OpenPCDet my CenterPoint config for KITTI dataset

This config file is modified from both official code released for KITTI and OpenPCDet version for WAYMO. And if anyone wants to get a pre-training model, please let me know. Update:

2022-06-30: I've uploaded the pre-trained model centerpoint@KITTI to the Google Drive. It is available at Performance and Models section in the repo

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

DATA_CONFIG:
    _BASE_CONFIG_: cfgs/dataset_configs/kitti_dataset.yaml

MODEL:
    NAME: CenterPoint

    VFE:
        NAME: MeanVFE

    BACKBONE_3D:
        NAME: VoxelResBackBone8x

    MAP_TO_BEV:
        NAME: HeightCompression
        NUM_BEV_FEATURES: 256

    BACKBONE_2D:
        NAME: BaseBEVBackbone

        LAYER_NUMS: [5]
        LAYER_STRIDES: [1]
        NUM_FILTERS: [128]
        UPSAMPLE_STRIDES: [2]
        NUM_UPSAMPLE_FILTERS: [256]

    DENSE_HEAD:
        NAME: CenterHead
        CLASS_AGNOSTIC: False

        CLASS_NAMES_EACH_HEAD: [
            [ 'Car', 'Pedestrian', 'Cyclist' ]
        ]

        SHARED_CONV_CHANNEL: 64
        USE_BIAS_BEFORE_NORM: True
        NUM_HM_CONV: 2 #  heatmap
        SEPARATE_HEAD_CFG:
            HEAD_ORDER: ['center', 'center_z', 'dim', 'rot']
            HEAD_DICT: {
                'center': {'out_channels': 2, 'num_conv': 2}, # offset
                '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: 4
            NUM_MAX_OBJS: 500
            GAUSSIAN_OVERLAP: 0.1
            MIN_RADIUS: 2

        LOSS_CONFIG:
            LOSS_WEIGHTS: {
                'cls_weight': 1.0,
                'loc_weight': 2.0,
                'code_weights': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
            }
        POST_PROCESSING:
            SCORE_THRESH: 0.1
            POST_CENTER_LIMIT_RANGE: [-75.2, -75.2, -2, 75.2, 75.2, 4]
            MAX_OBJ_PER_SAMPLE: 500
            NMS_CONFIG:
                NMS_TYPE: nms_gpu
                NMS_THRESH: 0.7
                NMS_PRE_MAXSIZE: 4096
                NMS_POST_MAXSIZE: 500

    POST_PROCESSING:
        RECALL_THRESH_LIST: [0.3, 0.5, 0.7]
        SCORE_THRESH: 0.1
        OUTPUT_RAW_SCORE: False

        EVAL_METRIC: kitti

        NMS_CONFIG:
            MULTI_CLASSES_NMS: False
            NMS_TYPE: nms_gpu
            NMS_THRESH: 0.01
            NMS_PRE_MAXSIZE: 4096
            NMS_POST_MAXSIZE: 500


OPTIMIZATION:
    BATCH_SIZE_PER_GPU: 4
    NUM_EPOCHS: 80

    OPTIMIZER: adam_onecycle
    LR: 0.003
    WEIGHT_DECAY: 0.01
    MOMENTUM: 0.9

    MOMS: [0.95, 0.85]
    PCT_START: 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

Jan 13 '22 12:01 OuyangJunyuan

Can you show the result on kitti val set?

Jan 14 '22 07:01 EmiyaNing

the result get from tools/test.py are as follwing:

Car [email protected], 0.70, 0.70:
bbox AP:92.4371, 89.7997, 89.1875
bev  AP:89.0399, 87.5611, 86.4736
3d   AP:87.3393, 79.4999, 77.3796
aos  AP:92.42, 89.72, 89.07
Car [email protected], 0.70, 0.70:
bbox AP:97.1744, 93.6886, 92.1958
bev  AP:92.7800, 89.0816, 88.0308
3d   AP:88.9477, 81.6019, 79.1999
aos  AP:97.15, 93.59, 92.05
Car [email protected], 0.50, 0.50:
bbox AP:92.4371, 89.7997, 89.1875
bev  AP:92.3280, 91.4396, 89.4304
3d   AP:92.3222, 89.8208, 89.3696
aos  AP:92.42, 89.72, 89.07
Car [email protected], 0.50, 0.50:
bbox AP:97.1744, 93.6886, 92.1958
bev  AP:97.1386, 94.7012, 93.3835
3d   AP:97.1264, 93.9522, 93.1182
aos  AP:97.15, 93.59, 92.05
Pedestrian [email protected], 0.50, 0.50:
bbox AP:74.4678, 71.9525, 69.1918
bev  AP:60.7042, 59.1109, 55.1364
3d   AP:57.3194, 54.2235, 50.6287
aos  AP:72.65, 69.44, 66.21
Pedestrian [email protected], 0.50, 0.50:
bbox AP:75.1197, 72.9895, 69.8245
bev  AP:60.5372, 58.2984, 54.0787
3d   AP:55.7938, 53.0594, 48.9036
aos  AP:73.24, 70.18, 66.46
Pedestrian [email protected], 0.25, 0.25:
bbox AP:74.4678, 71.9525, 69.1918
bev  AP:78.9672, 78.2260, 75.2896
3d   AP:78.8915, 77.9085, 75.0464
aos  AP:72.65, 69.44, 66.21
Pedestrian [email protected], 0.25, 0.25:
bbox AP:75.1197, 72.9895, 69.8245
bev  AP:81.4859, 80.1064, 76.6350
3d   AP:81.3868, 79.6834, 76.2700
aos  AP:73.24, 70.18, 66.46
Cyclist [email protected], 0.50, 0.50:
bbox AP:87.5590, 75.4489, 71.9498
bev  AP:88.6395, 68.9562, 65.1039
3d   AP:85.7151, 65.6614, 62.1453
aos  AP:87.46, 74.95, 71.55
Cyclist [email protected], 0.50, 0.50:
bbox AP:91.6160, 76.7370, 72.7161
bev  AP:89.0516, 69.3149, 65.6969
3d   AP:85.7833, 65.5833, 62.2746
aos  AP:91.49, 76.25, 72.24
Cyclist [email protected], 0.25, 0.25:
bbox AP:87.5590, 75.4489, 71.9498
bev  AP:87.8537, 72.0860, 68.6397
3d   AP:87.8537, 72.0850, 68.6397
aos  AP:87.46, 74.95, 71.55
Cyclist [email protected], 0.25, 0.25:
bbox AP:91.6160, 76.7370, 72.7161
bev  AP:91.3929, 73.1575, 69.0485
3d   AP:91.3929, 73.1573, 69.0482
aos  AP:91.49, 76.25, 72.24

Jan 15 '22 15:01 OuyangJunyuan

Would you please provide centerpoint-pointpillars config?

Feb 16 '22 12:02 YoushaaMurhij

Would you please provide centerpoint-pointpillars config?

it's sry that i didn't try. But it's easy to implement, just using pillar as VFE and checking that BACKBONE_2D feature map size equal to Voxel Grid Size_XY / CenterHead stride

Feb 17 '22 02:02 OuyangJunyuan

@OuyangJunyuan , Thanks for your comment! I have implemented it.

Feb 17 '22 08:02 YoushaaMurhij

It's seems that I can't reproduce your result in my enviroment( RTX 3090 + cuda11.0 + spconv1.2.1 + OpenPCDet v0.5.0). I'm puzzed that the POST_CENTER_LIMIT_RANGE is not suit for KITTI in your configure file.

Mar 11 '22 02:03 EmiyaNing

I need a pre-training mode, can you give me?

Mar 25 '22 03:03 hhhmrcscs

Hi, can you share your pretrained model? Thanks [email protected]

Mar 28 '22 02:03 angryhen

I need a pre-training mode, can you give me?

Hi, can you share your pretrained model? Thanks [email protected]

has been sent~

Mar 28 '22 03:03 OuyangJunyuan

I need a pre-training mode, can you give me?

any email will be provided？

Mar 28 '22 03:03 OuyangJunyuan

[email protected], Thanks

Mar 28 '22 04:03 hhhmrcscs

I need a pre-training mode, can you give me?

Hi, can you share your pretrained model? Thanks [email protected]

has been sent~

Hi, I used your pretrain-model to test kitti testing velodyne bin, but got error result. python demo.py --cfg_file cfgs/kitti_models/centerpoints.yaml --ckpt ../checkpoints/centerpoint_kitti_80.pth --data_path ../testing/velodyne/000003.bin

Mar 28 '22 07:03 angryhen

@OuyangJunyuan , Thanks for your comment! I have implemented it.

Hi, can you share your yaml file? Thanks [email protected]

Apr 19 '22 02:04 HuangCongQing

Aha, the CenterPoint's yaml file I used in kitti just like aforemention. You can try it, and train&test in normal way.

Apr 23 '22 09:04 OuyangJunyuan

I need a pre-training mode, can you give me?

Hi, can you share your pretrained model? Thanks [email protected]

has been sent~

Hi, I used your pretrain-model to test kitti testing velodyne bin, but got error result. python demo.py --cfg_file cfgs/kitti_models/centerpoints.yaml --ckpt ../checkpoints/centerpoint_kitti_80.pth --data_path ../testing/velodyne/000003.bin

you can set NMS_THRESH to 0.01 in MODEL.DENSE_HEAD.POST_PROCESSING config. Centerpoint model do nms in centerhead, so the config of MODEL.POST_PROCESSING is useless.

Apr 23 '22 14:04 Rapisurazurite

This issue is stale because it has been open for 30 days with no activity.

May 24 '22 02:05 github-actions[bot]

Hi, can you share your pretrained model? Thanks a lot! [email protected]

May 31 '22 13:05 YS-Kuang

@OuyangJunyuan 你好，可以分享一下预训练模型吗 [email protected]

Jun 29 '22 16:06 ghb0224

@OuyangJunyuan 你好，可以分享一下预训练模型吗 [email protected]

Hi, can you share your pretrained model? Thanks a lot! [email protected]

sry for replying late. the pretrain model can obttain is available here

Jun 30 '22 01:06 OuyangJunyuan

font{
    line-height: 1.6;
}
ul,ol{
    padding-left: 20px;
    list-style-position: inside;
}




    Sorry,I have not received.
    

    
        
    


 

    
                ***@***.***
    




---- Replied Message ----



  
    
     From 
    
    
        ***@***.***>
        
    
  
  
    
     Date 
    
    
    6/30/2022 09:25
    
  
  
    
     To 
    
    
     
      
        ***@***.***>
        
      
    
  
  
    
     Cc 
    
    
      
        ***@***.***>
        ,
      
      
        ***@***.***>
        
      
    
  
  
    
     Subject 
    
    
          Re: [open-mmlab/OpenPCDet] my CenterPoint config for KITTI dataset (Issue #750)

@OuyangJunyuan 你好，可以分享一下预训练模型吗 @.***

Hi, can you share your pretrained model? Thanks a lot! @.***

sry for replying late. the pretrain model can obttain is available here

—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you commented.Message ID: @.***>

Jun 30 '22 02:06 ghb0224

Hi, I have a question about 'FEATURE_MAP_STRIDE: 4'. It seems to be 8, since the backbone downsamples the feature map to 1/8. I have read the code of centerhead and config file of that on Waymo, I think 8 should be the right value. However, stride=4 gets better results than 8. And I have also conducted experiments on stride=1 and stride=2, they get nearly 0 mAP. Hope to get your reply.

Jul 16 '22 11:07 xiaoxin83121

Hi, I have a question about 'FEATURE_MAP_STRIDE: 4'. It seems to be 8, since the backbone downsamples the feature map to 1/8. I have read the code of centerhead and config file of that on Waymo, I think 8 should be the right value. However, stride=4 gets better results than 8. And I have also conducted experiments on stride=1 and stride=2, they get nearly 0 mAP. Hope to get your reply.

Indeed, what you thought are approximately right about the Backbone3D. But there is one thing you has ignored, i.e., BEV_Backbone. Please kindly refer to configuration file that you can find UPSAMPLE_STRIDES: [2] in it, resulting in 1/8 * 2 = 1/4 here.

Jul 18 '22 01:07 OuyangJunyuan

This issue is stale because it has been open for 30 days with no activity.

Aug 18 '22 02:08 github-actions[bot]

This issue was closed because it has been inactive for 14 days since being marked as stale.

Sep 02 '22 02:09 github-actions[bot]

@OuyangJunyuan 你好，可以分享一下预训练模型吗 [email protected]

Hi, can you share your pretrained model? Thanks a lot! [email protected]

sry for replying late. the pretrain model can obttain is available here

hello, can you share your pretrained model ? I dont't find it at the issue.

Mar 15 '23 15:03 DezeZhao

@OuyangJunyuan 你好，可以分享一下预训练模型吗 [email protected]

Hi, can you share your pretrained model? Thanks a lot! [email protected]

sry for replying late. the pretrain model can obttain is available here

hello, can you share your pretrained model ? I dont't find it at the issue.

[email protected], thank u.

Mar 15 '23 15:03 DezeZhao

嗨，你能分享一下你的预训练模型吗？谢谢 [email protected]

Hello, can you share the pretrained model you obtained? Thanks [email protected]

Sep 07 '23 07:09 HezyC

@OuyangJunyuan , May I ask how do you achieve it? I tried to achieve it, but the loss has been kept around 8, and all AP are 0.

Jan 19 '24 11:01 lifan67