[자율주행 AI 챌린지] Voxel_RCNN VoxelNext 개선방안 보고서

[이승현]

Voxel_RCNN 제출결과

1. voxel_size[0.075,0.075,0.15] 파인튜닝 결과

0.075_80epoch_결과

해상도0.075 파인튜닝 결과

0.075 파인튜닝 추론시간

# custom_av_dataset.yaml

DATASET: 'CustomAvDataset'

DATA_PATH: '../data/custom_av'

#POINT_CLOUD_RANGE: [-70.4, -70.4, -4.0, 70.4, 70.4, 4.0]

POINT_CLOUD_RANGE: [-75.0, -75.0, -2.0, 75.0, 75.0, 4.0] #

MAP_CLASS_TO_KITTI: {

    'Vehicle': 'Car',

    'Pedestrian': 'Pedestrian',

    'Cyclist': 'Cyclist',

}

DATA_SPLIT: {

    'train': train,

    'test': val

}

INFO_PATH: {

    'train': [custom_av_infos_train.pkl],

    'test': [custom_av_infos_val.pkl],

}

POINT_FEATURE_ENCODING: {

    encoding_type: absolute_coordinates_encoding,

    used_feature_list: ['x', 'y', 'z'],

    src_feature_list: ['x', 'y', 'z', 'intensity'],

}

DATA_AUGMENTOR:

    DISABLE_AUG_LIST: ['placeholder','line_downsample','random_world_translation','random_points_dropout'] #,gt_sampling','random_world_scaling','random_world_flip','random_world_rotation',placeholder','random_points_dropout',,

    AUG_CONFIG_LIST:

        - NAME: gt_sampling

          USE_ROAD_PLANE: False

          DB_INFO_PATH:

              - custom_av_dbinfos_train.pkl

          PREPARE: {

             filter_by_min_points: ['Vehicle:5', 'Pedestrian:5', 'Cyclist:5'],

          }

          SAMPLE_GROUPS: ['Vehicle:10', 'Pedestrian:16', 'Cyclist:14']

          NUM_POINT_FEATURES: 4

          DATABASE_WITH_FAKELIDAR: False

          REMOVE_EXTRA_WIDTH: [0.0, 0.0, 0.0]

          LIMIT_WHOLE_SCENE: True

        - NAME: random_world_flip

          ALONG_AXIS_LIST: ['x', 'y']

        # - NAME: random_world_rotation

        #   WORLD_ROT_ANGLE: [-0.78539816, 0.78539816]

        # - NAME: random_world_scaling

        #   WORLD_SCALE_RANGE: [0.95, 1.05]

        # 약한 증강용

        - NAME: random_world_rotation

          WORLD_ROT_ANGLE: [-0.39269908, 0.39269908]    # ±22.5°로 축소

        - NAME: random_world_scaling

          WORLD_SCALE_RANGE: [0.98, 1.02]  

        - NAME: random_world_translation # 추가한 증강기법

          NOISE_TRANSLATE_STD: [0.2, 0.2, 0.2]

        - NAME: random_points_dropout # 추가한 증강기법

          DROP_RATE: 0.15         # ← 25% 제거

          ENSURE_MIN_KEEP: 90000  # ← 너무 과하게 줄어드는 것 방지

          PROB: 0.4               # ← 50% 확률로 적용

        - NAME: line_downsample

          TARGET_LINES: 64

          LINE_BINS: 128

          TOLERANCE: 6          # 60~68 라인이면 스킵

          MIN_BIN_OCCUPANCY: 120

          ELEVATION_FROM: 'auto'

          PROB: 0.5             # 프레임마다 시도하되, 함수 내부가 알아서 스킵/적용

DATA_PROCESSOR:

    - NAME: mask_points_and_boxes_outside_range

      REMOVE_OUTSIDE_BOXES: True

    - NAME: shuffle_points

      SHUFFLE_ENABLED: {

        'train': True,

        'test': True

      }

    - NAME: transform_points_to_voxels

      VOXEL_SIZE: [0.075, 0.075, 0.15] # 0.1,0.1,0.2 0.08 0.06

      MAX_POINTS_PER_VOXEL: 5

      MAX_NUMBER_OF_VOXELS: {

        'train': 250000, #150000

        'test': 300000 #150000

      }

# custom_voxel_rcnn.yaml

CLASS_NAMES: ['Vehicle', 'Pedestrian', 'Cyclist']

DATA_CONFIG:

    _BASE_CONFIG_: cfgs/dataset_configs/custom_av_dataset.yaml

    # DATA_PROCESSOR:

    #     -   NAME: mask_points_and_boxes_outside_range

    #         REMOVE_OUTSIDE_BOXES: True

    #         STRICT_MASK: True

    #     -   NAME: shuffle_points

    #         SHUFFLE_ENABLED: {

    #             'train': True,

    #             'test': True

    #         }

        # -   NAME: transform_points_to_voxels_placeholder

        #     VOXEL_SIZE: [ 0.08, 0.08, 0.15 ] #0.1 0.15

MODEL:

    NAME: VoxelRCNN

    VFE:

        NAME: DynMeanVFE

    BACKBONE_3D:

        NAME: VoxelBackBone8x

    MAP_TO_BEV:

        NAME: HeightCompression

        NUM_BEV_FEATURES: 256

    BACKBONE_2D:

        NAME: BaseBEVBackbone

        LAYER_NUMS: [5, 5]

        LAYER_STRIDES: [1, 2]

        NUM_FILTERS: [128, 256]

        UPSAMPLE_STRIDES: [1, 2]

        NUM_UPSAMPLE_FILTERS: [256, 256]

    DENSE_HEAD:

        NAME: CenterHead

        CLASS_AGNOSTIC: False

        CLASS_NAMES_EACH_HEAD: [

            [ 'Vehicle', 'Pedestrian', 'Cyclist' ]

        ]

        SHARED_CONV_CHANNEL: 64

        USE_BIAS_BEFORE_NORM: True

        NUM_HM_CONV: 3 #2

        SEPARATE_HEAD_CFG:

            HEAD_ORDER: [ 'center', 'center_z', 'dim', 'rot' ]

            HEAD_DICT: {

                'center': { 'out_channels': 2, 'num_conv': 2 },

                'center_z': { 'out_channels': 1, 'num_conv': 2 },

                'dim': { 'out_channels': 3, 'num_conv': 2 },

                'rot': { 'out_channels': 2, 'num_conv': 2 },

            }

        TARGET_ASSIGNER_CONFIG:

            FEATURE_MAP_STRIDE: 8

            NUM_MAX_OBJS: 500

            GAUSSIAN_OVERLAP: 0.11 #0.15  해상도를 올렸기 때문에 너무작으면 가우시안 overlap를 조정함

            MIN_RADIUS: 1 # 1 보행자와 자전거의 recall값을 올리기 위해서

        LOSS_CONFIG:

            LOSS_WEIGHTS: {

                'cls_weight': 2.3, #1.0 2.0

                'loc_weight': 2.3, #2.0

                'code_weights': [ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 ]

            }

        CLASS_WEIGHTS:         # heatmap 분류 손실용

            Vehicle:   1.0

            Pedestrian: 2.7    

            Cyclist:   3.0 #2.7      

        LOC_CLASS_WEIGHTS:    

            Vehicle:   1.0

            Pedestrian: 1.2      

            Cyclist:   1.5      

        POST_PROCESSING:

            SCORE_THRESH: 0.1 # 0.1

            POST_CENTER_LIMIT_RANGE: [ -75.0, -75.0, -2.0, 75.0, 75.0, 4.0 ]  #-74.80, -65.92, -3.0, 74.80,65.92, 4.0 #-75.2, -75.2, -2, 75.2, 75.2, 4 $#65.92

            MAX_OBJ_PER_SAMPLE: 500 #500

            NMS_CONFIG:

                NMS_TYPE: nms_gpu

                NMS_THRESH: 0.7 #0.7

                NMS_PRE_MAXSIZE: 4096

                NMS_POST_MAXSIZE: 400

        DEBUG_LABEL_HIST: True

    ROI_HEAD:

        NAME: VoxelRCNNHead

        CLASS_AGNOSTIC: True #True

        SHARED_FC: [256, 256]

        CLS_FC: [256, 256]

        REG_FC: [256, 256]

        DP_RATIO: 0.3

        NMS_CONFIG:

            TRAIN:

                NMS_TYPE: nms_gpu

                MULTI_CLASSES_NMS: False

                NMS_PRE_MAXSIZE: 9000

                NMS_POST_MAXSIZE: 512

                NMS_THRESH: 0.65 # 0.8

            TEST:

                NMS_TYPE: nms_gpu

                MULTI_CLASSES_NMS: False

                NMS_PRE_MAXSIZE: 4096 #1024

                NMS_POST_MAXSIZE: 512 #100

                NMS_THRESH: 0.6 #0.7

                # NMS_PRE_MAXSIZE: 4096

                # NMS_POST_MAXSIZE: 300

                # NMS_THRESH: 0.85

        ROI_GRID_POOL:

            FEATURES_SOURCE: ['x_conv2', 'x_conv3', 'x_conv4']

            PRE_MLP: True

            GRID_SIZE: 8 #8

            POOL_LAYERS:

                x_conv2:

                    MLPS: [ [ 64, 64 ] ]

                    QUERY_RANGES: [ [ 3, 3, 2 ] ]

                    POOL_RADIUS: [ 0.4 ]

                    NSAMPLE: [ 16 ]

                    POOL_METHOD: max_pool

                x_conv3:

                    MLPS: [ [ 64, 64 ] ]

                    QUERY_RANGES: [ [ 3, 3, 2 ] ]

                    POOL_RADIUS: [ 0.8 ]

                    NSAMPLE: [ 16 ]

                    POOL_METHOD: max_pool

                x_conv4:

                    MLPS: [ [ 64, 64 ] ]

                    QUERY_RANGES: [ [ 3, 3, 2 ] ]

                    POOL_RADIUS: [ 1.6 ]

                    NSAMPLE: [ 16 ]

                    POOL_METHOD: max_pool

        TARGET_CONFIG:

            BOX_CODER: ResidualCoder

            ROI_PER_IMAGE: 192 #128 양성 풀 확대

            FG_RATIO: 0.5

            #SAMPLE_ROI_BY_EACH_CLASS: True

            CLS_SCORE_TYPE: roi_iou

            CLS_FG_THRESH: 0.70 #0.75 양성 확대

            CLS_BG_THRESH: 0.25

            CLS_BG_THRESH_LO: 0.1

            HARD_BG_RATIO: 0.8

            REG_FG_THRESH: 0.55 #0.55 작은 박스 IOU도 허용

        LOSS_CONFIG:

            CLS_LOSS: BinaryCrossEntropy

            REG_LOSS: smooth-l1

            CORNER_LOSS_REGULARIZATION: True

            LOSS_WEIGHTS: {

                'rcnn_cls_weight': 1.0,

                'rcnn_reg_weight': 1.0,

                'rcnn_corner_weight': 1.0,

                'code_weights': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]

            }

    POST_PROCESSING:

        RECALL_THRESH_LIST: [0.3, 0.5, 0.7]

        SCORE_THRESH: 0.12 #0.1 #0.2 #0.12

        OUTPUT_RAW_SCORE: False

        EVAL‎_METRIC: waymo

        NMS_CONFIG:

            MULTI_CLASSES_NMS: False  #False

            NMS_TYPE: nms_gpu

            NMS_THRESH: 0.6 #0.7

            NMS_PRE_MAXSIZE: 4096

            NMS_POST_MAXSIZE: 500

OPTIMIZATION:

    BATCH_SIZE_PER_GPU: 2

    NUM_EPOCHS: 20 #30

    OPTIMIZER: adam_onecycle

    LR: 0.0010 #0.003 #0.006

    WEIGHT_DECAY: 0.001

    MOMENTUM: 0.9

    MOMS: [0.95, 0.85]

    PCT_START: 0.1 #0.35 #0.35 #0.4

    DIV_FACTOR: 10

    DECAY_STEP_LIST: [35, 45]

    LR_DECAY: 0.1

    LR_CLIP: 0.0000001

    LR_WARMUP: False #True

    WARMUP_EPOCH: 1

    GRAD_NORM_CLIP: 10

0.075 해상도에서 ROI_GRID와 HM_CONV 을 6->8 2->3으로 늘리는 방법을 채택하여 소형 객체의 AP값을 올리는걸 기대하였는데  강한 증강 이후에 약한 증강으로 일반화를 노렸지만 ROI_GRID와 HM_CONV가 이미 충분히 학습이 된 상태에서 약한증강으로 튜닝을 하여 특징을 잃어버리는 결과를 가져온거 같다. 

2. voxel_size[0.08,0.08,0.15] 파인튜닝 결과

30epoch(강한18+약한12)

18epoch

0.08해상도 추론시간

# custom_av_dataset_finetuning.yaml

DATASET: 'CustomAvDataset'

DATA_PATH: '../data/custom_av'

#POINT_CLOUD_RANGE: [-70.4, -70.4, -4.0, 70.4, 70.4, 4.0]

POINT_CLOUD_RANGE: [-75.52, -75.52, -2.0, 75.52, 75.52, 4.0] #-74.88, -74.88, -2.0, 74.88, 74.88, 4.0

MAP_CLASS_TO_KITTI: {

    'Vehicle': 'Car',

    'Pedestrian': 'Pedestrian',

    'Cyclist': 'Cyclist',

}

DATA_SPLIT: {

    'train': train,

    'test': val

}

INFO_PATH: {

    'train': [custom_av_infos_train.pkl],

    'test': [custom_av_infos_val.pkl],

}

POINT_FEATURE_ENCODING: {

    encoding_type: absolute_coordinates_encoding,

    used_feature_list: ['x', 'y', 'z'],

    src_feature_list: ['x', 'y', 'z', 'intensity'],

}

DATA_AUGMENTOR:

    DISABLE_AUG_LIST: ['placeholder','random_points_dropout','line_downsample','random_world_translation'] #

    AUG_CONFIG_LIST:

        - NAME: gt_sampling

          USE_ROAD_PLANE: False

          DB_INFO_PATH:

              - custom_av_dbinfos_train.pkl

          PREPARE: {

             filter_by_min_points: ['Vehicle:5', 'Pedestrian:5', 'Cyclist:5'],

          }

          SAMPLE_GROUPS: ['Vehicle:8', 'Pedestrian:16', 'Cyclist:14']

          NUM_POINT_FEATURES: 4

          DATABASE_WITH_FAKELIDAR: False

          REMOVE_EXTRA_WIDTH: [0.0, 0.0, 0.0]

          LIMIT_WHOLE_SCENE: True

        - NAME: random_world_flip

          ALONG_AXIS_LIST: ['x', 'y']

        # - NAME: random_world_rotation

        #   WORLD_ROT_ANGLE: [-0.78539816, 0.78539816]

        # - NAME: random_world_scaling

        #   WORLD_SCALE_RANGE: [0.95, 1.05]

        # 약한 증강용

        - NAME: random_world_rotation

          WORLD_ROT_ANGLE: [-0.39269908, 0.39269908]    # ±22.5°로 축소

        - NAME: random_world_scaling

          WORLD_SCALE_RANGE: [0.98, 1.02]  

        - NAME: random_world_translation # 추가한 증강기법

          NOISE_TRANSLATE_STD: [0.2, 0.2, 0.2]

        - NAME: random_points_dropout # 추가한 증강기법

          DROP_RATE: 0.15         # ← 25% 제거

          ENSURE_MIN_KEEP: 90000  # ← 너무 과하게 줄어드는 것 방지

          PROB: 0.4               # ← 50% 확률로 적용

        - NAME: line_downsample

          TARGET_LINES: 64

          LINE_BINS: 128

          TOLERANCE: 6          # 60~68 라인이면 스킵

          MIN_BIN_OCCUPANCY: 120

          ELEVATION_FROM: 'auto'

          PROB: 0.5             # 프레임마다 시도하되, 함수 내부가 알아서 스킵/적용

DATA_PROCESSOR:

    - NAME: mask_points_and_boxes_outside_range

      REMOVE_OUTSIDE_BOXES: True

    - NAME: shuffle_points

      SHUFFLE_ENABLED: {

        'train': True,

        'test': True

      }

    - NAME: transform_points_to_voxels

      VOXEL_SIZE: [0.08, 0.08, 0.15] # 0.1,0.1,0.2

      MAX_POINTS_PER_VOXEL: 5

      MAX_NUMBER_OF_VOXELS: {

        'train': 200000, #150000

        'test': 200000 #150000

      }

# custom_voxel_rcnn_0.08.yaml

CLASS_NAMES: ['Vehicle', 'Pedestrian', 'Cyclist']

DATA_CONFIG:

    _BASE_CONFIG_: cfgs/dataset_configs/custom_av_dataset_0.08.yaml

    # DATA_PROCESSOR:

    #     -   NAME: mask_points_and_boxes_outside_range

    #         REMOVE_OUTSIDE_BOXES: True

    #         STRICT_MASK: True

    #     -   NAME: shuffle_points

    #         SHUFFLE_ENABLED: {

    #             'train': True,

    #             'test': True

    #         }

        # -   NAME: transform_points_to_voxels_placeholder

        #     VOXEL_SIZE: [ 0.08, 0.08, 0.15 ] #0.1 0.15

MODEL:

    NAME: VoxelRCNN

    VFE:

        NAME: DynMeanVFE

    BACKBONE_3D:

        NAME: VoxelBackBone8x

    MAP_TO_BEV:

        NAME: HeightCompression

        NUM_BEV_FEATURES: 256

    BACKBONE_2D:

        NAME: BaseBEVBackbone

        LAYER_NUMS: [5, 5]

        LAYER_STRIDES: [1, 2]

        NUM_FILTERS: [128, 256]

        UPSAMPLE_STRIDES: [1, 2]

        NUM_UPSAMPLE_FILTERS: [256, 256]

    DENSE_HEAD:

        NAME: CenterHead

        CLASS_AGNOSTIC: False

        CLASS_NAMES_EACH_HEAD: [

            [ 'Vehicle', 'Pedestrian', 'Cyclist' ]

        ]

        SHARED_CONV_CHANNEL: 64

        USE_BIAS_BEFORE_NORM: True

        NUM_HM_CONV: 2

        SEPARATE_HEAD_CFG:

            HEAD_ORDER: [ 'center', 'center_z', 'dim', 'rot' ]

            HEAD_DICT: {

                'center': { 'out_channels': 2, 'num_conv': 2 },

                'center_z': { 'out_channels': 1, 'num_conv': 2 },

                'dim': { 'out_channels': 3, 'num_conv': 2 },

                'rot': { 'out_channels': 2, 'num_conv': 2 },

            }

        TARGET_ASSIGNER_CONFIG:

            FEATURE_MAP_STRIDE: 8

            NUM_MAX_OBJS: 500

            GAUSSIAN_OVERLAP: 0.11 #0.1  해상도를 올렸기 때문에 너무작으면 가우시안 overlap를 조정함

            MIN_RADIUS: 1 # 2 보행자와 자전거의 recall값을 올리기 위해서

        LOSS_CONFIG:

            LOSS_WEIGHTS: {

                'cls_weight': 2.3, #1.0 2.0

                'loc_weight': 2.3, #2.0

                'code_weights': [ 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 ]

            }

        CLASS_WEIGHTS:         # heatmap 분류 손실용

            Vehicle:   1.0

            Pedestrian: 2.3    

            Cyclist:   2.7      

        LOC_CLASS_WEIGHTS:    

            Vehicle:   1.0

            Pedestrian: 1.2      

            Cyclist:   1.5      

        POST_PROCESSING:

            SCORE_THRESH: 0.1 # 0.1

            POST_CENTER_LIMIT_RANGE: [ -75.52, -75.52, -2.0, 75.52, 75.52, 4.0 ]  #-74.88, -74.88, -2.0, 74.88, 74.88, 4.0

            MAX_OBJ_PER_SAMPLE: 500 #500

            NMS_CONFIG:

                NMS_TYPE: nms_gpu

                NMS_THRESH: 0.7 #0.7

                NMS_PRE_MAXSIZE: 4096

                NMS_POST_MAXSIZE: 400

        DEBUG_LABEL_HIST: True

    ROI_HEAD:

        NAME: VoxelRCNNHead

        CLASS_AGNOSTIC: True #True

        SHARED_FC: [256, 256]

        CLS_FC: [256, 256]

        REG_FC: [256, 256]

        DP_RATIO: 0.3

        NMS_CONFIG:

            TRAIN:

                NMS_TYPE: nms_gpu

                MULTI_CLASSES_NMS: False

                NMS_PRE_MAXSIZE: 9000

                NMS_POST_MAXSIZE: 512

                NMS_THRESH: 0.65 # 0.8

            TEST:

                NMS_TYPE: nms_gpu

                MULTI_CLASSES_NMS: False

                NMS_PRE_MAXSIZE: 4096 #1024

                NMS_POST_MAXSIZE: 512 #100

                NMS_THRESH: 0.6 #0.7

                # NMS_PRE_MAXSIZE: 4096

                # NMS_POST_MAXSIZE: 300

                # NMS_THRESH: 0.85

        ROI_GRID_POOL:

            FEATURES_SOURCE: ['x_conv2', 'x_conv3', 'x_conv4']

            PRE_MLP: True

            GRID_SIZE: 8 #6

            POOL_LAYERS:

                x_conv2:

                    MLPS: [ [ 64, 64 ] ]

                    QUERY_RANGES: [ [ 3, 3, 2 ] ]

                    POOL_RADIUS: [ 0.4 ]

                    NSAMPLE: [ 16 ]

                    POOL_METHOD: max_pool

                x_conv3:

                    MLPS: [ [ 64, 64 ] ]

                    QUERY_RANGES: [ [ 3, 3, 2 ] ]

                    POOL_RADIUS: [ 0.8 ]

                    NSAMPLE: [ 16 ]

                    POOL_METHOD: max_pool

                x_conv4:

                    MLPS: [ [ 64, 64 ] ]

                    QUERY_RANGES: [ [ 3, 3, 2 ] ]

                    POOL_RADIUS: [ 1.6 ]

                    NSAMPLE: [ 16 ]

                    POOL_METHOD: max_pool

        TARGET_CONFIG:

            BOX_CODER: ResidualCoder

            ROI_PER_IMAGE: 192 #128 양성 풀 확대

            FG_RATIO: 0.5

            #SAMPLE_ROI_BY_EACH_CLASS: True

            CLS_SCORE_TYPE: roi_iou

            CLS_FG_THRESH: 0.68 #0.75 , 0.70 양성 확대

            CLS_BG_THRESH: 0.25

            CLS_BG_THRESH_LO: 0.1

            HARD_BG_RATIO: 0.8

            REG_FG_THRESH: 0.55 #0.55 작은 박스 IOU도 허용

        LOSS_CONFIG:

            CLS_LOSS: BinaryCrossEntropy

            REG_LOSS: smooth-l1

            CORNER_LOSS_REGULARIZATION: True

            LOSS_WEIGHTS: {

                'rcnn_cls_weight': 1.0,

                'rcnn_reg_weight': 1.0,

                'rcnn_corner_weight': 1.0,

                'code_weights': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]

            }

    POST_PROCESSING:

        RECALL_THRESH_LIST: [0.3, 0.5, 0.7]

        SCORE_THRESH: 0.12 #0.1 #0.2 #0.12

        OUTPUT_RAW_SCORE: False

        EVAL‎_METRIC: waymo

        NMS_CONFIG:

            MULTI_CLASSES_NMS: False  #False

            NMS_TYPE: nms_gpu

            NMS_THRESH: 0.6 #0.7

            NMS_PRE_MAXSIZE: 4096

            NMS_POST_MAXSIZE: 500

OPTIMIZATION:

    BATCH_SIZE_PER_GPU: 2

    NUM_EPOCHS: 30 #30

    OPTIMIZER: adam_onecycle

    LR: 0.0012 #0.006

    WEIGHT_DECAY: 0.001

    MOMENTUM: 0.9

    MOMS: [0.95, 0.85]

    PCT_START: 0.15 #0.35 #0.4

    DIV_FACTOR: 10

    DECAY_STEP_LIST: [35, 45]

    LR_DECAY: 0.1

    LR_CLIP: 0.0000001

    LR_WARMUP: False

    WARMUP_EPOCH: 1

    GRAD_NORM_CLIP: 10

0.075 해상도에서는 연산량 + 추론시간 대비 이득이 크지가 않아서 이전에 49.3점수를 받은 ckpt를 받아서 30epoch을 finetuning하는 방식으로 돌아왔다. 하지만 기존의 점수와 크게 다르지 않고 오히려 점수가 떨어진것을 확인 할 수 있다. 때문에 0.08해상도에서 학습을 돌리고 ROI_GRID=8을 유지하면서 HM_CONV를 늘려 소형객체 AP값을 올려보는 방식으로 50점의 도전을 진행할거같다.

VoxelNext 제출결과

0.075해상도 결과

0.075 추론결과

# custom_voxelnext_0.075.yaml

CLASS_NAMES: ['Vehicle', 'Pedestrian', 'Cyclist']

DATA_CONFIG:

    _BASE_CONFIG_: cfgs/dataset_configs/custom_av_dataset.yaml

MODEL:

    NAME: VoxelNeXt

    VFE:

        NAME: MeanVFE

    BACKBONE_3D:

        NAME: VoxelResBackBone8xVoxelNeXt

        SPCONV_KERNEL_SIZES: [5, 5, 3, 3]

        OUT_CHANNEL: 256 #256

        CHANNELS: [32, 64, 128, 256, 256] # 마지막 256

    DENSE_HEAD:

        NAME: VoxelNeXtHead

        IOU_BRANCH: True #True

        CLASS_AGNOSTIC: False

        INPUT_FEATURES: 256

        CLASS_NAMES_EACH_HEAD: [

            ['Vehicle', 'Pedestrian', 'Cyclist']

        ]

        SHARED_CONV_CHANNEL: 256 #256

        USE_BIAS_BEFORE_NORM: True

        NUM_HM_CONV: 2 # 2 VoxelNext-K3에서 3x3으로 변형했는데 2퍼 상승됨

        SEPARATE_HEAD_CFG:

            HEAD_ORDER: ['center', 'center_z', 'dim', 'rot']

            HEAD_DICT: {

                'center': {'out_channels': 2, 'num_conv': 2},

                'center_z': {'out_channels': 1, 'num_conv': 2},

                'dim': {'out_channels': 3, 'num_conv': 2},

                'rot': {'out_channels': 2, 'num_conv': 2},

                'iou': {'out_channels': 1, 'num_conv': 2},

            }

        RECTIFIER: [0.68, 0.71, 0.65]

        TARGET_ASSIGNER_CONFIG:

            FEATURE_MAP_STRIDE: 8

            NUM_MAX_OBJS: 500

            GAUSSIAN_OVERLAP: 0.1 # 0.1 -> 0.2 -> 0.1

            MIN_RADIUS: 1 # 2 -> 3 -> 2

        LOSS_CONFIG:

            CLASS_LOSS_WEIGHTS: [1.5, 3.0, 2.5] # 2.0 1.7

            # - Pedestrian/Cyclist의 long-tail 보정: 분류 히트맵 손실에 더 큰 가중

            # - Vehicle은 충분히 자주 등장하므로 1.0 유지

            # FOCAL_ALPHA: 0.25          # 클래스 불균형 완화(포컬 논문 기본값)

            # FOCAL_GAMMA: 2.0           # hard-example에 더 집중

            LOSS_WEIGHTS: {

                'cls_weight': 1.3,

                'loc_weight': 2.0,

                iou_weight: 1.5,

                'code_weights': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]

            }

        # LOSS_CONFIG:

        #     LOSS_WEIGHTS: {

        #       'cls_weight': 1.5,        # 1.0 → 1.5 : coarse grid에서 분류 신호 강화

        #       'loc_weight': 1.5,        # 2.0 → 1.5 : 과도한 회귀 가중 완화(불안정성 감소)

        #       # [cx, cy, cz, w, l, h, rot_sin, rot_cos] 가정

        #       'code_weights': [0.75, 0.75, 1.0, 1.2, 1.2, 1.2, 1.1, 1.1]

        #       # - center xy: 0.75로 약간 낮춤(양자화 한계 고려)

        #       # - dim / rot: ↑ (IoU 개선에 직결, AP/APH 방어)

        #     }

        POST_PROCESSING:

            SCORE_THRESH: 0.1 # 0.1 -> 0.2~0.3

            POST_CENTER_LIMIT_RANGE: [-75.0, -75.0, -2.0, 75.0, 75.0, 4.0]

            MAX_OBJ_PER_SAMPLE: 500

            NMS_CONFIG:

                NMS_TYPE: nms_gpu

                MULTI_CLASSES_NMS: False

                NMS_THRESH: [0.65, 0.7, 0.7] #0.7

                NMS_PRE_MAXSIZE: [2048, 1024, 1024] #4096  #[4096]

                NMS_POST_MAXSIZE: [200, 150, 150] #500

    POST_PROCESSING:

        RECALL_THRESH_LIST: [0.3, 0.5, 0.7]

        EVAL‎_METRIC: waymo

OPTIMIZATION:

    BATCH_SIZE_PER_GPU: 1

    NUM_EPOCHS: 40

    OPTIMIZER: adam_onecycle

    LR: 0.003 #0.003 #0.0015

    WEIGHT_DECAY: 0.01

    MOMENTUM: 0.9

    MOMS: [0.95, 0.85]

    PCT_START: 0.35 #0.35

    DIV_FACTOR: 10

    DECAY_STEP_LIST: [15, 25]

    LR_DECAY: 0.1

    LR_CLIP: 0.0000001

    LR_WARMUP: True

    WARMUP_EPOCH: 1

    GRAD_NORM_CLIP: 10

# custom_av_dataset_voxelnext_0.075.yaml

DATASET: 'CustomAvDataset'

DATA_PATH: '../data/custom_av'

#POINT_CLOUD_RANGE: [-70.4, -70.4, -4.0, 70.4, 70.4, 4.0]

POINT_CLOUD_RANGE: [-75.0, -75.0, -2.0, 75.0, 75.0, 4.0] #

MAP_CLASS_TO_KITTI: {

    'Vehicle': 'Car',

    'Pedestrian': 'Pedestrian',

    'Cyclist': 'Cyclist',

}

DATA_SPLIT: {

    'train': train,

    'test': val

}

INFO_PATH: {

    'train': [custom_av_infos_train.pkl],

    'test': [custom_av_infos_val.pkl],

}

POINT_FEATURE_ENCODING: {

    encoding_type: absolute_coordinates_encoding,

    used_feature_list: ['x', 'y', 'z'],

    src_feature_list: ['x', 'y', 'z', 'intensity'],

}

DATA_AUGMENTOR:

    DISABLE_AUG_LIST: ['placeholder','line_downsample','random_world_translation','random_points_dropout'] #gt_sampling','random_world_scaling','random_world_flip','random_world_rotation',placeholder','random_points_dropout','line_downsample' 'random_points_dropout',

    AUG_CONFIG_LIST:

        - NAME: gt_sampling

          USE_ROAD_PLANE: False

          DB_INFO_PATH:

              - custom_av_dbinfos_train.pkl

          PREPARE: {

             filter_by_min_points: ['Vehicle:5', 'Pedestrian:5', 'Cyclist:5'],

          }

          SAMPLE_GROUPS: ['Vehicle:10', 'Pedestrian:18', 'Cyclist:14']

          NUM_POINT_FEATURES: 4

          DATABASE_WITH_FAKELIDAR: False

          REMOVE_EXTRA_WIDTH: [0.0, 0.0, 0.0]

          LIMIT_WHOLE_SCENE: True

        - NAME: random_world_flip

          ALONG_AXIS_LIST: ['x', 'y']

        - NAME: random_world_rotation

          WORLD_ROT_ANGLE: [-0.78539816, 0.78539816]

        - NAME: random_world_scaling

          WORLD_SCALE_RANGE: [0.95, 1.05]

        # 약한 증강용

        # - NAME: random_world_rotation

        #   WORLD_ROT_ANGLE: [-0.39269908, 0.39269908]    # ±22.5°로 축소

        # - NAME: random_world_scaling

        #   WORLD_SCALE_RANGE: [0.98, 1.02]  

        - NAME: random_world_translation # 추가한 증강기법

          NOISE_TRANSLATE_STD: [0.2, 0.2, 0.2]

        - NAME: random_points_dropout # 추가한 증강기법

          DROP_RATE: 0.15         # ← 25% 제거

          ENSURE_MIN_KEEP: 90000  # ← 너무 과하게 줄어드는 것 방지

          PROB: 0.5               # ← 50% 확률로 적용

        - NAME: line_downsample

          TARGET_LINES: 64

          LINE_BINS: 128

          TOLERANCE: 6          # 60~68 라인이면 스킵

          MIN_BIN_OCCUPANCY: 120

          ELEVATION_FROM: 'auto'

          PROB: 0.5             # 프레임마다 시도하되, 함수 내부가 알아서 스킵/적용

DATA_PROCESSOR:

    - NAME: mask_points_and_boxes_outside_range

      REMOVE_OUTSIDE_BOXES: True

    - NAME: shuffle_points

      SHUFFLE_ENABLED: {

        'train': True,

        'test': True

      }

    - NAME: transform_points_to_voxels

      VOXEL_SIZE: [0.075, 0.075, 0.15] # 0.1,0.1,0.2 0.08 0.06

      MAX_POINTS_PER_VOXEL: 5

      MAX_NUMBER_OF_VOXELS: {

        'train': 200000, #150000

        'test': 300000 #150000

      }

1등 takeout팀의 모델이 voxelnext로 추정되어 선택하였다. 먼저 소형객체의 AP값 확보를 위해 0.075의 해상도를 채택하였고 IOU_BREANCH 옵션이 있는데 이 옵션은 0.075해상도 + 4070SUPER Ti 성능(VRAM 16G)에서는 감당이 안되어서 옵션을 끄고 진행하였다. HM 파라미터들은 해상도에 맞게 조정하여 40epoch 학습을 돌렸지만 자전거 클래스는 잘 찾지만 보행자를 잘 못찾는 현상을 발견해 현재 해상도를 0.1로 조정하고 IOU_BREANCH 옵션을 키고 학습을 진행중에 있다.