_BASE_: "Base.yaml"
SOLVER:
  TYPE: "sgd"
  IMS_PER_BATCH: 2
  BASE_LR: 0.0001
  STEPS: (5000, 12000)
  MAX_ITER: 10000 #4972
  WARMUP_ITERS: 0
  CHECKPOINT_PERIOD: 1000
TEST:
  EVAL_PERIOD: 100000
VIS_PERIOD: 100
DATASETS:
  TRAIN: ('SUNRGBD_train', 'SUNRGBD_val')
  TEST: ('SUNRGBD_test',)
  CATEGORY_NAMES: ('chair', 'table', 'cabinet', 'car', 'lamp', 'books', 'sofa', 'pedestrian', 'picture', 'window', 'pillow', 'truck', 'door', 'blinds', 'sink', 'shelves', 'television', 'shoes', 'cup', 'bottle', 'bookcase', 'laptop', 'desk', 'cereal box', 'floor mat', 'traffic cone', 'mirror', 'barrier', 'counter', 'camera', 'bicycle', 'toilet', 'bus', 'bed', 'refrigerator', 'trailer', 'box', 'oven', 'clothes', 'van', 'towel', 'motorcycle', 'night stand', 'stove', 'machine', 'stationery', 'bathtub', 'cyclist', 'curtain', 'bin')
MODEL:
  DEPTH_ON: True #whether to use the depth anything concated features # if do not use this, then we can use ["p2", "p3", "p4", "p5", "p6"], [[32], [64], [128], [256], [512]], otherwise only ["p2", "p3", "p4", "p5"], [[32], [64], [128], [256]]
  DEVICE: 'cpu'
  FPN:
    IN_FEATURES: ["p2", "p3", "p4", "p5"]
  RPN:
    IN_FEATURES: ["p2", "p3", "p4", "p5"]
  ANCHOR_GENERATOR:
    SIZES: [[32], [64], [128], [256]]  # One size for each in feature map
  ROI_HEADS:
    NAME: 'ROIHeads3DScore' # name of the class that is the 3d predictor
    IN_FEATURES: ["p2", "p3", "p4", "p5"]
    NUM_CLASSES: 50
    POSITIVE_FRACTION: 0.25 # we can use this to control the ratio of positive to negative sampled cubes in
  ROI_CUBE_HEAD:
    NAME: 'CubeHead' # name of the 3d head
    DIMS_PRIORS_ENABLED: True
    POOLER_TYPE: 'ROIAlignV2'
    POOLER_RESOLUTION: 7
  META_ARCHITECTURE: 'RCNN3D_combined_features' # name of the overall arch that calls the ROI_HEADS.NAME and ROI_CUBE_HEAD.NAME