CenterPoint Implementation

How to setup CenterPoint

First clone the repo (preferably to your persistent folder git clone https://github.com/asmoly/CenterPoint_Nautilus.git

Then go into the repo root folder and run source install_conda.sh if you don’t already have conda installed. Make sure to restart your terminal after this.

Next, from the repo root folder run source setup.sh, this will install all the necessary dependencies for OpenPCDet.

Then run, python scripts/make_test_data.py, verify that the custom_datasetappeared in the data/ directory. This python file is a helpful file to look at to understand the dataset format. This program generates the random dataset. It creates random point clouds and then inserts object scatters into the point clouds.

# For future reference this is the format the the custom dataset should be

data/custom/
  ImageSets/
    train.txt
    val.txt
  points/
    000001.npy
    000002.npy
  labels/
    000001.txt
    000002.txt

# The points files are 2D numpy arrays with each element being [x, y, z, intensity]
# The labels are text files with each line being: x y z dx dy dz heading_angle class_name
# The image sets are just text files with the ids of the point clouds you want for the train and val set

# These are some other notable config files neccessary for the custom dataset
tools\cfgs\dataset_configs\custom_dataset.yaml
tools\cfgs\custom_models\centerpoint_custom.yaml

Run python -m pcdet.datasets.custom.custom_dataset create_custom_infos tools/cfgs/dataset_configs/custom_dataset.yaml from the repo root folder. This sets up the custom dataset for training.

Then run this from the tools/ directory

python train.py \
  --cfg_file cfgs/custom_models/centerpoint_custom.yaml \
  --batch_size 1 \
  --epochs 2 \
  --workers 2

Here you can specify the training parameters.

Checkpoints should be stored here: CenterPoint_Nautilus/output/custom_models/centerpoint_custom/default/ckpt/

You can test a checkpoint using:

python [demo.py](http://demo.py/) \
	--cfg_file cfgs/custom_models/centerpoint_custom.yaml \
	--ckpt ../output/custom_models/centerpoint_custom/default/ckpt/checkpoint_epoch_2.pth \
	--data_path ../data/custom/points/001000.npy \
	--ext .npy

You should see something like this pop up (This was a mostly untrained model so the output is pretty random but the trained models look much better):