Human Parsing for Reid : human parsing networks trained on Look into Person dataset. LIP dataset has 20 body parts,while in this project we merge those 20 parts into 5 parts: background, head, upper-body, lower-body, shoes.
Input Size : 256 x 128
torch0.4.1
torchvision0.2.1
We support Deeplabv2, EDAnet, Tiny-FCN-4s, BiSeNet, Enet, ESPnet, Enet. Weights for Deeplabv2, EDAnet, Tiny-FCN-4s are available.
Model | meanIOU( 5 parts) | Parameters(MB) | Download(code 61ww) |
---|---|---|---|
Deeplabv2 | 67.92% | 42.76 | model |
EDAnet | 65.42% | 0.68 | model |
Tiny-FCN-4s | 65.21% | 0.71 | model |
Test on your own images: test.py
Train your network: train_uni.py
Prepare 4 parts datasets and show overlaps: /preprocess
from left to right : tiny-FCN-4s, EDAnet, deeplabv2-res101, ground truth