Traditional computer vision and multimodal LLMs for cloud detection to support navigation

Publication date : Oct 05, 2026, 5:13PM

Cloud detection and characterization are crucial for aiding air navigation. Cloud cover, the location of observed formations, or their structure can provide valuable information.

This internship is part of the Climate Impulse project, which aims to achieve the first non-stop, zero-emission around-the-world flight. The goal is to explore automatic image analysis methods to detect and describe cloud conditions observed from an aircraft.

Classical computer vision approaches, such as detection or classification models like YOLO, allow for quick identification of objects in an image. However, they require annotated data.

Meanwhile, multimodal language models can interpret an image and generate a contextualized description. They could serve as a complementary approach to locate clouds within the image or provide a global understanding of the scene.

The internship's objective is to compare these two approaches, evaluate their performance and limitations, and study their complementarity for scene analysis in navigation assistance.

Work to be done:

1) State of the Art and Data
Conduct a bibliographic review on cloud detection, classification, and characterization via computer vision.
Identify, select, or enrich datasets of sky images, aerial or onboard. Special attention will be given to the possibility of creating a synthetic dataset tailored to the use cases studied.
2) Traditional Computer Vision Approach
Implement one or more reference models like YOLO in Python/PyTorch.
Train and evaluate these models for cloud formation detection and/or classification.
Measure performance, robustness under varied visual conditions, and inference time.
3) Exploration of Multimodal LLMs
Test several multimodal models via API and/or locally for open-weight models.
Design prompts to obtain a structured scene description: presence of clouds, estimated coverage, location, observable features, and uncertainty level.
Assess the relevance, reliability, latency, and limitations of these models.
4) Comparison and Demonstrator
Compare traditional and multimodal approaches based on accuracy, robustness, explainability, cost, and computation time.
Study an hybrid approach where a detector localizes cloud zones and a multimodal LLM provides interpretation or scene summarization.
Develop a demonstrator visualizing detection results and generated descriptions.

• Final-year engineering or master's student in artificial intelligence.
• Interest in Deep Learning, Computer Vision.
• Curiosity for multimodal LLMs and generative AI approaches.
• Rigor in experimentation.
• Good Python skills.

The Innovation Division's ambition is to push the boundaries of Orange's innovation and strengthen its technological leadership by leveraging our research capabilities to foster responsible innovation serving humanity, inform the Group's long-term strategic choices, and influence the global digital ecosystem.

We train experts in today's and tomorrow's technologies and ensure continuous improvement in the performance of our services and efficiency. The Innovation Division brings together, worldwide, 6,000 employees dedicated to research and innovation, including 740 researchers. With a global vision and a diverse range of profiles (researchers, engineers, designers, developers, data scientists, sociologists, graphic designers, marketers, cybersecurity experts…), the women and men of Innovation listen and serve the countries, regions, and business units to make Orange a trusted multiservice operator.

Within Orange Innovation, you will be part of a research team with dual expertise in artificial intelligence and signal processing. You will work alongside researchers engaged in innovative and foundational topics shaping the future of telecommunications, particularly for 6G.

Desired start date : Mar 01, 2027, 12:00AM

At Orange, only your skills matter.

Regardless of your age, gender, background, origin, religion, sexual orientation, disability, neurodiversity, or appearance, we actively encourage diversity within our teams, as it is a collective strength and a driver of innovation.
Orange is a disability-inclusive employer: please feel free to let us know about any specific needs you may have.