This individual project focuses on image segmentation using a subset of the ADE20K dataset. The assessment requires students to investigate viable image segmentation models, analyse and visualise the provided data, develop appropriate data preprocessing procedures, train a segmentation model and critically evaluate the resulting performance. The assignment is worth 25% of the module assessment and is designed to assess students’ ability to apply research methods to a data science problem, communicate analytical findings effectively and select appropriate methods while understanding their advantages and limitations. Students are required to work with the provided ADE20K dataset and produce a model capable of semantic segmentation for four specified object classes: person, car, book and airplane. The model must identify which regions of an unlabelled image correspond to these four classes. The assignment specifically requires semantic segmentation rather than simply identifying whether an object is present. Treating the task as binary segmentation, where all objects are considered a single foreground class, will result in a significant penalty. The project requires students to investigate different approaches and models, perform exploratory data analysis, develop suitable preprocessing procedures and customise their selected model or models to improve performance. Students may use more than one model, although at least one model must be partially or fully trained by the student. Where multiple models are used, comparison with pretrained models is encouraged. Potential approaches include segmentation architectures such as U-Net and Mask R-CNN, with appropriate model selection justified through relevant literature and experimental evidence. The report should contain at least four core sections: an Introduction incorporating a literature review, Data Description and Exploratory Data Analysis, Methodology, and Results and Discussion. The literature review should cite at least three relevant research papers. The results section should include evaluation of three test images using the developed model and comparison with relevant published literature. Students are expected to provide a critical analysis of their work rather than simply reporting numerical results, explaining the reasons for observed outcomes and considering how modelling and implementation choices affected performance. The assessment also requires a documented Google Colab notebook containing the implemented steps used to train and evaluate the model. The notebook should be accessible to markers and should demonstrate the preprocessing, model development, training and evaluation process. The report must be between 700 and 1,000 words, excluding references and code, and should include code in text format as an appendix rather than screenshots. Figures and tables should have clear captions explaining what they present and their source. Assessment criteria include literature review and context, dataset description and exploratory data analysis, implementation quality, quality of analysis and critical discussion, and report quality and structure. Strong work is expected to demonstrate appropriate model selection, effective preprocessing and augmentation, suitable evaluation metrics such as Intersection over Union (IoU) and Dice score, meaningful visualisation and a critical interpretation of results and limitations.
Image Segmentation · Semantic Segmentation · ADE20K Dataset · Data Science · Exploratory Data Analysis · EDA · Data Pre-processing · Data Visualisation · Machine Learning · Deep Learning · Segmentation Models · U-Net
Megaminds has supported academic requirements in data science, data science and related disciplines.