About

A Tensorflow implementation of Deeplabv3plus, trained on VOC2012 data set.

Difference from the paper Deeplabv3plus

• training strategy

  We use small batch_size = 8, and don't train the BN layer.

• atrous convolution

  We don't employ atrous convolution in the resnet model.

Set up

Prepare dataset

• Config the path in the input_data.py

The VOCdevkit directory should be as follows:

|--VOCdevkit
        |--train.txt         (10582 lines)
        |--val.txt           (1449 lines)
        |--test.txt          (1456 lines)
        |--train_raw.txt     (1464 lines)
        |--trainval_raw.txt  (2913 lines)
        |--VOC2012
            |--JPEGImages            (33260 images)
            |--SegmentationClass
            |--SegmentationClassAug  (12031 images)
            |--SegmentationObject
            |--ImageSets
            |--Annotations

• JPEGImages is needed. Download it from http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar for trainval set, and http://host.robots.ox.ac.uk:8080/eval/downloads/VOC2012test.tar for test data.
• SegmentationClassAug is needed. Download it from https://www.dropbox.com/s/oeu149j8qtbs1x0/SegmentationClassAug.zip?dl=0

Prepare pretrained resnet-101 model

• Config the path in the deeplab_model.py

The resnet_v2_101_2017_04_14 directory should be as follows:

|--resnet_v2_101_2017_04_14
               |--eval.graph
               |--resnet_v2_101.ckpt
               |-train.graph

resnet_v2_101.ckpt is needed. Download the directory(pretrained model) from http://download.tensorflow.org/models/resnet_v2_101_2017_04_14.tar.gz

Dependent libraries

• Tensorflow 1.11 and Python 2.7(3.7) have been tested.
• Anaconda
• Opencv

Inference

• download checkpoint from here

• randomly select images

  python predict.py

• select one image

  python predict.py --prediction_on_which val --filename 2007_002400

Evaluation

• download checkpoint from here

python evaluate.py

• for val data, generate prediction results and get mIoU.
• for test data, generate prediction results.

Re-Train to reproduce the result

• rm -r checkpint
• rm -r summary

python train.py

Results

on val set

Repo(%)	Paper(%)
75.84	79.35

sheep	horse	tv/monitor	bicycle	aeroplane	cow	dining table	bus	potted plant	background	dog	cat	person	train	bottle	car	chair	sofa	bird	boat	motorbike
0.85	0.84	0.67	0.41	0.88	0.86	0.56	0.92	0.59	0.94	0.86	0.9	0.84	0.88	0.78	0.82	0.41	0.54	0.88	0.66	0.83

predicted images

Note: left image is raw image, right image is the predicted image.

learning rate

loss

Reference

• https://github.com/rishizek/tensorflow-deeplab-v3-plus
• https://github.com/DrSleep/tensorflow-deeplab-resnet
• https://github.com/tensorflow/models/tree/master/research/deeplab