To automate this process, this paper proposes a novel self learning framework which automatically finetunes a generic pre-trained model to any new environment. With this, the framework enables automated labeling of new dataset types. The method first explores the video frames captured from a new target environment to generate the candidate vessel samples. Afterwards, it exploits a temporal filtering concept to verify the correctly generated candidates as new labels for learning, while removing the false positives. Finally, the system updates the vessel model using the provided self-learning dataset. Experimental results on our real-world evaluation dataset show that generalizing a finetuned Single Shot Detector to a new target domain using the proposed self-learning framework increases the average precision and the F1-score by 12% and 5%, respectively. Additionally, the proposed temporal filter reduced the noisy detections in a sensitive setting from 58% to only 5%.
Original language | English |
---|---|
Title of host publication | Representations, Analysis and Recognition of Shape and Motion from Imaging Data - 7th International Workshop, RFMI 2017, Revised Selected Papers |
Editors | Faouzi Ghorbel, Liming Chen, Boulbaba Ben Amor |
Place of Publication | Cham |
Publisher | Springer |
Pages | 121-135 |
Number of pages | 15 |
ISBN (Electronic) | 978-3-030-19816-9 |
ISBN (Print) | 978-3-030-19815-2 |
DOIs | |
Publication status | Published - 5 May 2019 |
Event | 7th International Workshop on Representations, Analysis and Recognition of Shape and Motion from Imaging Data, RFMI 2017 - Savoie, France Duration: 17 Dec 2017 → 20 Dec 2017 |
Publication series
Name | Communications in Computer and Information Science |
---|---|
Volume | 842 |
ISSN (Print) | 1865-0929 |
Conference
Conference | 7th International Workshop on Representations, Analysis and Recognition of Shape and Motion from Imaging Data, RFMI 2017 |
---|---|
Country/Territory | France |
City | Savoie |
Period | 17/12/17 → 20/12/17 |
Funding
Acknowledgement. This work is supported by the European ITEA APPS project. We thank the company Vinotion for providing the Botlek dataset to us. We also show our gratitude to the company NVIDIA for granting us a “TITAN X PASCAL” GPU.
Keywords
- Automated dataset creation
- ConvNet
- Convolutional networks (CNN)
- Maritime surveillance
- Self-learning
- Vessel detection