the result is bad that I train the fcos3d with nuscenes v1.0-mini

[Will-yh]

Hello mmdetection3d team,

I want to train fcos3d with nuscenes dataset, while the whole datasets is really huge, sometimes it is hard for me to download all the data. Therefore, I train the fcos3d with nuscenes v1.0-mini data. The following is the steps:

mmdetection3d version: release v0.17.1

1. I downloaded the data from nuscenes website:
   

2. unzip the v1.0-mini.tgz

3. modify the v1.0-mini inside v1.0-mini to v1.0-trainval

4. symbolic link the datasets: ln -s v1.0-mini mmdetection3d/data

5. modify tools/create_data.py
   elif args.dataset == 'nuscenes' and args.version != 'v1.0-mini':
   train_version = f'{args.version}-trainval'
   nuscenes_data_prep(
   root_path=args.root_path,
   info_prefix=args.extra_tag,
   version=train_version,
   dataset_name='NuScenesDataset',
   out_dir=args.out_dir,
   max_sweeps=args.max_sweeps)
   #test_version = f'{args.version}-test'
   #nuscenes_data_prep(
   # root_path=args.root_path,
   # info_prefix=args.extra_tag,
   # version=test_version,
   # dataset_name='NuScenesDataset',
   # out_dir=args.out_dir,
   # max_sweeps=args.max_sweeps)

6. Create the date with: python tools/create_data.py nuscenes --root-path ./data/nuscenes --out-dir ./data/nuscenes --extra-tag nuscenes --workers 20

7. the created folder as follow:
   

8. the trained model:
   

9. test the model: python demo/mono_det_demo.py demo/data/nuscenes/n015-2018-07-24-11-22-45+0800__CAM_BACK__1532402927637525.jpg demo/data/nuscenes/n015-2018-07-24-11-22-45+0800__CAM_BACK__1532402927637525_mono3d.coco.json configs/fcos3d/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d_finetune.py work_dirs/fcos3d_r101_caffe_fpn_gn-head_dcn_2x8_1x_nus-mono3d/latest.pth

   The result shows as below:
   
   Compare with the result of the demo, this result is really bad.

I'm wondering if the bad result is caused by the training datasets?

[Tai-Wang]

There is only one minor problem in your implementation: you can set --version to v1.0-mini to generate infos only for that mini-data version, but I guess this point won't affect your results.

I did not try to train the model with the mini version, but I guess it is possible that the limited data results in such bad performance. Someone has tried with 1/10 full dataset before and there was also a serious performance degradation. If you are interested in training models with a small dataset, you can try with KITTI, and the PRs #964 and #1014 for FCOS3D++ are ready to be merged into v1.0.0.dev0. The only difference is the evaluation metric is different such that the finally expected results can be a little different (KITTI expects only the most accurate ones while nuScenes expects more comprehensive predictions.)

[Tai-Wang]

Close due to inactivity.