best result on nuscenes

[TianhaoFu]

hi,
i noticed that your rank in nuscenes lidar track is 2

CenterPoint-VID
mAP: 0.674
NDS: 0.718

What config should I use to train the model to achieve this effect?

[tianweiy]

this one is not my entry

[TianhaoFu]

thanks.

Which ranking on the nuscene list is yours now? Can you share the training strategy and related configuration?

@tianweiy

[tianweiy]

the leaderboard uses old version (and worse). The most recent models get 66.8 / 68.5 (single / flip) #249 on val set. The test set should be similar or higher. The configs are already provided.

If you want to get to SOTA performance, you will need fusion based methods (e.g. https://github.com/tianweiy/MVP)

[TianhaoFu]

ok, thanks!

[TianhaoFu]

the leaderboard uses old version (and worse). The most recent models get 66.8 / 68.5 (single / flip) #249 on val set. The test set should be similar or higher. The configs are already provided.
If you want to get to SOTA performance, you will need fusion based methods (e.g. https://github.com/tianweiy/MVP)

Hi, i want to know does the result of 68.5 use dcn? thanks!
@tianweiy

[tianweiy]

No, I don't think dcn helps much

[tianweiy]

I basically just run this config (w/ w/o flip test) with pretrained model https://github.com/tianweiy/CenterPoint/blob/master/configs/mvp/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_scale.py

[TianhaoFu]

I basically just run this config (w/ w/o flip test) with pretrained model https://github.com/tianweiy/CenterPoint/blob/master/configs/mvp/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_scale.py

can you provide the per-trained model? thanks!

[tianweiy]

i just use this model https://github.com/tianweiy/CenterPoint/blob/master/configs/nusc/README.md (first one)

[TianhaoFu]

so you trained totally 40 epoch?
could you provide your training command? like

python -m torch.distributed.launch --nnodes=1 ./tools/train.py configs/mvp/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_scale.py --resume_from epoch_20.pth

thanks!

[tianweiy]

I only trained 20 epochs, all training follows a standard schedule (and already specified in the config).

You can get the number by just testing the pretrained model (which trained for 20 epoch)

[tianweiy]

run config < i mean testing

[TianhaoFu]

got it .
Do you think adding double flip augmentation will make the performance better?
thanks!

[tianweiy]

the 68.5 is with flip aug, 66.8 is without flip aug. Flip definitely helps at the cost of 4x latency

[TianhaoFu]

does the flip mean double flip?
should i add the double_flip=ture during training and testing

[tianweiy]

we always do flip aug during training.

By default, we don't do flip aug during testing.

To enable it, we need to add multiple flip flags in the config. See this file for an example https://github.com/tianweiy/CenterPoint/blob/master/configs/nusc/voxelnet/nusc_centerpoint_voxelnet_0075voxel_dcn_flip.py

[TianhaoFu]

does the flip aug during testing help improve performance?

[tianweiy]

yes

the 68.5 is with flip aug, 66.8 is without flip aug.

[TianhaoFu]

the 68.5 is with flip aug, 66.8 is without flip aug. Flip definitely helps at the cost of 4x latency

could you provide this test config with flip aug?
thanks!

[tianweiy]

will do in a bit (after i finish graphics hw, probably 1 to 2 hrs)

[TianhaoFu]

i noticed that in https://github.com/tianweiy/CenterPoint/blob/master/configs/mvp/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_scale.py config.

there is no voxel_generator config, why?

byw, what the difference between https://github.com/tianweiy/CenterPoint/blob/master/configs/mvp/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_scale.py and https://github.com/tianweiy/CenterPoint/blob/master/configs/nusc/voxelnet/nusc_centerpoint_voxelnet_0075voxel_fix_bn_z.py

[tianweiy]

accuracy they are the same. the one inside mvp folder uses dynamic voxelization, so we can remove the old voxelizer

[tianweiy]

will do in a bit (after i finish graphics hw, probably 1 to 2 hrs)

up now