liuxinren456852/dcpcr

DCPCR: Deep Compressed Point Cloud Registration in Large-Scale Outdoor Environments

DCPCR

DCPCR: Deep Compressed Point Cloud Registration in Large-Scale Outdoor Environments

How to get started (with docker)

Install nvida-docker.

Data

You can download the compressed apollo dataset from here and link the dataset to the docker container by configuring the Makefile /dcpcr/Makefile

DATA=<path-to-your-data>

For visualization and finetuning on the uncompressed data, you first have to download the apollo data and use the script '/dcpcr/scripts/apollo_aggregation.py' to compute the dense point clouds. This requires around 500 GB, see compression is nice ;) You can also visualize the registration on the compressed data, but it's hard to see stuff, due to the low resolution.

Building the docker container

For building the Docker Container simply run

make build

in the root directory.

Running the Code

The first step is to run the docker container inside dcpcr/:

make run

The following commands assume to be run inside the docker container.

Training

For training a network we first have to create the config file with all the parameters. An example of this can be found in /dcpcr/config/config.yaml. To train the network simply run

python3 trainer -c <path-to-your-config>

Evaluation

Evaluating the network on the test set can be done by:

python3 test.py -ckpt <path-to-your-checkpoint>

All results will be saved in a dictonary in the dcpcr/experiments. When finetuning with the compressed data we used -dt 1, while -dt 5 for the uncompressed.

Qualitative results

In dcpcr/scripts/qualitative are some scripts to visualize the results.

Pretrained models

The pretrained weights of our models can be found here

How to get started (without Docker)

Installation

A list of all dependencies and install instructions can be derived from the Dockerfile. Use pip3 install -e . to install dcpcr.

Running the code

The scripts can be run as before inside the docker container. Only the dcpcr/config/data_config.yaml might need to be updated.