Reproduce “ConInfer: Context-Aware Inference for Training-Free Open-Vocabulary Remote Sensing Segmentation” on OpenResearch

About this paper

Wenyang Chen, Zhanxuan Hu, Yaping Zhang, Hailong Ning, Yonghang Tai

ConInfer introduces a training-free, context-aware inference framework for open-vocabulary remote sensing segmentation, addressing the limitations of independent patch-level predictions. It integrates semantic priors from Vision-Language Models with structural regularization from Vision Foundation Models, achieving an average 7.21% mIoU improvement on multi-class segmentation and 10.89% IoU improvement on single-class extraction over strong baselines.

Key insights

• Explicitly incorporating global contextual information during inference is crucial for generating consistent and accurate open-vocabulary remote sensing segmentation maps, moving beyond isolated patch predictions.
• Vision Foundation Models like DINOv3, trained purely on visual data, effectively capture fine-grained structural details and contextual regularities in remote sensing imagery, complementing the semantic understanding of VLMs.
• A joint, iterative optimization process that allows semantic priors to guide structural context modeling and vice-versa is essential for deriving calibrated predictions that balance semantic accuracy with spatial coherence.

Summary by alphaXiv.