document-understanding

Geospatial foundation models (GFMs) offer powerful representations for Earth observation, yet effective adaptation to dense prediction tasks remains underexplored. We introduce KAO-TerraMind, a targeted adaptation framework that systematically investigates three factors, backbone optimisation, loss design, and decoder architecture. Through controlled ablations, we show that decoder-focused adaptation achieves the best performance, improving overall accuracy from 78.44% to 78.90%, mIoU from 40.82% to 41.13%, F1 from 52.57% to 52.80%, and mAP from 52.63% to 53.03%. In contrast, stronger backbone optimisation alone provides no measurable mIoU improvement, while an increased Dice loss benefits under-represented classes. These findings highlight decoder design and multi-level feature integration as effective strategies for adapting geospatial foundation models to dense prediction. We will release the finetuning code and evaluation protocol to facilitate reproducible research.

CHULA-OCR introduces an uncertainty-aware framework for recognizing text in Thai land title deeds, addressing challenges such as degraded scans, complex document layouts, handwritten annotations, and visually similar Thai characters. By integrating an entropy-guided weighting mechanism, the framework prioritizes reliable character regions and reduces the influence of ambiguous visual information. Experimental results demonstrate improved robustness compared with strong OCR and document understanding baselines across diverse document conditions. Furthermore, uncertainty estimation enables the identification of unreliable predictions for human verification, supporting efficient and reliable digitization of large-scale land administration archives.