[ORE] Clarification needed about training procedure

[kinivi]

Dear authors,

Thank you for opensourcing your work. I have quick questions about ORE application that needed clarification:

• Have you used BOTH mavl and RPN pseudolables during training? Or default ORE pseudolabelling was turned off?
• Did you use default training settings for the ORE model?

Thank you for the help in advance!

[mmaaz60]

Hi @kinivi,

1. Default pseudo labeling was turned off.
2. Yes, we use the default training setting for ORE.

I hope it will help, do let me know if you have any questions. Thank You

[kinivi]

@mmaaz60 thank you for the rapid response. I have a follow-up question.
ORE
have you modified the training script beyond configuration? I have trained using the following conf on the modified annotations (added unk my mvit). However I found that model doesn't detect any unknowns as-is. It starts detection only after fitting Weibull distribution using VOC07 val set.

MODEL:
  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-50.pkl"
DATASETS:
  TRAIN: ("voc_2007_train", "voc_2012_trainval") # t1_voc_coco_2007_train, t1_voc_coco_2007_ft
  #TEST: ('voc_coco_2007_test', 't1_voc_coco_2007_known_test')   # voc_coco_2007_test, t1_voc_coco_2007_test, t1_voc_coco_2007_val
SOLVER:
  STEPS: (12000, 16000)
  MAX_ITER: 18000
  WARMUP_ITERS: 100
OUTPUT_DIR: "./output/udet_mavl"
OWOD:
  PREV_INTRODUCED_CLS: 0
  CUR_INTRODUCED_CLS: 20
  ENABLE_THRESHOLD_AUTOLABEL_UNK: False
DATALOADER:
  NUM_WORKERS: 7

Could you please elaborate on what you after generating pseudolabells for the ORE?

Thank you in advance.