This repository contains code to learn a model for semantic segmentation of 3D LiDAR scans
This software is released under GPLv3. If you use it in academic work, please cite:
@article{dewan-deeptemporalseg,
author = {Ayush Dewan and Wolfram Burgard},
title = {DeepTemporalSeg: Temporally Consistent Semantic Segmentation of 3D LiDAR Scans},
booktitle = {https://arxiv.org/abs/1906.06962},
year = {2019},
url = {http://deep-temporal-seg.informatik.uni-freiburg.de/dewan_deep_temporal_seg.pdf}
}
- Tensorflow
- Pyhton 3.6
./download_dataset.sh
This will download the following datasets:
- tfrecord files for the dataset from https://github.com/BichenWuUCB/SqueezeSeg
- tfrecords files for our dataset generated from the KITTI tracking benchmark. The details regarding the dataset are described in the paper.
All the files required for training and testing the model is in python_scripts folder. To train the model following script has to be executed.
train_seg.py
Parameters
--model_name (default: lidar_segmentation)
--train_record_filename
--validation_record_filename
--log_dir
--path_to_store_models (default: learned_models/)
--learning_rate (default: 0.0001)
--eta (default: 0.0005)
--total_epochs (default: 200)
--batch_size (default: 2)
--image_height (default: 64)
--image_width (default: 324)
--num_channels (default: 5)
--num_classes (default: 4)
--growth (default: 16)
--gpu (default: 0)
python train_seg.py --model_name lidar_segmentation --train_record_filename ../datasets/squeeze_seg_train/ --validation_record_filename ../datasets/squeeze_seg_validation/ --image_width 512 --batch_size 2
./download_models.sh
This will download the models trained on the dataset from https://github.com/BichenWuUCB/SqueezeSeg and KITTI tracking benchmark
test.py
Parameters
--model_name
--validation_record_filename
--is_visualize (default: no)
--image_width (default: 512)
--gpu (default: 0)
python test.py --model_name ../models/squeeze_seg --validation_record_filename ../datasets/squeeze_seg_validation/squeeze_seg_validation.records --is_visualize yes
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent