Mots-Clés

Explaining object detection models

Description

 

Duration: Flexible Mar/Apr – Aug/Sep

Context

 

In a biological context, it is useful to understand the internal workings of deep learning models and to identify the most essential features or reasoning for the classification or regression tasks. Fortunately, many methods have been developed in the last decade to tackle the problem of the explainability of DL models, such as feature relevance, local or global explanations, and visualizations. However, these methods cannot be directly translated to the object detection tasks given many technical difficulties:

 

• No transformation of input to the output via the usual gradient

• Dependability on the specific characteristics (anchor based or free)

• Dependability on the architectures

• Computationally challenging

• Explaining category as well as location of the objects

 

Objectives

 

The objectives of this internship would be:

• perform literature review of the existing methods for explaining object detection models

• compare different methods

• identify pitfalls

• apply it to the yolov8 and RetinaNet object detector developed in the team

• highlight the advantages and disadvantages of these approach on these detectors

• propose solutions to resolve different technical problems

 

Qualifications

 

• Ability to familiarize with the code quickly

• Python, C++

• Machine learning

• OS: Linux

• Latex

• Ability to work as independent as well as a part of team

• Creative and good communication skills

 

Application procedure:

Please send your application as one pdf including CV and your grades to misbah.razzaq@inrae.fr

 

Thank you for your interest.