cfg.yml

PIPELINE:
  DATA:
    ROOT: 'kitti'
    SCRIPT: 'kitti_object'
    LOADER: 'kittiObjectDataset'
    NUM_WORKERS: 1
  MODEL:
    TASK: detection
    SCRIPT: 'faster_rcnn_lidar3d.bv_net'
    NAME: 'bv_net'
  TRAIN:
    START_EPOCH: 0
    END_EPOCH: 15
    BATCH_SIZE: 4

    OPTIMIZER: adam
    LEARNING_RATE: 0.00005
    WEIGHT_DECAY: 0.0005
    DOUBLE_BIAS: false
    BIAS_DECAY: false
    MOMENTUM: 0.9
    LR_DECAY_GAMMA: 0.1
    LR_DECAY_EPOCH: 10

    CLIP_GRAD: -1

    PRETRAINED: False
    PRETRAINED_WEIGHT: '' 
    RESUME: false
    RESUME_EPOCH: 0
    FINE_TUNE: false
    FINE_TUNE_STRICT: false
    FINE_TUNE_MODEL: ''
    FINE_TUNE_STATE_DICT: true
    SNAPSHOT_INTERVAL: 10000
    SNAPSHOT_PREFIX: results/snapshots/bv_net

    EVAL_INTERVAL: -1
    DISP_INTERVAL: 200

  VAL:
    BATCH_SIZE: 4
  TEST:
    BATCH_SIZE: 4
MODEL:
  DETECTION: 
    CLASSES: 
      - __background__
      - car
    WITH_FC6: true
    RESNET:
      FIX_BN_AFFINE: true
      FIX_BN: true
    FORMAT_NAME: true
    WITH_DROPOUT: true
    BN_MOMENTUM: 0.1
    CLASS_AGNOSTIC: false
    POOLING_SIZE: 7
    CROP_RESIZE_WITH_MAX_POOL: true
    LIGHT_HEAD: true
    PSROI_ALIGN: true
    PSROI_ALIGN_TYPE: 1
    POOLING_MODE: crop
    IM_LOG: false
    RCNN_CLS_WEIGHT: 1.0
    RCNN_BOX_WEIGHT: 1.0
    TRUNCATED: false
    PRETRAINED: false
    PRETRAINED_WEIGHT: ''
    FC_CHANNEL: 512
    NUM_LAYERS: 0
    NET_TYPE: net1
    ###########proposal target config
    BBOX_NORMALIZE_MEANS:
      - 0.0
      - 0.0
      - 0.0
      - 0.0
    BBOX_NORMALIZE_STDS:
      - 0.1
      - 0.1
      - 0.2
      - 0.2
    BBOX_INSIDE_WEIGHTS:
      - 1.0
      - 1.0
      - 1.0
      - 1.0
    INCLUDE_GTBOX: false
    RPN_BATCH_SIZE: 128
    FG_FRACTION: 0.25
    BBOX_NORMALIZE_TARGETS_PRECOMPUTED: true
    FG_THRESH: 0.7
    BG_THRESH_HI: 0.7 # 0.5 default 
    BG_THRESH_LO: 0.0

    ###########rpn config
    ANCHOR_SCALES:
      - 2
      - 4
      - 8
      - 16
      - 32
    ANCHOR_RATIOS:
      - 0.5
      - 1
      - 2
    FEAT_STRIDE:
      - 16

    RPN_CLS_WEIGHT: 1.0
    RPN_BOX_WEIGHT: 1.0
    ## proposal config
    RPN_PRE_NMS_TOP_N: 6000
    RPN_POST_NMS_TOP_N: 300
    RPN_NMS_THRESH: 0.7
    RPN_MIN_SIZE: 5
    USE_GPU_NMS: true

    ## anchor target layer config
    RPN_CLOBBER_POSITIVES: false
    RPN_NEGATIVE_OVERLAP: 0.3
    WITH_IGNORE: false
    RPN_IGNORE_OVERLAP: 0.5
    RPN_POSITIVE_OVERLAP: 0.7
    RPN_BBOX_OVERLAP: -1.0
    RPN_FG_FRACTION: 0.5
    RPN_BATCHSIZE: 256
    RPN_BBOX_INSIDE_WEIGHTS:
      - 1.0
      - 1.0
      - 1.0
      - 1.0
    RPN_POSITIVE_WEIGHT: -1.0
  
  LIDAR_BV3D_DET:   
    CLASSES: ['__background__', 'car']
    CLASS_AGNOSTIC: True
    LIGHT_HEAD: True
    ROI_CONV: False
    USE_IGNORE: True
    USE_FILTER: False

    FC_CHANNEL: 1024
    FC1_CHANNEL: 0
    FC2_CHANNEL: 0
    LIGHT_HEAD: True
    PSROI_ALIGN: True
    ROI_POOL_TYPE: 'maxpool'
    # Birdview projection parameters
    BV_PARAM:
      # height of input birdview feature map
      HEIGHT: 800
      # width of input birdview feature map
      WIDTH: 704
      # Y offset
      ORI_Y: 40.0
      # X offset
      ORI_X: 0.0
      # Y resolution
      RES_Y: 0.1
      # X resolution
      RES_X: 0.1
      # depth (channel) of input voxel feature
      DEPTH: 10
      # Z offset
      ORI_Z: 0.0
      # Z resolution
      RES_Z: 0.4
      # camera height
      MIN_HEIGHT: -2.0
      # max height
      MAX_HEIGHT: 2.0
      # hand-crafted feature types
      NUM_SLICES: 1
      USE_HEIGHT_SLICE: True
      USE_MEAN_HEIGHT: True
      USE_MAX_HEIGHT: False
      USE_INTENSITY_SLICE: True
      USE_MEAN_INTENSITY: True
      USE_TOP_INTENSITY: False
      USE_DENSITY: True
      USE_NONEMPTY: True
      USE_DISTANCE: True
      USE_DIRECTION: True
      USE_RGB_SLICE: False
      OUTPUT_SCORE: True
      RM_EMPTY_ANCHORS: False
      USE_CENTER: False
    BIRDVIEW:
      USE_VOXEL: False
      USE_POINTNET: False
      ANCHOR_SIZES: [[3.9,1.6], [1.6,3.9]]  # meter
      NUM_ANCHORS: 2
      IMU_HEIGHT: 1.73   # meter
      RPN_FEAT_STRIDE: 2
      RCNN_FEAT_STRIDE: 2
      PRECOMPUTED: False
      PROPOSAL_TYPE: 'corner'    # 'corner', 'center'
      MULTI_BIN_START: 1.0
      ANGLE_TYPE: 'global'
    # anchor target config
    ANCHOR_ORIENTED_OVERLAP: False

    # rpn config
    RPN_ADAPTIVE_THRESH: True
    RPN_POSITIVE_HI: 0.7
    RPN_POSITIVE_LO: 0.3
    RPN_NEGATIVE_HI: 0.6
    RPN_NEGATIVE_LO: 0.2
    # If an anchor statisfied by positive and negative conditions set to negative
    RPN_CLOBBER_POSITIVES: False
    # Max number of foreground examples
    RPN_FG_FRACTION: 0.5
    # Total number of examples
    RPN_BATCHSIZE: 256
    # Give the positive RPN examples weight of p * 1 / {num positives}
    # and give negatives a weight of (1 - p)
    # Set to -1.0 to use uniform example weighting
    RPN_POSITIVE_WEIGHT: -1.0

    ## proposal 3d config
    # NMS threshold used on RPN proposals
    RPN_NMS_THRESH: 0.7
    # Number of top scoring boxes to keep before apply NMS to RPN proposals
    RPN_PRE_NMS_TOP_N: 12000
    # Number of top scoring boxes to keep after applying NMS to RPN proposals
    RPN_POST_NMS_TOP_N: 2000
    # Proposal height and width both need to be greater than RPN_MIN_SIZE (at orig image scale)
    RPN_MIN_SIZE: 8
    ## Use oriented corner box nms for rpn
    RPN_ORIENTED_NMS: False
    USING_PY_METHOD: False
    RPN_CONV: False

    ## rcnn config
    RCNN_CLS_WEIGHT: 1.0
    RCNN_BBOX_WEIGHT: 1.0

  CLASSIFICATION: ''
  PARSING: ''
  PRED: ''
  STEREO: ''
  TRACKING: ''