一、引言
本次探讨的问题是：如何通过OpenPCDet改变预测的类别？
假设：我们原来预测的类别是Car，如何仅在yaml文件中添加一些配置，使得模型能够同时预测出Pedestrians和Cyclists。
下面就这个问题简单复现一下。
二、复现过程
首先我们以VirConv举例：
在VirConv.yaml文件中，上面的Class只有['Car']类别，因此首先我们要做的是要把其他两个类别加进来：
CLASS_NAMES: ['Car','Pedestrian', 'Cyclist'] # prediction class然后这样运行起来，预测的结果将会是：
也就是说，除了car类别以外，没有任何一个类别可以被成功预测出来。这是因为我们没有添加anchor生成器的一些参数：
DENSE_HEAD:
NAME: AnchorHeadSingle
CLASS_AGNOSTIC: False
USE_DIRECTION_CLASSIFIER: True
DIR_OFFSET: 0.78539
DIR_LIMIT_OFFSET: 0.0
NUM_DIR_BINS: 2
# 在这里加入预测类别的框
ANCHOR_GENERATOR_CONFIG: [ # prediction anchor generation
{
'class_name': 'Car',
'anchor_sizes': [[3.9, 1.6, 1.56]],
'anchor_rotations': [0, 1.57],
'anchor_bottom_heights': [-1.78],
'align_center': False,
'feature_map_stride': 8,
'matched_threshold': 0.6,
'unmatched_threshold': 0.45
},
{
'class_name': 'Pedestrian',
'anchor_sizes': [[0.8, 0.6, 1.73]],
'anchor_rotations': [0, 1.57],
'anchor_bottom_heights': [-0.6],
'align_center': False,
'feature_map_stride': 8,
'matched_threshold': 0.5,
'unmatched_threshold': 0.35
},
{
'class_name': 'Cyclist',
'anchor_sizes': [[1.76, 0.6, 1.73]],
'anchor_rotations': [0, 1.57],
'anchor_bottom_heights': [-0.6],
'align_center': False,
'feature_map_stride': 8,
'matched_threshold': 0.5,
'unmatched_threshold': 0.35
}
]
加入后就没啥问题了：
2024-04-22 20:20:45,398 INFO Car AP@0.70, 0.70, 0.70:
bbox AP:99.0242, 89.8920, 89.7941
bev AP:90.5798, 88.8974, 88.8224
3d AP:90.2740, 87.3502, 87.2354
aos AP:98.71, 89.46, 89.26
Car AP_R40@0.70, 0.70, 0.70:
bbox AP:99.4942, 95.7422, 95.6544
bev AP:96.6726, 92.3975, 92.3774
3d AP:96.1565, 88.9776, 88.9854
aos AP:99.16, 95.20, 94.96
Car AP@0.70, 0.50, 0.50:
bbox AP:99.0242, 89.8920, 89.7941
bev AP:98.9991, 97.3256, 89.7406
3d AP:98.9965, 89.8333, 89.7366
aos AP:98.71, 89.46, 89.26
Car AP_R40@0.70, 0.50, 0.50:
bbox AP:99.4942, 95.7422, 95.6544
bev AP:99.4721, 97.7615, 95.6498
3d AP:99.4714, 95.6845, 95.6274
aos AP:99.16, 95.20, 94.96
Pedestrian AP@0.50, 0.50, 0.50:
bbox AP:59.1360, 50.5516, 42.3628
bev AP:59.0754, 50.3845, 42.3691
3d AP:58.5248, 49.8588, 42.0124
aos AP:32.92, 28.50, 24.50
Pedestrian AP_R40@0.50, 0.50, 0.50:
bbox AP:61.8966, 50.3211, 43.5737
bev AP:61.6259, 50.0281, 43.1782
3d AP:60.9391, 49.4208, 42.6651
aos AP:30.05, 23.97, 20.56
Pedestrian AP@0.50, 0.25, 0.25:
bbox AP:59.1360, 50.5516, 42.3628
bev AP:68.7782, 59.6901, 51.3097
3d AP:68.7782, 59.6901, 51.3086
aos AP:32.92, 28.50, 24.50
Pedestrian AP_R40@0.50, 0.25, 0.25:
bbox AP:61.8966, 50.3211, 43.5737
bev AP:68.3972, 56.2436, 46.9825
3d AP:68.3972, 56.2436, 46.9822
aos AP:30.05, 23.97, 20.56
Cyclist AP@0.50, 0.50, 0.50:
bbox AP:60.3033, 42.8206, 34.7173
bev AP:59.6499, 41.9951, 34.3846
3d AP:57.8227, 41.0860, 33.6334
aos AP:29.64, 22.72, 19.19
Cyclist AP_R40@0.50, 0.50, 0.50:
bbox AP:61.2353, 37.5554, 35.1295
bev AP:58.4128, 36.8041, 34.4358
3d AP:56.3943, 35.5600, 31.3543
aos AP:26.01, 15.93, 14.79
Cyclist AP@0.50, 0.25, 0.25:
bbox AP:60.3033, 42.8206, 34.7173
bev AP:59.7645, 42.0996, 34.3846
3d AP:59.7645, 42.0996, 34.3846
aos AP:29.64, 22.72, 19.19
Cyclist AP_R40@0.50, 0.25, 0.25:
bbox AP:61.2353, 37.5554, 35.1295
bev AP:60.5544, 36.9441, 34.5113
3d AP:60.5544, 36.9441, 34.5113
aos AP:26.01, 15.93, 14.79三、 注意
还有一点值得注意的是，yaml文件还有一个地方提及了三个类别，也就是gt-sampling：
AUG_CONFIG_LIST:
- NAME: gt_sampling
AUG_WITH_IMAGE: True # use PC-Image Aug
JOINT_SAMPLE: True # joint sample with point
KEEP_RAW: False # keep original PC
POINT_REFINE: True # refine points with different calib
BOX_IOU_THRES: 0.5
IMG_AUG_TYPE: by_depth
AUG_USE_TYPE: annotation
IMG_ROOT_PATH: semi/image_2
USE_ROAD_PLANE: False
DB_INFO_PATH:
- kitti_dbinfos_trainsemi.pkl
PREPARE: {
filter_by_min_points: ['Car:5', 'Pedestrian:5', 'Cyclist:5'],
filter_by_difficulty: [-1],
}
# 这是gt-sampling增强用到的参数
SAMPLE_GROUPS: ['Car:15', 'Pedestrian:15', 'Cyclist:15']
NUM_POINT_FEATURES: 8
DATABASE_WITH_FAKELIDAR: False
REMOVE_EXTRA_WIDTH: [0.0, 0.0, 0]
LIMIT_WHOLE_SCENE: False
- NAME: random_world_rotation
WORLD_ROT_ANGLE: [-0.78539816, 0.78539816]
- NAME: random_world_flip
ALONG_AXIS_LIST: ['x']
- NAME: random_world_scaling
WORLD_SCALE_RANGE: [0.95, 1.05]
参数的意思也很简单，就是从别的帧里采样多少个目标过来，Car:15就是采样十五个汽车。实际上，这些与我们上述的更改与预测什么类别并没有太大联系，只能说一定程度上会影响预测精度。