search

traveller59 / second.pytorch

SECOND for KITTI/NuScenes object detection
MIT License
1.73k stars 722 forks source link

what is the config file used for pretrained model pp_model_for_nuscenes_pretrain/voxelnet-296960.tckpt? #198

Open chowkamlee81 opened 5 years ago

chowkamlee81 commented 5 years ago

Hi,

The pretrained model pp_model_for_nuscenes_pretrain/voxelnet-296960.tckpt, downloaded.

What is the configuration file used to develop this? Kindly help

cvnerds commented 5 years ago

maybe configs/pointpillars/pp_pretrain.config

zovathio commented 5 years ago

could you find out? for configs/pointpillars/pp_pretrain.config I got a dimension mismatch error

HanFa commented 5 years ago

I also get the shape mismatch errors for RPN with config file configs/poiintpillars/pp_pretrain.config and all config files in folder configs/nuscenes. Could anyone provide the proper config?

yinjunbo commented 5 years ago

I'd like to know which config file yield the best results on nuScenes. Does anyone figure it out?

lawrencekiba commented 5 years ago

I have this dimension mismatch issue as well.

AhmedMagdyHendawy commented 5 years ago

Anyone knows which config file?

ICS-India commented 4 years ago

@AhmedMagdyHendawy @chowkamlee81 @traveller59

in all.pp.largea.config/all.pp.mida.config/all.pp.lowa.config if you do the following modification you will be able to load the network
1) In RPN change the upsample stride to this "upsample_strides : [1, 2, 4]" 2) In target assigner keep only one class ("I kept car alone ")

Then you should be able to load the network and evaluate the samples . But i got zero accuracy with NUSCENES mini dataset . Let me know if anybody able to get reasonable accuracy this pretrained network . original config may be different ?

yuanzheng-s commented 4 years ago

@ICS-India hi, i follow the way of your modification,and build_network successful! but the inference result in the kittiviewer web is strange, the predict bbox have large difference with the ground truth,have you met this before