Skip to content

Training dataest ratio and Panoptic segmentation reproduce #17

Description

@qjadud1994

Thank you for sharing impressive work! @congvvc


What is the training dataset ratio used for HyperSeg training?

For example, PSALM followed this training dataset ratio:
coco_panoptic : ref_coco : coco_interactive : mm_conv = 1 : 1 : 1 : 1

Likewise, InstructSeg followed this training dataset ratio:
ref_coco : mm_conv : ref_vos : reason : reason_vos = 4 : 2 : 4 : 1 : 3


The main reason for this question is that I couldn't reproduce the 61.2 COCO panoptic PQ score. I used the COCO panoptic segmentation training code from PSALM and trained HyperSeg, but the maximum PQ score of HyperSeg was 56.0.
Can you advise me on this result? Is there any difference from PSALM for panoptic segmentation training?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions