3DCV Final Project

This project is forked from https://github.com/wayveai/fiery.

Environments

1. git clone this repo and create the necessary directorys

   https://github.com/sty61010/fiery.git
   cd fiery/
   mkdir -p data tensorboard_logs

2. Download NuScenes dataset and place it under the data directory.

3. Create the python environment from bev18.yml

   conda env create -n fiery --file bev18.yml

4. Install mmdetection3d package. For more information, please refer to mmdet3d document.

   git clone https://github.com/open-mmlab/mmdetection3d.git
   cd mmdetection3d
   pip install -v -e .

How to test the model

Download the pretrained weight here and extract them into tensorboard_logs.

In the fiery directory, run

bash test.sh tensorboard_logs/model 20.ckpt

It will run evaluation on the trainval dataset and report the mAP score.

Expected result:

How to train the model

There are many configs in the fiery/config folder, you can train the model by

python train.py --config CONFIG GPUS "[0]" ...[Other Parameters]

To preproduce our pretrained model, run

python train.py --config tensorboard_logs/model/hparams.yml