Seeing Trouble: Detecting Computer Vision Failures Across Domains

Overview

Computer vision systems are increasingly used in real-world applications such as robotics, infrastructure inspection, environmental monitoring, healthcare, transport, and agriculture. However, these systems can make errors when conditions change, for example due to unusual lighting, weather, camera viewpoint, object appearance, image quality, or unfamiliar environments.

This project will investigate how to identify when a computer vision model is likely to be wrong, without relying on strong assumptions about the data or requiring very large labelled datasets to set decision thresholds. The project is broadly connected to reliable artificial intelligence and trustworthy computer vision across different application domains.

You will work with existing computer vision models and datasets, or a project-specific application dataset, to evaluate methods for detecting errors using model confidence. The project will explore how well different confidence or consistency-based signals can indicate when computer vision model predictions should be trusted, flagged for review, or treated with caution.

Research engagement

You will engage with applied research in reliable computer vision. This may include a focused literature review on uncertainty estimation, error detection, and robustness of these techniques under domain shift; hands-on experimentation with computer vision models; and evaluation of methods that aim to detect unreliable predictions with limited labelled data.

The project can be adapted to different computer vision tasks, such as image classification, object detection, segmentation, or others, depending on your background, interest areas and the available project data.

Research activities

You will work with the supervisor and, where appropriate, other researchers working on computer vision applications. Activities may include:

• Running existing computer vision models on one or more datasets from different domains.
• Implementing and testing methods for identifying unreliable predictions, such as confidence scores, prediction consistency, uncertainty estimates, or simple calibration approaches.
• Comparing methods using metrics such as error detection accuracy, precision, recall, data and compute requirements.
• Preparing a short final report and demonstration of the findings.

Research skills

You will gain experience working with state-of-the-art computer vision models, Python-based machine learning workflows, and working with real image datasets. You will also gain skills in applied computer vision research, experimental evaluation,  and communicating research findings clearly.

Outcomes

Expected outcomes include an experimental evaluation of two or more error-detection approaches across different domains, an analysis of common failure modes, and recommendations for designing more reliable computer vision systems across domains.

Skills and experience

You should have some programming experience, preferably in Python.  Some familiarity with machine learning or computer vision would be helpful, including tools like Pytorch, but the project can be scoped to match your background. The project is suitable for students interested in artificial intelligence, computer vision, robotics, trustworthy AI, or real-world machine learning.

Start date

2 November, 2026

End date

19 February, 2027

Location

QUT Gardens Point or Online

Additional information

You will receive regular supervision and guidance throughout the project. Depending on the final project direction, you may have access to example code and desktop machines that can run large computer vision models.

Keywords

• computer vision
• reliable AI
• trustworthy AI
• image classification
• object detection
• machine learning
• artificial intelligence

Contact

Dimity Miller

N/A

d24.miller@qut.edu.au