Internship subject at LIFO, Orléans

22 December 2023

---

Catégorie : Stagiaire

---

Few-shot object detection for robot manipulation

Location : LIFO Laboratory, University of Orléans, France

Supervisors : Rim Rahali, Trung Anh Dang, Vincent Nguyen

Duration : 6 mois

Allowances : Legal internship allowances (around €600 per month)

Context:

With deep learning, recent advancements in computer vision hold great promise for the field of robotics. However, the practical application of these breakthroughs is not straightforward, particularly in adapting to new tasks that demand substantial annotated data, and memory, power and time for the re-training.

Few-Shot Object Detection (FSOD), an emerging approach, helps detectors adapt to unseen classes from limited object instances of novel categories. This internship aims to empower the FSOD approach and address specific challenges in a robotics context marked by limited data and annotations, dynamic environments, and the need for efficient fine-tuning.

• Misclassifying novel instances
• Losing knowledge of the models on previously learned objects
• The problem of scale variations

Objectives and missions :

We aim at developing strategies to prevent misclassifying novel instances as similar classes, while counteract knowledge loss regarding previously learned objects. We can also study the methods for tackling domain shift challenges in dynamic environments and handling a common scenario in real-world application: scale variations in objects.

The internship consists of 2 parts: 1) Analyze recent methods for FSOD. 2) Implement and improve a selected method.

Required skills :

• Deep learning, programming.
• Computer science, machine learning or applied mathematics profile.
• Ability to write reports for a regular, clear and effective restitution of the work carried out;

Application:

Application to be sent to vincent.nguyen@univ-orleans.frwith the subject “FSOD Internship Application” accompanied with a CV and a cover letter, as well as grades from previous academic years.