p0l0satik / SphericalProjectionPlaneSegmentator

Dataloader for plane segmentation pipeline with spherical projection

SphericalProjection

Plane segmentation pipeline with spherical projection

Data preparation

Downloading dataset

Download Semantic Kitti dataset. You will need sequence 00.

Download EVOPS plane labels for Semantic Kitti dataset.

Put velodyne and labels as well as evops labels renamed to plane_labels into one directory. In the end the directory with data should look like this.

├── labels
│   └── 000000.label
├── plane_labels
│   └── label-000000.npy
└── velodyne
    └── 000000.bin

You can find an example in examples/data/kitti/00.

Preparing dataset

The next step is running data preprocessing script. This can be done with the preprocessing.py script. It takes the following parameters:

Required

• --dataset path to your dataset folder
• --dataset_len length of your dataset
• --save_path path to store ready to use data

Optional

By default, preprocessing applies ransac filtering and considers road labels as planes. However, you can change that.

• --no_ransac turns off ransac filtering
• --no_road road labels will not be considered as planes
• --visualise_ransac ransac filtering steps will be visualised

An example of preprocessing script run, applied to the data in examples directory:

python3 preprocessing.py --dataset examples/data/kitti/00/ --save_path examples/data/dataset_ready/ --dataset_len 1

Training network

A train run is configured with a config file in a .yaml format. You can find an example in runs_configs directory. If you wish to train the network with your own parameters create a copy of an existing run and change the parameters you like.

This project uses wandb for logging data during train. In order to use it too you need:

1. Log into your wandb account
2. Set use_wandb to True in config
3. Fill in wandb_project with your project name In case you don't want to use wandb train pipline will log with the help of console.

In order to run train successfully you also need to fill the following variables in config:

• dataset_dir - the path to your prepared dataset
• checkpoint_save_path - the path where to store your network checkpoints

After the run is configured you can run train with run_network.py. This script takes only one required parameter - path to your config file. Here is an example:

python3 run_network --config runs_configs/exp1_basic.yaml

Visualising results

The results can be visualized with visualise_prediction.py. It takes the following parameters.

Required

• --chpt path to your network checkpoint
• --dataset path to the dataset you used to train your network
• --config path to the config you used for network training

Optional

• --device by default the network prediction is run on cpu for visualizing results on any device. You can change this.
• --step you can skip similar frames of the dataset configuring step. Default value is 5.
• --iterations number of samples you wish to visualise. Default is 1.

An example of visualising script run will produce 3 visualisations (15 samples with step 5) :

python3 visualise_prediction.py --chpt exp_2_basic_unet_ransac_14 --dataset dataset/kitti/prep/ransac_01 --iterations 15