0
1

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?

More than 3 years have passed since last update.

【3D OD LiDAR編】SECOND : Sparse Convolutionで高速化

Last updated at Posted at 2020-10-03

SECOND: Sparsely Embedded Convolutional Detection

![image.png](https://qiita-image-store.s3.ap-northeast-1.amazonaws.com/0/482094/ca557f95-571e-1edc-0353-b70b8d49d8cd.png)

VoxelNetとほとんど同じなSECOND。
違う点だけを見ていきたいと思う。

  1. Feature Learning Network => Voxel Feature and Coordinate + Voxel Feature Extractor
  2. Convolutional Middle Layers => Sparse Conv Layer
  3. Region Proposal Network => Region Proposal Network

新規性

Sparse Convolutional Middle Extractor

VoxelNetではConvolutional Middle Layersと呼ばれていた、3Dを2Dに畳み込む処理。 Sparse Convolutionを採用する事で高速化。

*Sparse Convolutionの詳細は別記事で追記

Region Proposal Network

![image.png](https://qiita-image-store.s3.ap-northeast-1.amazonaws.com/0/482094/e6076400-6140-bdf9-41c1-77c647145461.png)

2D Object Detectionの話だから軽く流すが、Multi ResolutionでLocalとGlobalの特徴量を得られる構成にした。

Sine-Error Loss for Angle Regression

image.png
角度のlossを計算する時0とPIで同じ角度を示すのに、値としてジャンプしてしまうのでsinやcosで解決しましょうっていう、よくあるやつ。

結果

![image.png](https://qiita-image-store.s3.ap-northeast-1.amazonaws.com/0/482094/8678b078-9470-aa3a-2815-d366973513b0.png) スピ-ドはSparse Convolutionalを3D=>2Dの畳み込みに採用し向上 精度は角度のlossにsinを使ったことと、2DのDetectionをMulti Resolutionにしたことによって向上 って感じかな〜。

結論

・VoxelNetを踏襲しつつ、精度も速度もシンプルな変更で上げていた。

参考文献

SECOND: Sparsely Embedded Convolutional Detection https://pdfs.semanticscholar.org/5125/a16039cabc6320c908a4764f32596e018ad3.pdf
0
1
0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
0
1

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?