Congratulations to Hong-Hanh Nguyen-Le (UCD), Cohort 5 ML-Labs student who won the top prize at the UCC AI Quest 2023 last month along with her team members and fellow PhD students from Trinity College and DCU. The competition is organised every year by UCC in conjunction with Insight, CRT AI and Cadence. We asked Hanh to tell us more about the challenge they had to solve and their solution: “this challenge explores the use of semantic segmentation for vegetation recognition in aerial images. By deeply utilizing exploratory data analysis techniques, we first identified three main challenges in the dataset provided by the competition: limited data, coarse mask annotations, and long-tailed sample difficulty distribution. To address them, we propose a hybrid approach with both data-centric inspired improvements and model-centric improvement. Specifically, we applied color augmentation techniques to encourage models to learn shape features instead of solely relying on color in such samples. Additionally, a patch-based learning strategy is implemented, and predictions are upscaled to enhance alignment with coarse labels. Regarding model-centric improvements, hierarchical stacking of diverse model families, and a novel Repeated Warmup Learning Rate (RWLR) strategy are introduced to improve performance and stabilize training.”

Another ML-Labs Cohort 5 team composed of Oluwuabukola Adegboro (DCU), Nika Gorchakova (DCU), Abigail Naa Amankwaa Abeo (DCU), Vaaibhavi Singh (UCD) and Oluwadara Adedeji (UCD) was in the top 10 finalists. Well done to all involved!