picture

Kevin Bascol

Research engineer at Bluecime

Contact

work

Bluecime S.A.S.,
445 Rue Lavoisier,
38330 Montbonnot-Saint-Martin, FRANCE

email

kevin(dot)bascol(at)bluecime(dot)com

computer

Github LinkedIn

News

In line with my thesis, from May 2020 I am fully integrated to Bluecime's team as a research engineer, pursuing my work to improve SIVAO and helping develop new products.


Thesis

Deep learning for anomaly detection in chairlifts videos


My PhD is under the supervision of Elisa Fromont and Rémi Emonet at Laboratoire Hubert Curien in Saint-Etienne (France).

Thanks to a CIFRE contract my PhD is in collaboration with Bluecime, a company based in Grenoble (France).
Bluecime proposes a system with a camera placed at the first tower of a chairlift and, by image processing techniques, an alarm is triggered if the passengers are potentially at risks (for example if they didn't close the restraining bar).

My thesis consists in using machine learning methods (mostly deep learning) to decide if the alarm should be triggered or not. Using machine learning could drastically reduce the time and resources required for the set-up of the current system.

Bluecime has designed a camera-based system to monitor the boarding station of chair-lifts in ski resorts, which aims at increasing the safety of all passengers. This already successful system does not use any machine learning component and requires an expensive configuration step. Machine learning is a subfield of artificial intelligence which deals with studying and designing algorithms that can learn and acquire knowledge from examples for a given task. Such a task could be classifying safe or unsafe situations on chairlifts from examples of images already labeled with these two categories,called the training examples. The machine learning algorithm learns a model able to predict one of these two categories on unseen cases. Since 2012, it has been shown that deep learning models are the best suited machine learning models to deal with image classification problems when many training data are available. In this context, this PhD thesis, funded by Bluecime, aims at improving both the cost and the effectiveness of Bluecime’s current system using deep learning. We first propose to formalize the Bluecime problem as a classification task with different training settings emulating use cases. We also propose a deep learning baseline providing competitive results in most of the settings,for a low configuration cost. We then propose different approaches to improve our baseline method.First, a data augmentation strategy to improve the robustness of our model. Then, two methods to better optimize the F-measure, a performance measure used in anomaly detection and better suited to evaluate our imbalanced problem than the usual accuracy measure. Finally, we propose selection strategies for the training data to improve results on newly installed chairlift for which no labeled training data is available. With this work we also show negative but interesting results on domain adaption in case of different imbalanced class distributions between the source and target domains.

@phdthesis{bascol2019phd,
  TITLE = {{Multi-source domain adaptation on imbalanced data: application to the improvement of chairlifts safety}},
  AUTHOR = {Bascol, Kevin},
  URL = {https://hal.archives-ouvertes.fr/tel-02417994},
  SCHOOL = {{Universit{\'e} jean Monnet}},
  YEAR = {2019},
  MONTH = Dec,
  KEYWORDS = {Machine leaning ; Deep learning ; Domain adaptation ; Imbalanced data},
  TYPE = {Theses},
  PDF = {https://hal.archives-ouvertes.fr/tel-02417994/file/Manuscrit_BASCOL.pdf},
  HAL_ID = {tel-02417994},
}
        
Manuscript pdf


Publications

International conferences:


Improving Domain Adaptation By Source Selection

Domain adaptation consists in learning from a source data distribution a model that will be used on a different target data distribution. The domain adaptation procedure is usually unsuccessful if the source domain is too different from the target one. In this paper, we study domain adaptation for image classification with deep learning in the context of multiple available source domains. We propose a multisource domain adaptation method that selects and weights the sources based on inter-domain distances. We provide encouraging results on both classical benchmarks and a new real world application with 21 domains.

@inproceedings{bascol2019source,
  TITLE = {{Improving Domain Adaptation By Source Selection}},
  AUTHOR = {Bascol, Kevin and Emonet, R{\'e}mi and Fromont, Elisa},
  BOOKTITLE = {{2019 IEEE International Conference on Image Processing (ICIP)}},
  YEAR = {2019},
}
Kevin Bascol, Rémi Emonet, Elisa Fromont - ICIP 2019 pdf

From Cost-Sensitive Classification to Tight F-measure Bounds

The F-measure is a classification performance measure, especially suited when dealing with imbalanced datasets, which provides a compromise between the precision and the recall of a classifier. As this measure is non convex and non linear, it is often indirectly optimized using cost-sensitive learning (that affects different costs to false positives and false negatives). In this article, we derive theoretical guarantees that give tight bounds on the best F-measure that can be obtained from cost-sensitive learning. We also give an original geometric interpretation of the bounds that serves as an inspiration for CONE, a new algorithm to optimize for the F-measure. Using 10 datasets exhibiting varied class imbalance, we illustrate that our bounds are much tighter than previous work and show that CONE learns models with either superior F-measures than existing methods or comparable but in fewer iterations.

@inproceedings{bascol2019fmeasure,
  TITLE = {{From Cost-Sensitive Classification to Tight F-measure Bounds}},
  AUTHOR = {Bascol, Kevin and Emonet, R{\'e}mi and Fromont, Elisa and Habrard, Amaury and METZLER, Guillaume and Sebban, Marc},
  BOOKTITLE = {{AISTATS 2019 - 22nd International Conference on Artificial Intelligence and Statistics}},
  ADDRESS = {Naha, Okinawa, Japan},
  SERIES = {The 22nd International Conference on Artificial Intelligence and Statistics},
  VOLUME = {89},
  NUMBER = {1},
  PAGES = {1245-1253},
  YEAR = {2019},
}
Kevin Bascol, Rémi Emonet, Elisa Fromont, Amaury Habrard, GuillaumeMetzler, Marc Sebban - AISTATS 2019 pdf

Improving Chairlift Security with Deep Learning

This paper shows how state-of-the-art deep learning methods can be combined to successfully tackle a new classification task related to chairlift security using visual information. In particular, we show that with an effective architecture and some domain adaptation components, we can learn an end-to-end model that could be deployed in ski resorts to improve the security of chairlift passengers. Our experiments show that our method gives better results than already deployed hand-tuned systems when using all the available data and very promising results on new unseen chairlifts.

@inproceedings{bascol2017chairlift,
  title={Improving Chairlift Security with Deep Learning},
  author={Bascol, Kevin and Emonet, R{\'e}mi and Fromont, Elisa and Debusschere, Raluca},
  booktitle={International Symposium on Intelligent Data Analysis (IDA 2017)},
  year = {2017}
}
Kevin Bascol, Rémi Emonet, Elisa Fromont, Raluca Debusschere - IDA 2017 pdf

Unsupervised Interpretable Pattern Discovery in Time Series Using Autoencoders

We study the use of feed-forward convolutional neural networks for the unsupervised problem of mining recurrent temporal patterns mixed in multivariate time series. Traditional convolutional autoencoders lack interpretability for two main reasons: the number of patterns corresponds to the manually-fixed number of convolution filters, and the patterns are often redundant and correlated. To recover clean patterns, we introduce different elements in the architecture, including an adaptive rectified linear unit function that improves patterns interpretability, and a group-lasso regularizer that helps automatically finding the relevant number of patterns. We illustrate the necessity of these elements on synthetic data and real data in the context of activity mining in videos.

@inproceedings{bascol2016unsupervised,
  title={Unsupervised Interpretable Pattern Discovery in Time Series Using Autoencoders},
  author={Bascol, Kevin and Emonet, R{\'e}mi and Fromont, Elisa and Odobez, Jean-Marc},
  booktitle={The joint IAPR International Workshops on Structural and Syntactic Pattern Recognition (SSPR 2016)},
  year={2016}
}
Kevin Bascol, Rémi Emonet, Elisa Fromont, Jean-Marc Odobez - SSPR 2016 pdf


French conference:


Un algorithme de pondération de la F-Mesure par pondération des erreurs de classfication.

Nous proposons un algorithme d’optimisation de la F-Mesure avec des garanties théoriques utilisable avec toute méthode d’apprentissage par pondération des erreurs. CONE, notre algorithme, génère itérativement un ensemble de coûts à partir de l’ensemble d’entraînement de telle sorte que le classifieur final ait une F-Mesure proche de l’optimale. L’optimalité de la F-Mesure obtenue est exprimée à l’aide d’une borne supérieure plus fine que celle présentée dans [Param-bath et al. 2014] De plus, nous montrons que les coûts obtenus à chaque itération de CONE permettent de réduire drastiquement l’espace de recherche et ainsi de converger rapidement vers les paramètres optimaux.L’efficacité de notre méthode est montrée à la fois en terme de F-Mesure mais aussi de vitesse de convergence sur plusieurs jeux de données déséquilibrés.

@inproceedings{bascol2019cone,
  TITLE = {{Un algorithme de pond{\'e}ration de la F-Mesure par pond{\'e}ration des erreurs de classfication.}},
  AUTHOR = {Bascol, Kevin and Emonet, R{\'e}mi and Fromont, Elisa and Habrard, Amaury and Metzler , Guillaume and Sebban, Marc},
  URL = {https://hal.archives-ouvertes.fr/hal-01803183},
  BOOKTITLE = {{Conf{\'e}rence pour l'Apprentissage Automatique}},
  ADDRESS = {Saint-Etienne du Rouvray, France},
  YEAR = {2018},
}
Kevin Bascol, Rémi Emonet, Elisa Fromont, Amaury Habrard, GuillaumeMetzler, Marc Sebban - CAp 2018