Master thesis: Semantic Communications with Foundation Models for Vehicle and Road User Detection

Bandwidth-limited devices in future wireless networks could save spectrum by sending information relevant to a task rather than complete images. This thesis at Ericsson explores how compression and wireless-channel noise affect frozen vision foundation-model features used to detect vehicles and other road users. The aim is to assess whether these representations can support reliable object detection within a semantic-communication system.

The assignment

• Review research on semantic communication, DeepJSCC, vision foundation models and JPEG compression.
• Extract spatial patch features using frozen CLIP-ViT-B/16 and DINOv2-ViT-B/14 models for six COCO categories: cars, trucks, buses, motorcycles, people and bicycles.
• Develop a pipeline with a learned linear compression bottleneck, quantization, a simulated additive white Gaussian noise channel and a lightweight anchor-free decoder for bounding boxes and class labels.
• Train the bottleneck and detection decoder jointly, introducing noise at randomly selected signal-to-noise ratios.
• Evaluate both foundation-model approaches against a task-specific encoder-decoder and conventional alternatives such as JPEG.
• Assess detection quality in relation to channel conditions, including mean average precision versus signal-to-noise ratio and results for individual classes.

Who we are looking for

• Master’s student in electrical engineering, computer science, telecommunications or a closely related discipline.
• Strong Python and PyTorch programming skills.
• Solid knowledge of deep learning, including convolutional networks, transformers, training methods and loss functions.
• Able to plan experiments, work independently and critically assess findings.
• Fluent in spoken and written English.
• Experience with object detection, wireless communication concepts or pretrained vision models is beneficial.

What we offer

• Supervision from Cristian Vaca Rubio and Maxim Teslenko.
• An opportunity to explore emerging technology alongside a team focused on innovation and challenging technical problems.

Practical details

Location: Stockholm. The project is intended for one student. Contact: Aneta Vulgarakis. Start date, project duration and application deadline are not specified.