The architecture of Visual Chronicles is designed to overcome the challenge of analyzing datasets that are too large for traditional MLLMs to process directly. It employs a two-stage approach: first, local analysis is performed to detect changes between pairs of images at specific locations, leveraging the semantic understanding capabilities of MLLMs. This is followed by a global aggregation stage, where detected changes are clustered and summarized into broader trends. This bottom-up methodology allows the system to scale efficiently, providing insights without the need for pre-defined target subjects or extensive labeled training data. Visual Chronicles has demonstrated superior performance compared to traditional unsupervised image analysis tools, particularly in its ability to identify nuanced and contextually relevant changes.
Beyond temporal change detection, Visual Chronicles supports flexible querying, enabling users to search for trends related to specific concepts, such as retail store openings or closures, or to identify unusual or noteworthy features within a city. The system can connect visual discoveries to socioeconomic events or policy decisions, offering a bridge between image-based evidence and real-world implications. Its interactive interface allows users to explore before-and-after image pairs, visualize change locations, and filter results by time window or semantic relevance. This makes Visual Chronicles not only a technical achievement in large-scale image analysis but also a practical platform for exploratory research and decision support.
Key features include:
- Massive-scale image analysis using multimodal large language models
- Two-stage pipeline for local change detection and global trend aggregation
- Supports open-ended and concept-specific queries without labeled training data
- Interactive visualization of trends, before/after image pairs, and change locations
- Connects visual trends to socioeconomic events and policy decisions