Skip to content

README.md

Latest commit

 

History

History
176 lines (143 loc) · 6.67 KB

README.md

File metadata and controls

176 lines (143 loc) · 6.67 KB

Hand Bone Image Segmentation

This project is the Naver Boost Camp CV11 team's submission code for Hand Bone Image Segmentation competition.
Given an image of hand bone, it is a matter of segmenting 29 bone parts.

Team Members


Environment

  • OS : Linux Ubuntu 18.04.5
  • GPU : Tesla V100 (32GB)

Folder Structure

├─eda
├─ensemble
├─mmdetection
├─mmdetection3
├─UniverseNet
├─yolov8
├─multilabel_kfold.py
└─streamlit



Usage

Install Requirements

  • pip install -r requirements.txt



train.sh

  1. Move the path to the tools folder where the train.sh file is located

  2. Write python3 train.py command in train.sh file

    python3 train.py --model FCN --loss bce_loss --epochs 100

    All arguments for config training

     python3 train.py [-h] [--seed SEED] [--loss LOSS] [--model MODEL] [--epochs EPOCHS] [--val_every VAL_EVERY] [--train_batch TRAIN_BATCH] [--train_workers TRAIN_WORKERS] [--wandb WANDB] [--encoder ENCODER]
             [--save_dir SAVE_DIR] [--model_path MODEL_PATH] [--debug DEBUG] [--transform TRANSFORM] [--acc_steps ACC_STEPS] [--dataclean DATACLEAN]

  3. Run

    nohup sh train.sh

Result

Loss experiment result

Metric : Dice coefficient

model encoder loss wandb public
MAnet resnet101 bce 0.9516 0.9456
MAnet resnet101 smp_dice 0.9513 0.9489
MAnet resnet101 calc 0.9523 0.9484
MAnet resnet101 smp_focal 0.9506 0.9452
MAnet resnet101 bce_with_logit 0.9492
MAnet resnet101 jaccard 0.9538 0.9496
MAnet resnet101 tversky 0.9524 0.9490
MAnet resnet101 comb(bce: 0.5, dice: 0.5) 0.9526 0.9500
MAnet resnet101 comb1(bce: 0.33, dice:0.33, jaccard: 0.33) 0.9524 0.9490
MAnet resnet101 comb2(bce: 1, dice: 3, jaccard: 6) 0.9528 0.9500
MAnet resnet101 comb3(bce: 0.1, dice:0.6, jaccard: 0.3) 0.9531 0.9503

We use comb3 as our loss. You can use it by set argument when training "--loss comb_loss"

Augmentation experiment result

Metric : Dice coefficient

augmentation wandb 리더보드
ElasticTransform(300) 0.9527 0.9507
Rotate(limit=45) 0.9524 0.9502
RandomContrast(limit=[0,0.5], p=1) 0.9617 0.9497 (inference transform: 0.9500)
RandomContrast (limit=0.2, p=0.5) 0.9515 0.9484
RandomContrast(limit=[0,0.5], p=0.5) 0.9493 0.9471
Normalize 0.9486 0.9465
ElasticTransform(400) 0.9524 0.9450
Crop 0.9479
CenterCrop 0.9343 0.9243 (inference transform: 0.2927)
Equalize and remove black (200) 0.6426(100) →0.7891 (200)

We use ElasticTransform(300), Rotate(limit=45) and RandomContrast(limit=[0,0.5], p=1) as our final augmentation.

TTA experiment result

Metric : Dice coefficient

TTA 리더보드
TTA 적용 안했을 때 0.9710
HorizontalFlip 0.9710
HorizontalFlip, Multiply([0.9,1,1.1,1.2]) 0.9717
HorizontalFlip, Multiply([0.9,1,1.1,1.2]), Rotate 90 0.9585

We use HorizontalFlip, Multiply([0.9,1,1.1,1.2]) as our final TTA combination

Final Solution

Metric : Dice coefficient image

  • Final submission : Public : 0.9743(2nd) / Private : 0.9749(2nd)