Question about ScanRefer validation scenes (sceneXXXX_00) vs. ScanNetV2 partitions #23
Description
Hello,
I noticed a potential inconsistency in the dataset splits.
In scanrefer_filtered_val.json, all scenes are annotated as sceneXXXX_00. However, the ReferIt3D data follows the partitions defined in scannetv2_val.txt, like sceneXXXX_00 and sceneXXXX_01.
Could this lead to visibility issues? For example, could a caption in ScanRefer describe an object that is structurally part of _00 but effectively outside the visible region of the partition defined in the standard validation list? I am worried that the model might be trying to ground objects that are not present in the input point cloud due to this mismatch.
Any clarification on how this is handled would be appreciated.