search

boostcampaitech3 / level2-semantic-segmentation-level2-cv-14

level2-semantic-segmentation-level2-cv-14 created by GitHub Classroom
0 stars 2 forks source link

[팁] 실험 모델 1차 설정 #6

Closed agwmon closed 2 years ago

agwmon commented 2 years ago

개요 및 목적

mmsegmentation에 있는 모델(backbone 제외)들에 대해 간단하게 정리했습니다~ 개인적으로 독특하고, 중요하다고 생각되는 모델들에 중요 표시를 해두었습니다.

이를 바탕으로 고려해보시면 좋을 것 같습니다.

작업 상세 내용

type 특이사항 cityspace +flip, ms 중요
FCN (CVPR'2015/TPAMI'2017) 마지막 단에 Fully connected network head 78.46 80.42 V
ERFNet (T-ITS'2017) Efficient trade-off
UNet (MICCAI'2016/Nat. Methods'2019) with deeplabv3 79.77 81.06 V
PSPNet (CVPR'2017) pyramid pooling module head 79.76 81.01
DeepLabV3 (ArXiv'2017) spatial pyramid pooling head 80.48 81.21 V
BiSeNetV1 (ECCV'2018) real-time(down & upsample) 79.69 80.89
PSANet (ECCV'2018) pointwise attention ->imformation propagation 원활? 79.69 80.89
DeepLabV3+ (CVPR'2018) low level + concat head 80.97 82.03 V
UPerNet (ECCV'2018) 이미지의 컨셉을 이해 -> 여러 head(scene, part …) (sota 모델에서 자주 쓰임) head swin이랑 자주씀 81.49 81.49 V
ICNet (ECCV'2018) real-time & cascade 76.29
NonLocal Net (CVPR'2018) nonlocal means filter -> spatial featue weighted sum 79.4
EncNet (CVPR'2018) context encoding -> feature(Channel) 별 attention 78.55
Semantic FPN (CVPR'2019) 75.8
DANet (CVPR'2019) dual attention - position & channel 80.52 V
APCNet (CVPR'2019) adaptive pyramid context - global image representation 79.64
EMANet (ICCV'2019) attention 연산 효율적으로 79.62
CCNet (ICCV'2019) context information -> concat , 연산 효율 79.45
DMNet (ICCV'2019) dynamic multi-scale filter with context 79.64
ANN (ICCV'2019) nonlocal block & pyramid pooling & attention 79.32
GCNet (ICCVW'2019/TPAMI'2020) se-net 79.18 80.71
FastFCN (ArXiv'2019) joint pyramid upsampling - 다른 head(PSP, deeplab …) deeplab v3 79.52 80.91
Fast-SCNN (ArXiv'2019) real-time 70.96
ISANet (ArXiv'2019/IJCV'2021) sparse self-attention 80.61 81.59
OCRNet (ECCV'2020) 81.35 82.7 V
DNLNet (ECCV'2020) attention, whiten 80.41
PointRend (CVPR'2020) point based & iterative 78.3 79.97 V
CGNet (TIP'2020) real-time
BiSeNetV2 (IJCV'2021) real-time
STDC (CVPR'2021) real-time
SETR (CVPR'2021) encoder - decoder, transformer layer 79.21 81.02
DPT (ArXiv'2021) dense viT / 작은 dataset에 잘됨 V
Segmenter (ICCV'2021) ViT, mask transformer decoder V
SegFormer (NeurIPS'2021) global & local attention & light V
K-Net (NeurIPS'2021) SOTA swinL+upernet V

참고 문서

https://github.com/open-mmlab/mmsegmentation https://paperswithcode.com/sota/semantic-segmentation-on-ade20k Cityscape를 benchmark로 사용한 모델들은 주로 오래된 모델입니다.