Skip to content

Chai77/Cylinder3D_Coda

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

61 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

TODO

  • Fix issue in converted network (see below)
  • Test with Cuda 11.6/11.7 support
  • Spconv 2.2 (FP32 disabled by default, see their instructions on how to enable FP32 if necessary.)

Description

This repository is forked from the original implementation. This repositiory was created to encourage continued research and provide a significantly faster implementation.

Spconv-V2 made major changes to the convolution operation and how weights are read. Spconv-V1 is not supported anymore. Following an unsuccessfull effort to restructure the weights, Cylinder3D was retrained on SemanticKITTI to produced new Spconv-v2 weights.

Note: the version released publicly by the authors is from their first paper and does not include the pointwise refinement module or weighted cross entropy. The CVPR version has not been made publically available. The mIOU of the retrained Spconv2 Version is 63.2, compared to 64.3 mIOU in page 7 in the paper. The original implementation does not contain manual seeding. To achieve the 63.2 mIOU result, the training regimen had to be changed slightly, as the Paper results could not be reproduced (See Training).

Improvements

  • Network code updated to Spconv-v2.1.x Credit here goes to the code by @min2209 in Issue.
  • Spconv-v2.x receives continued support. Speedup of 50 - 80% compared to original implementation
  • Mixed precision support for further speedup during training

It is likely that further improvements can be made with a more careful choice of hyperparameters when training.

Issue

Fixed Typo in converted model. Weights are updated.

reaA = resA.replace_feature(self.bn1(resA.features)) should be: resA = resA.replace_feature(self.bn1(resA.features))

Installation

Weights

The weights with mIOU 63.2 (Validation, vs 64.3 Original) can be downloaded here and should be placed into the ./network folder.

Weights are trained according to the original Cylinder3D, and not according to knowledge distillation (Cylinder3D PVKD).

Training

  • 40 epochs
  • 0.00707 base LR with sqrt_k scaling rule (equals to original 0.001 at batchsize = 2, equals 0.00489 at batchsize = 24)
  • AdamW with Weight Decay 0.001
  • CosineDecay Schedule
  • Batch Size 24 (Better result possible with lower batch size, batch size chosen for economical reasons.)

Requirements

Also tested with CUDA 11.3, just "downgrade" pytorch, spconv and torch-scatter.

Tested on Ubuntu 20.04 LTS. Recommend pip install over conda install.

Data Preparation

SemanticKITTI

./
├── 
├── ...
└── path_to_data_shown_in_config/
    ├──sequences
        ├── 00/           
        │   ├── velodyne/	
        |   |	├── 000000.bin
        |   |	├── 000001.bin
        |   |	└── ...
        │   └── labels/ 
        |       ├── 000000.label
        |       ├── 000001.label
        |       └── ...
        ├── 08/ # for validation
        ├── 11/ # 11-21 for testing
        └── 21/
	    └── ...

Training

  1. modify the config/semantickitti.yaml with your custom settings. We provide a sample yaml for SemanticKITTI
  2. train the network by running "sh train.sh"

If you find this work useful in your research, please consider citing the original authors papers:

@article{zhu2020cylindrical,
  title={Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation},
  author={Zhu, Xinge and Zhou, Hui and Wang, Tai and Hong, Fangzhou and Ma, Yuexin and Li, Wei and Li, Hongsheng and Lin, Dahua},
  journal={arXiv preprint arXiv:2011.10033},
  year={2020}
}

About

Cylinder3D fork for use with UT CODa.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published