For F3DT2D, we applied the same threshold-based filtering strategy as WAPR.v2.1 (MUSE), with a filtering threshold of 0.35, while keeping all other settings unchanged. Since the original F3DT2D implementation does not provide masks, we generated them using SAM. For a small number of bounding boxes where valid masks could not be obtained, we used the filled bounding-box regions as the corresponding masks.