This is an implementation of DeepLab-LargeFOV in TensorFlow for semantic image segmentation on PASCAL VOC dataset.
This code is based on the implementation from tensorflow-deeplab-lfov. Please check this repository for details.
- Works with TensorFLow>=1.0
- Weight decay
- Tracks training with TensorBoard
- Fully functional evaluation code
- Achieve the performance reported in ArXiv
The Post-processing step with DenseCRF
Atrous spatial pyramid pooling (ASPP)
TensorFlow>=1.0 is supported
To install the required python packages (except TensorFlow), run
pip install -r requirements.txtor for a local installation
pip install -user -r requirements.txtmAP: 59.22999966028858 (62.25 reported in ArXiv)
background: 88.5654446653
aeroplane: 66.6891497821
bicycle: 27.5425855685
bird: 71.9108573055
boat: 51.4274847257
bottle: 62.3651852486
bus: 78.4051023721
car: 70.6123718826
cat: 75.3887068995
chair: 25.7829156204
cow: 53.5545469656
diningtable: 48.3028937323
dog: 69.360881792
horse: 52.9500919802
motorbike: 64.8494736002
person: 72.9275538607
pottedplant: 35.1361763734
sheep: 65.3341241448
sofa: 37.3502489281
train: 71.4086065257
tvmonitor: 53.9655908926
Optimizer: SGD
Batch Size: 10
Learning rate: 1e-3
Lr_decay_step: 5000
Total_step: 20000
Momentum: 0.9
Weight decay: 0.0005
You can download two already converted models (model.ckpt-init and model.ckpt-pretrained) here.