Description
Deep neural networks have become the de facto approach for computer vision in a large number of applications due to the exceptional performance they manifest “in the lab.” The widespread adoption of these methods has also benefited the agricultural robotics sector by providing it with high quality visual models that constitute the perceptive component of the robots [1]. However, the deployment of these models in the real world comes with a large spectrum of challenges [2, 3].
One such important challenge is that of domain shift [4]. This problem arises when the source data, on which the neural network has been trained, follows a different distribution than that of the target data, on which the model is to be deployed. Domain shift often leads to an important degradation in the predictive capacity of the neural network, sometimes rendering it unusable. One or many of a large collection of factors may be behind this domain shift. For example, in the agricultural domain, it can be due to changes in the environmental conditions (sunny vs. cloudy weather), changes in the morphological aspect of the observed species (at different growth stages, for example), or even differences in sensor characteristics due to a change of camera, for instance [5].
A large number of domain adaptation methods exist in the literature [5, 6, 7, 8]. This internship aims at exploring different domain adaptation methods for adapting neural networks trained on images captured using a source camera to be deployed on images captured using a target camera. Here, the source and the target present different optical and color characteristics. The internship work will consist of:
- Conducting a bibliographic review of current state-of-the-art domain adaptation methods,
- Implementing and adapting these methods,
- Deploying and validating these methods on real-world agricultural images.
If you are interested in conducting experimental research in deep learning for applications with high impact in agriculture and the environment, this is the right internship for you.
Candidate Profile
- Final-year Master’s or Ecole d’ingénieurs student in machine learning, statistics, image processing, computer science or related fields ;
- Strong interest in applied research in deep learning, computer vision and robotics. Interest and motivation for agriculture and the environment are a strong plus.
- Good to exceptional programming skills in Python, and at least some knowledge and experience with the main deep learning libraries (PyTorch, TensorFlow, Keras…).
- Motivation to spend six months in one of the most beautiful cities in France: Bordeaux, with potential visits to the great city of Lyon.
Internship Organisation
This internship is financed by EXXACT Robotics and hosted at the IMS Laboratory in Bordeaux. The intern will be supervised by Jean-Pierre Da Costa (IMS), Lionel Bombrun (IMS) and Paul Melki (EXXACT Robotics).
The internship duration is 6 months starting February/March 2026 with a monthly gross salary of 1300 euros.
If you’re thrilled to join the team, please apply here directly.
References
[1] Kamilaris, A. & Prenafeta-Boldù, F. X. (2018). Deep Learning in Agriculture: A Survey, Computers and Electronics in Agriculture, 147. DOI: 10.1016/j.compag.2018.02.016.
[2] Paleyes, A., Urma, R.-G. & Lawrence, N. D. (2023). Challenges in Deploying Machine Learning: A Survey of Case Studies, ACM Computing Surveys, 55, DOI: 10.1145/ 3533378
[3] Hendrycks, D. et al. (2022). Unsolved Problems in ML Safety, arXiv:2109.13916.
[4] Zhou, K. et al. (2023). Domain Generalization: A Survey, IEEE Transactions on Pattern Analysis and Machine Intelligence, 45, DOI: 10.1109/TPAMI.2022.3195549.
[5] Hu, X., Chen, S. & Zhang, D. (2025). Domain Adaptation in Agricultural Image Analysis: A Comprehensive Review from Shallow Models to Deep Learning, arXiv:2506.05972.
[6] Zhao, S. et al. (2024). More is Better: Deep Domain Adaptation with Multiple Sources, International Joint Conference on Artificial Intelligence (IJCAI), 9, DOI: 10.24963/ijcai.2024/923.
[7] Yang, Y. & Soatto, S. (2020). FDA: Fourier Domain Adaptation for Semantic Segmentation, Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Virtual.
[8] Magistri, et al. (2023). From One Field to Another – Unsupervised Domain Adaptation for Semantic Segmentation in Agricultural Robotics, Computers and Electronics in Agriculture, 212, DOI: 10.1016/j.compag.2023.108114.
