Multi-modal Sensor Fusion and Semi-automated Labelling Systems
PhD research on Data-Centric Learning and Dataset Quality for Autonomous Driving Perception
My PhD research focused on applied deep learning for multi-modal sensor data, particularly in the context of automated driving. I developed methods for sensor fusion and semi-automated data labelling, bridging machine learning research with scalable, production-relevant systems. My work emphasized efficiency and practicality in handling large, complex datasets.
Research Focus
My research focuses on a central challenge in modern machine learning and autonomous driving: how to develop reliable perception systems when training data is imperfect, incomplete, or costly to obtain. Across my work, I have investigated the interaction between dataset quality, labeling processes, and deep-learning-based perception, with the goal of improving model robustness while reducing dependence on expensive manual annotation. This research sits at the intersection of computer vision, machine learning, data-centric AI, and autonomous driving perception.
A key motivation is that state-of-the-art perception systems rely heavily on large-scale labeled datasets, yet the generation and maintenance of high-quality annotations remains one of the most significant bottlenecks in industrial AI development. While advances in neural network architectures continue to improve performance, practical deployment often depends equally on the quality, consistency, and scalability of the underlying data. This observation motivated a series of contributions addressing dataset quality from complementary perspectives.
The foundation of this research is the broader field of multi-modal perception for autonomous driving, where information from cameras, LiDAR, radar, and other sensors must be fused to enable robust scene understanding. In the paper Deep Multi-modal Object Detection and Semantic Segmentation for Autonomous Driving: Datasets, Methods, and Challenges, we analyzed the emerging landscape of multi-sensor perception systems, reviewed major benchmarks and datasets, and examined the advantages and limitations of different fusion strategies. The work highlighted that the success of autonomous driving perception is not only determined by model architecture but also by sensor quality, annotation quality, and dataset coverage, thereby motivating a stronger focus on data-centric approaches to perception development.
Building on this perspective, the paper Estimating Labeling Quality with Deep Object Detectors introduced methods for using trained object detectors themselves as tools for dataset auditing and quality assessment. Instead of treating annotations as unquestionable ground truth, we demonstrated that detector predictions can be leveraged to estimate annotation quality, identify missing labels, and detect inconsistencies in large-scale datasets. This work established a feedback loop between model development and data validation, enabling more efficient quality assurance processes for industrial perception datasets. The approach reduces the need for exhaustive manual inspection and supports scalable dataset management in safety-critical applications such as autonomous driving.
The next step was to move beyond quality estimation towards active dataset improvement. In the paper Iterative Label Improvement: Robust Training by Confidence Based Filtering and Dataset Partitioning, we proposed a data-centric learning framework that automatically improves training labels through iterative self-training and confidence-based relabeling. The method combines pseudo-label generation, confidence estimation, and dataset partitioning to progressively increase label quality while simultaneously improving model performance. Unlike many competing approaches, the framework does not require additional teacher models or perfectly curated reference datasets, making it highly applicable in real-world industrial settings. Furthermore, the approach naturally extends to semi-supervised learning scenarios by exploiting large amounts of unlabeled data.
Taken together, these contributions form a coherent research agenda centered on data quality as a fundamental driver of machine learning performance. Rather than viewing datasets as static inputs to model development, the research treats datasets as evolving assets that can be analyzed, validated, and continuously improved using machine learning itself. This perspective anticipates the principles of what is now widely known as Data-Centric AI, where improvements in data quality often yield greater practical benefits than increases in model complexity.
The broader impact of this work lies in enabling more scalable and reliable AI systems for autonomous driving and other safety-critical domains. By combining expertise in perception systems, dataset engineering, label quality estimation, and robust learning under noisy supervision, the research contributes toward reducing annotation costs, improving model reliability, and accelerating the deployment of machine learning systems in real-world environments.