CAGroup3D.yaml

CLASS_NAMES: [ 'cabinet', 'bed', 'chair', 'sofa', 'table', 'door', 'window',
               'bookshelf', 'picture', 'counter', 'desk', 'curtain',
               'refrigerator', 'showercurtrain', 'toilet', 'sink', 'bathtub',
               'garbagebin']

DATA_CONFIG: 
    _BASE_CONFIG_: cfgs/dataset_configs/scannet_dataset.yaml

VOXEL_SIZE: &VOXEL_SIZE 0.02
N_CLASSES: &N_CLASSES 18
SEMANTIC_THR: &SEMANTIC_THR 0.15

MODEL:
    NAME: CAGroup3D
    VOXEL_SIZE: *VOXEL_SIZE
    SEMANTIC_MIN_THR: 0.05
    SEMANTIC_ITER_VALUE: 0.02
    SEMANTIC_THR: *SEMANTIC_THR
    BACKBONE_3D:
        NAME: BiResNet
        IN_CHANNELS: 3
        OUT_CHANNELS: 64

    DENSE_HEAD:
        NAME: CAGroup3DHead
        IN_CHANNELS: [64, 128, 256, 512]
        OUT_CHANNELS: 64
        SEMANTIC_THR: *SEMANTIC_THR
        VOXEL_SIZE: *VOXEL_SIZE
        N_CLASSES: *N_CLASSES
        N_REG_OUTS: 6
        CLS_KERNEL: 9
        WITH_YAW: False
        USE_SEM_SCORE: False # if feed sem scores to the second-stage, default: False
        EXPAND_RATIO: 3
        ASSIGNER:
            NAME: CAGroup3DAssigner
            LIMIT: 27
            TOPK: 18
            N_SCALES: 4
        LOSS_OFFSET:
            NAME: SmoothL1Loss
            BETA: 0.04
            REDUCTION: sum
            LOSS_WEIGHT: 1.0
        LOSS_BBOX:
            NAME: IoU3DLoss
            WITH_YAW: False
            LOSS_WEIGHT: 1.0
        NMS_CONFIG:
            SCORE_THR: 0.01
            NMS_PRE: 1000
            IOU_THR: 0.5
    
    ROI_HEAD:
        NAME: CAGroup3DRoIHead
        NUM_CLASSES: *N_CLASSES
        MIDDLE_FEATURE_SOURCE: [3]
        GRID_SIZE: 7
        VOXEL_SIZE: *VOXEL_SIZE
        COORD_KEY: 2
        MLPS: [[64,128,128]]
        CODE_SIZE: 6
        ENCODE_SINCOS: False
        ROI_PER_IMAGE: 128
        ROI_FG_RATIO: 0.9
        REG_FG_THRESH: 0.3
        ROI_CONV_KERNEL: 5
        ENLARGE_RATIO: False
        USE_IOU_LOSS: False 
        USE_GRID_OFFSET: False
        USE_SIMPLE_POOLING: True
        USE_CENTER_POOLING: True
        LOSS_WEIGHTS:
            RCNN_CLS_WEIGHT: 1.0 # no use
            RCNN_REG_WEIGHT: 1.0 # set to 0.5 if use iou loss
            RCNN_IOU_WEIGHT: 1.0
            CODE_WEIGHT: [1., 1., 1., 1., 1., 1.]

    POST_PROCESSING:
        RECALL_THRESH_LIST: [0.25, 0.5]
        EVAL_METRIC: scannet

OPTIMIZATION:
    BATCH_SIZE_PER_GPU: 16 # 4x4 or 8x2
    NUM_EPOCHS: 1 #10

    OPTIMIZER: adamW 
    LR: 0.001
    WEIGHT_DECAY: 0.0001 
    DECAY_STEP_LIST: [7, 9] 
    LR_DECAY: 0.1
    GRAD_NORM_CLIP: 10

    # no use
    PCT_START: 0.4 
    DIV_FACTOR: 10 
    LR_CLIP: 0.0000001

    LR_WARMUP: False
    WARMUP_EPOCH: 1