LiheYoung / U-RISC

2nd place solution to the Ultra-high Resolution EM Images Segmentation Challenge hosted by BAAI and PKU

Ultra-high Resolution EM Images Segmentation Challenge

2nd place in the Ultra-high Resolution EM Images Segmentation Challenge hosted by BAAI and PKU

Environment

Hardware

• 4 NVIDIA Tesla V100 GPUs (32GB memory each)
• CPU memory 250GB

Packages

pip install -r requirements.txt

Data

Processed data can be downloaded here. Put it in top-level folder.

Model

Pretrained Model:

ResNet-50: Download, ResNet-101: Download, ResNet-152: Download

Simple Track

DFF, backbone ResNet-50

Complex Track

CASENet, backbone ResNet-152

Training

Simple Track

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset simple --model DFF --backbone resnet50 --batch-size 4 --lr 0.0014 --epochs 200 --crop-size 960 --aug --k 1

Complex Track

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --dataset complex --model CASENet --backbone resnet152 --batch-size 4 --lr 0.0014 --epochs 45 --crop-size 1280 --aug --kernel-size 9 --edge-weight 0.4

Testing and Ensembling

Simple Track

CUDA_VISIBLE_DEVICES=0 python test.py --dataset simple --model DFF --backbone resnet50

Complex Track

CUDA_VISIBLE_DEVICES=0 python test.py --dataset complex --model CASENet --backbone resnet152