Efficient multi-concept inference for training-free OVCD

Develop efficient inference schemes that enable CoRegOVCD and related training-free open-vocabulary change detection frameworks to handle multiple queried concepts simultaneously or with minimal computational overhead while preserving accuracy.

Background

The current pipeline performs concept-specific inference, which is effective and comparatively fast for single-query evaluation but can require repeated passes for multiple concepts. The efficiency analysis demonstrates competitive single-concept throughput, motivating research on mechanisms to amortize computation across concepts or to process multiple queries jointly without sacrificing accuracy.