2 Scientific objectives

Learning approaches, particularly those employing deep learning models, heavily depend on data for
e cient training. In essence, a model's overall performance is intricately linked to both the quantity
and quality of annotated data accessible during the training phase. The laborious and costly task of
gathering extensive sets of annotated data for detection is often challenging. Obtaining such training data
becomes an insurmountable hurdle in some situations. Regulatory constraints, such as those prevalent
in the medical eld, frequently impose limitations on the use of personal data. Similarly, in the military
domain, the classi cation of data renders its disclosure unfeasible. This poses considerable problems
for training data-intensive deep learning methods. Fortunately, there are more data-e cient learning
strategies available, known as few-shot learning (FSL). There are various approaches for few-shot learning,
but they generally adhere to a common fundamental principle. They rst acquire general knowledge about
a related task (source task) before ne-tuning to suit a speci c target task. In the detection issue, this
strategy is known as "few-shot object detection". A large dataset containing object annotations in the
database is available. The source task is to detect these objects, while the target task is to detect objects
of new classes with a limited number of available annotations. However FSOD represents a signi cant
challenge, especially for deep learning-based approaches, which typically require a large number of training
examples and are prone to over tting in data-limited contexts.
Within the LabCom, extensive prior research has concentrated on aerial images, speci cally emphasizing
small object detection within this domain ([1]-[4]). A limited number of studies have explored aerial
images more broadly. The objective is to scrutinize and assess recently developed methods since 2022 (e.g. [5]-[8]), with a particular emphasis on their performance in the realm of aerial images, especially those sourced from the DOTA database. Subsequently, our aim is to devise targeted solutions to address the challenges identified. The research plan encompasses the following key steps:

1. Literature Review and Conceptual Understanding: Conduct an in-depth exploration of concepts outlined in the scienti c literature relevant to our specific problem, aligning with the industrial needs of COSE.

2. Performance Evaluation: Assess and compare the e ectiveness of various methodologies on both public datasets and those carefully curated by COSE.

3. Innovative Object Detection Approaches: Introduce novel approaches for detecting objects in aerial images, with meticulous consideration of constraints imposed by limited information.

4. Algorithm Re nement and Benchmarking: Extend and re ne the algorithms, concentrating exclusively on datasets provided by COSE to establish benchmarks for detection algorithms.

3 References

1. Pierre Le Jeune and Anissa Mokraoui, "Cross-Scale Query-Support Alignment Approach for Small Object Detection in the Few-Shot Regime", IEEE International Conference on Image Processing 2023 (ICIP).

2. Pierre Le Jeune and Anissa Mokraoui, "Extension de l'Intersection over Union pour am eliorer la d etection d'objets de petite taille en r egime d'apprentissage few-shot", GRETSI 2023, XXIX eme Colloque Francophone de Traitement du Signal et des Images, Grenoble, France.

3. Pierre Le Jeune and Anissa Mokraoui, "Improving Few-Shot Object Detection through a Performance Analysis on Aerial and Natural Images," 2022 30th European Signal Processing Conference (EUSIPCO), Belgrade, Serbia, 2022, pp. 513-517, doi: 10.23919/EUSIPCO55093.2022.9909878.

4. Pierre Le Jeune, Mustapha Lebbah, Anissa Mokraoui and Hanene Azzag, "Experience feedback using Representation Learning for Few-Shot Object Detection on Aerial Images," 2021 20th IEEE

International Conference on Machine Learning and Applications (ICMLA), Pasadena, CA, USA, 2021, pp. 662-667, doi: 10.1109/ICMLA52953.2021.00110.

5. Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, Shih-Fu Chang, Few-shot object detection with fully cross-transformer. CVPR 2022.

6. Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, Shih-Fu Chang, Meta faster r-cnn: Towardsaccurate few-shot object detection with attentive feature alignment. AAAI 2022.

7. Gongjie Zhang, Zhipeng Luo, Kaiwen Cui Shijian Lu, Meta-detr: Few-shot object detection via uni ed image-level meta-learning. TPAMI 2022.

4 Information to apply
Duration of the post-doctoral: 12 months
Expected Start Date: 1/03/2024
Contact: Anissa MOKRAOUI, Fangchen FENG
Email: labcom.iriser@univ-paris13.fr
Application le: CV, cover letter, at least one letter of recommendation (please indicate the post-doc
o er reference in the subject of your email).