Bridging Geographic Bias in Urban Streetscape Inference via Lifelong Learning with Visual-Semantic Pivoting
Researchers have developed a new lifelong learning framework called HVSP-LL to address geographic bias in urban streetscape inference. This framework uses a visual-semantic pivoting module to organize landscape concepts and align image features with semantic anchors, enabling transferable representations. An equity-aware rehearsal mechanism sequentially absorbs new urban regions while minimizing perception gaps between cities. The system achieved a 6.1-point improvement over existing continual learning baselines on a benchmark spanning twelve cities and seven perceptual dimensions, significantly reducing the inter-city perception gap. AI
IMPACT This research offers a novel approach to mitigate bias in AI models used for urban planning and public health, potentially leading to more equitable decision-making.