- detect malaria parasites and white blood cells in thick blood smears
- origin paper: https://arxiv.org/abs/1904.07850
- origin repo: https://github.com/xingyizhou/CenterNet
- new datasets: Malaria-dataset (COCO-format)
- new backbone: ResNet18, ResNet34, MobileNetV2, MobileNetV3, MNASNet, EfficientNets
- new neck: FPN
- fix gaussian radius calculation in lib/utils/image.py (refer to https://github.com/princeton-vl/CornerNet)
- add cpu inference ability
model | resolution | precision | recall | f1 | fps |
---|---|---|---|---|---|
resnet18prm | 384 x 384 | 90.1 | 98.6 | 94.35 | 125 |
resnet18prm | 512 x 512 | 91.0 | 99.0 | 95.00 | 100 |
resnet18prm | 600 x 600 | 94.6 | 99.5 | 97.05 | 91 |
- pip install -r requirements.txt
- bash compile_coco.sh
- bash compile_nms.sh
malaria dataset are publicly available at:
- raw: ftp://lhcftp.nlm.nih.gov/Open-Access-Datasets/Malaria/Thick_Smears_150
- processed: https://drive.google.com/file/d/1JP9egCgDB36MblsgyklQlGlTH-iluOC3/view?usp=sharing
-
train: bash cmd_train.sh
-
eval: bash cmd_test.sh
-
demo: bash cmd_demo.sh