I can't train the multi-modal accuracy(KITTI val split in AP3D(R11))you mentioned in the paper

[qimingx]

my val result: epochs:90 3d AP:89.3103, 85.0520, 79.2089

your val result in Table 8 Focal Conv-F 3d: 89.82 85.22 85.19

I wonder why can't improve accuracy in hard situation and how many epochs you train?

[yukang2017]

Thanks for you interest in our work. It is also trained for 80 epochs. The accuracy in hard situation of KITTI is really unstable. We only get this one model for such a high result for hard situation of KITTI by chance. This is the training log for this result (epoch 70).

log_train_20211106-190258.txt

[qimingx]

Thanks for you interest in our work. It is also trained for 80 epochs. The accuracy in hard situation of KITTI is really unstable. We only get this one model for such a high result for hard situation of KITTI by chance. This is the training log for this result (epoch 70).

log_train_20211106-190258.txt

Thanks for answer. And I have another question that is how to amend voxel_rcnn_car.yaml? I replaced VoxelBackBone8x with VoxelBackBone8xFocal, but got a reduced accuracy compared with voxel_rcnn.

[yukang2017]

For multi-modal setting, you also need modify the data config part in the config yaml file. Because it involves multi-modal data augmentations. Please refer to this part https://github.com/dvlab-research/FocalsConv/blob/2fbfdda0179d7a29e7a304b4ac9d67e866ee05d9/OpenPCDet/tools/cfgs/kitti_models/voxel_rcnn_car_focal_multimodal.yaml#L3

[qimingx]

For multi-modal setting, you also need modify the data config part in the config yaml file. Because it involves multi-modal data augmentations. Please refer to this part

https://github.com/dvlab-research/FocalsConv/blob/2fbfdda0179d7a29e7a304b4ac9d67e866ee05d9/OpenPCDet/tools/cfgs/kitti_models/voxel_rcnn_car_focal_multimodal.yaml#L3

I mean single-modal setting. I replaced VoxelBackBone8x with VoxelBackBone8xFocal, but got a reduced accuracy compared with voxel_rcnn.

[yukang2017]

Is the VoxelBackBone8xFocal from the this config?

https://github.com/dvlab-research/FocalsConv/blob/2fbfdda0179d7a29e7a304b4ac9d67e866ee05d9/OpenPCDet/tools/cfgs/kitti_models/pv_rcnn_focal_lidar.yaml#L40

BACKBONE_3D:
    NAME: VoxelBackBone8xFocal
    MASK_MULTI: True
    ENLARGE_VOXEL_CHANNELS: 64

[qimingx]

Is the VoxelBackBone8xFocal from the this config?

https://github.com/dvlab-research/FocalsConv/blob/2fbfdda0179d7a29e7a304b4ac9d67e866ee05d9/OpenPCDet/tools/cfgs/kitti_models/pv_rcnn_focal_lidar.yaml#L40

BACKBONE_3D:
    NAME: VoxelBackBone8xFocal
    MASK_MULTI: True
    ENLARGE_VOXEL_CHANNELS: 64

yes,but baseline is voxel-rcnn

[yukang2017]

It's weird. I will train it by myself and try to fix this latter. Thanks.

In addition, would you please try to set ENLARGE_VOXEL_CHANNELS as 128 ? It would return better results based on my experience.

[qimingx]

It's weird. I will train it by myself and try to fix this latter. Thanks.

In addition, would you please try to set ENLARGE_VOXEL_CHANNELS as 128 ? It would return bette results based on my experience.

ok!

[yukang2017]

Hi, I have updated the code. Please use the below backbone config for lidar-only Voxel R-CNN.

    NAME: VoxelBackBone8xFocal
    USE_IMG: False
    ENLARGE_VOXEL_CHANNELS: 64
    KERNEL_SIZE: 5
    USE_STAGES: [1,]

I have trained for several times. You can get better results by this.

[qimingx]

I have trained for several times. You can get better results by this.

Ok, thank you!

[yukang2017]

Thanks for your patience. I currently close this issue. Feel free to open other issues and any further discussions.