search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.39k stars 652 forks source link

BoxInst-Performance-on-Pascal-VOC-2012 #580

Open Ray-rui opened 2 years ago

Ray-rui commented 2 years ago

Enjoyed your excellent work "BoxInst: High-Performance Instance Segmentation with Box Annotations" !

I am trying to reproduce the result of BoxInst on Pascal-VOC-2012. However, the result I got is different from that in the paper. I am wondering if you could share the config files.

Best regards.

tianzhi0549 commented 2 years ago

Hi, it has been a long time since we released this code. I am unsure if the following config is correct, but please try.

_BASE_: "../Base-BoxInst.yaml"
MODEL:
  WEIGHTS: "detectron2://ImageNetPretrained/MSRA/R-101.pkl"
  RESNETS:
    DEPTH: 101
  FCOS:
    NUM_CLASSES: 20
  BOXINST:
    ENABLED: True
    BOTTOM_PIXELS_REMOVED: 0
    PAIRWISE:
      WARMUP_ITERS: 1000
INPUT:
  MIN_SIZE_TRAIN: (480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800)
  MIN_SIZE_TEST: 800
DATASETS:
  TRAIN: ('voc_2007_trainval', 'voc_2012_trainval')
  TEST: ('voc_2007_test',)
SOLVER:
  STEPS: (12000, 16000)
  MAX_ITER: 18000  # 17.4 epochs
  WARMUP_ITERS: 100
OUTPUT_DIR: "output/BoxInst_VOC_R_101_FPN"