DeepLab-v1-Tensorflow

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.

TODO

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)

Requirements

TensorFlow>=1.0 is supported

To install the required python packages (except TensorFlow), run

pip install -r requirements.txt

or for a local installation

pip install -user -r requirements.txt

Best results at present

mAP: 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

parameters

Optimizer: SGD

Batch Size: 10

Learning rate: 1e-3

Lr_decay_step: 5000

Total_step: 20000

Momentum: 0.9

Weight decay: 0.0005

Caffe model