[모각코] 물랑룽즈 - 2회차

[모각코] 물랑룽즈 - 2회차

 

JetSeg: Efficient Real-Time Semantic Segmentation Model For Low-Power GPU-Embedded Systems https://arxiv.org/pdf/2305.11419.pdf / https://github.com/mmontielpz/jetseg

GitHub - mmontielpz/jetseg: NeurIPS 2023

 

2회차에는 퍼듀 프로젝트에 필요한 JetSeg 논문을 읽고 리뷰하였습니다.

 

JetSeg -

encoder: JetNet → Jetblock (light weight efficient block) reduces the # of parameters/ asymmetric, non-asymmetric convolution + depthwise-dilated convolution(JetConv)

decoder : RegSeg decoder

JetSeg에서 말하는 related work

→

Real-time segmentation에서는 말이죠?~

ENet → SegNet의 1%의 parameter ( residual layer를 bottleneck 모듈에 설치)

다른 제안들 → Pyramid feature extraction

FPENet: efficient feature pyramid for multi-scale context encoding

ESPNet: efficient spatial pyramid dilated convolutions

DFANet: mixes multibranch and spatial pyramid pooling + reduce complexity by reusing features enhancing feature representation

⇒ efficiency를 위해서 accuracy를 희생.

ADSCNet → 봐줄만한 trade-off를 하고자 노력

lightweight asymmetric depthwise separable convolution network that connects sets of dilated convolutional layers using dense dilated convolution connections

• mIoU of 67.5% / Cityscapes dataset at 76.9FPS 21.1 MFLOPS / 20K parameters

contribution

→ JetSeg → Encoder(JetNet), Block(JetBlocK), JetConv, JetLoss

→ JetConv → 공간정보 enhance, receptive field 늘리기, long-range dependencies 늘리기, dilated asymmetric & non-asymmetric convolution을 통해 spatial symmetry와 균형있는 특성 extraction 가능하게 함

→ JetNet - speed를 줄이지 않고 정보 extract 가능, JetConvs로 구성되어있음, dilated asymmetric depthwise conv와 non-asymmetric conv를 동시에 사용해서 # of parameter를 증가시키지 않고 더 많은 양의 정보를 얻을 수 있었음

→ JetBlock - efficient unit(feature map extraction), lightweight activation(TanhExp)

→ JetLoss - precision, recall, IoUB loss의 합침

JetSeg 모델 구조

• S0(residual layer 있음), S1 ⇒ high-level feature extraction
  1. Channel shuffle - reduce # of parameters
  2. JetConv - capture low and high-level features wout increasing model complexity
  3. CBAM - learn discriminative and robust features
• S2 residual layer 존재, CBAM → SAM(Spatial Attention Module)
  1. SAM → long-range dependencies, enhance spatial awareness
• S3 → s2의 절반만큼의 residual layer + efficient channel attention module(관련있는 채널간의 정보를 강조)