boostcampaitech4lv23cv2 / level2_semanticsegmentation_cv-level2-cv-13

level2_semanticsegmentation_cv-level2-cv-13 created by GitHub Classroom
5 stars 1 forks source link

Semantic Segmentation Competition CV-13

๐Ÿ•ต๏ธMembers

๐Ÿ“ž TEAM 031

์‹ ์žฌ์˜

์œ ์Šน์ข…

์œค์ƒ์ค€

์ด์„ฑ์šฐ

์ด์˜์„ญ

๐Ÿ—‘๏ธ์žฌํ™œ์šฉ ํ’ˆ๋ชฉ ๋ถ„๋ฅ˜๋ฅผ ์œ„ํ•œ Semantic Segmentation

image

๋ฐ”์•ผํ๋กœ ๋Œ€๋Ÿ‰ ์ƒ์‚ฐ, ๋Œ€๋Ÿ‰ ์†Œ๋น„์˜ ์‹œ๋Œ€, ์šฐ๋ฆฌ๋Š” ๋งŽ์€ ๋ฌผ๊ฑด์ด ๋Œ€๋Ÿ‰์œผ๋กœ ์ƒ์‚ฐ๋˜๊ณ , ์†Œ๋น„๋˜๋Š” ์‹œ๋Œ€๋ฅผ ์‚ด๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ด๋Ÿฌํ•œ ๋ฌธํ™”๋Š” '์“ฐ๋ ˆ๊ธฐ๋Œ€๋ž€','๋งค๋ฆฝ์ง€ ๋ถ€์กฑ'๊ณผ ๊ฐ™์€ ์—ฌ๋Ÿฌ ์‚ฌํšŒ๋ฌธ์ œ๋ฅผ ๋‚ณ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

๋ถ„๋ฆฌ์ˆ˜๊ฑฐ๋Š” ์ด๋Ÿฌํ•œ ํ™˜๊ฒฝ ๋ถ€๋‹ด์„ ์ค„์ผ ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ• ์ค‘ ํ•˜๋‚˜์ž…๋‹ˆ๋‹ค. ์ž˜ ๋ถ„๋ฆฌ๋ฐฐ์ถœ ๋œ ์“ฐ๋ ˆ๊ธฐ๋Š” ์ž์›์œผ๋กœ์„œ ๊ฐ€์น˜๋ฅผ ์ธ์ •๋ฐ›์•„ ์žฌํ™œ์šฉ๋˜์ง€๋งŒ, ์ž˜๋ชป ๋ถ„๋ฆฌ๋ฐฐ์ถœ ๋˜๋ฉด ๊ทธ๋Œ€๋กœ ํ๊ธฐ๋ฌผ๋กœ ๋ถ„๋ฅ˜๋˜์–ด ๋งค๋ฆฝ ๋˜๋Š” ์†Œ๊ฐ๋˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ฆฌ๋Š” ์‚ฌ์ง„์—์„œ ์“ฐ๋ ˆ๊ธฐ๋ฅผ Segmentationํ•˜๋Š” ๋ชจ๋ธ์„ ๋งŒ๋“ค์–ด ์ด๋Ÿฌํ•œ ๋ฌธ์ œ์ ์„ ํ•ด๊ฒฐํ•ด๋ณด๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ๋ฌธ์ œ ํ•ด๊ฒฐ์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ๋Š” ๋ฐฐ๊ฒฝ, ์ผ๋ฐ˜ ์“ฐ๋ ˆ๊ธฐ, ํ”Œ๋ผ์Šคํ‹ฑ, ์ข…์ด, ์œ ๋ฆฌ ๋“ฑ 11 ์ข…๋ฅ˜์˜ ์“ฐ๋ ˆ๊ธฐ๊ฐ€ ์ฐํžŒ ์‚ฌ์ง„ ๋ฐ์ดํ„ฐ์…‹์ด ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ์—ฌ๋Ÿฌ๋ถ„์— ์˜ํ•ด ๋งŒ๋“ค์–ด์ง„ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์˜ ๋ชจ๋ธ์€ ์“ฐ๋ ˆ๊ธฐ์žฅ์— ์„ค์น˜๋˜์–ด ์ •ํ™•ํ•œ ๋ถ„๋ฆฌ์ˆ˜๊ฑฐ๋ฅผ ๋•๊ฑฐ๋‚˜, ์–ด๋ฆฐ์•„์ด๋“ค์˜ ๋ถ„๋ฆฌ์ˆ˜๊ฑฐ ๊ต์œก ๋“ฑ์— ์‚ฌ์šฉ๋  ์ˆ˜ ์žˆ์„ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๋ถ€๋”” ์ง€๊ตฌ๋ฅผ ์œ„๊ธฐ๋กœ๋ถ€ํ„ฐ ๊ตฌํ•ด์ฃผ์„ธ์š”! ๐ŸŒŽ

๐Ÿ’พ Datasets

๐Ÿ—“๏ธTimeline

image

๐Ÿง‘โ€๐Ÿ’ปTeam Roles

์‹ ์žฌ์˜

    SMP์™€ MMSeg๋ฅผ ์œ„ํ•œ baseline ์ž‘์„ฑ

    Sweep Configuration, Hard-voting ensemble

์œ ์Šน์ข…

    U-Net, U-Net++, U-Net 3+ ๋ชจ๋ธ ๋ถ„์„

    OpenCV๋ฅผ ํ™œ์šฉํ•œ ์ „์ฒ˜๋ฆฌ

    k-fold ๊ตฌํ˜„

์œค์ƒ์ค€

    Detectron2, MMsegmentation ํ™˜๊ฒฝ ๊ตฌ์„ฑ ๋ฐ ํ•™์Šต

     ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ๋ณ„ ํ•™์Šต์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ์…‹ ๊ตฌ์„ฑ ๋ณ€๊ฒฝ

     Wandb Sweep ํ™˜๊ฒฝ ๊ตฌ์„ฑ

์ด์„ฑ์šฐ

    Pseudo labeling ์ฝ”๋“œ ์ž‘์„ฑ

    MMsegmentation ํ™˜๊ฒฝ ๊ตฌ์„ฑ ๋ฐ ํ•™์Šต

    EDA

์ด์˜์„ญ

    k-fold dataset ์ฝ”๋“œ ์ž‘์„ฑ

    mmsegmentation train ํ™˜๊ฒฝ ๊ตฌ์ถ•

    mIoU metric ๋ถ„์„

๐Ÿ”๏ธEnvironments

drawing GitHub

drawing Notion

drawing Jira

drawing WandB

โš™๏ธRequirements

Ubuntu 18.04.5 LTS
Intel(R) Xeon(R) Gold 5120 CPU @ 2.20GHz
NVIDIA Tesla V100-PCIE-32GB

conda install pytorch=1.7.1 cudatoolkit=11.0 torchvision -c pytorch  
pip install openmim  
mim install mmseg  

Link To Installation Guide

๐ŸŽ‰Results๐ŸŽ‰

Public LB : 11th (mAP 0.7502)

image

Private LB : 12th (mAP 0.7296)

image

๐Ÿ“ŒPlease Look at our Wrap-Up Report for more details

image