Reproduce “Multitask Learning for Large-scale Semantic Change Detection” on OpenResearch

About this paper

Rodrigo Caye Daudt, Bertrand Le Saux, Alexandre Boulch, Yann Gousseau

Researchers from ONERA and Télécom ParisTech introduced the High Resolution Semantic Change Detection (HRSCD) dataset, the first large-scale, pixel-wise labeled very high resolution resource for this task. They also proposed a multitask Fully Convolutional Neural Network (FCNN) with a sequential training scheme, which achieved a 25.49% Kappa coefficient for semantic change detection, outperforming existing state-of-the-art methods.

Key insights

• Multitask learning, integrating land cover mapping with change detection within a single network, significantly improves semantic change detection performance compared to separate or naive comparison methods.
• A sequential training approach effectively balances multitask learning objectives without the need for manual weighting hyperparameter tuning, making the process more efficient and practical.
• The availability of a large-scale, VHR dataset like HRSCD is crucial for the effective training and generalization of deep supervised learning models for semantic change detection, revealing performance gains that smaller datasets cannot.

Summary by alphaXiv.