minghanli / boxvis Goto Github PK

Thank you for your excellent work!
In your paper, the mentioned fully-supervised Mask2Former model for performance comparing achieves 43.2 mAP on YouTube-VIS-2021 validation subset. I tried to reproduce this baseline according to the cues from both the paper and the repository of yours, but I still get around 40.6 mAP as reported in the original Mask2Former paper.

I set the following training configs:

    cfg.INPUT.SAMPLING_FRAME_NUM = 3
    cfg.INPUT.SAMPLING_FRAME_RANGE = 10
    cfg.INPUT.MIN_SIZE_TRAIN = (320, 352, 384, 416, 448, 480, 512)
    cfg.SOLVER.IMS_PER_BATCH = 16
    cfg.SOLVER.BASE_LR = 0.0001
    cfg.SOLVER.STEPS = (12000,)
    cfg.SOLVER.MAX_ITER = 15000
    cfg.DATASETS.TRAIN: ("ytvis_2021_train", )
    cfg.DATASETS.TEST: ("ytvis_2021_val", )

Can you tell me how I can reproduce the fully-supervised Mask2Former precision of 43.2 mAP?

I want to train a BoxVIS model only with OVIS dataset, and I think the following params in yaml file should be adjusted as below:

    DATASETS:
        DATASET_RATIO: []
        TRAIN: ("ovis_train_org",),   # registered in the **buildin** file of original OVIS train subset
        TEST: ("ovis_val", )
    SEM_SEG_HEAD:
        NUM_CLASSES: 25
    MODEL:
        BoxVIS:
            BVISD_ENABLED: False

Are there any other config modifications I should adopt? Thanks for answering.