Awesome
Image Segmentation via Keras
Implementation of Segnet, FCN, UNet, PSPNet and other models in Keras. This repo is cloned and modify based on https://github.com/divamgupta/image-segmentation-keras. Click here to see original README
Important Dependencies
-
python 3.5+
-
Keras 2.0
-
python_opencv
-
Theano / Tensorflow / CNTK
-
numpy, matplotlib
Use following to install necessary dependencies
python setup.py install
Prepare data for segmentation
I use CUB-200-2011 and VOC-2012 for image segmentation. You should first download these dataset.
After downloading the dataset, you should put corresponding dataset into Datasets
, it should be like
-- Datasets
---- CUB_200_2011
------ images
------ segmentations
---- VOC
------ JEPGImages
------ SegmentationClass
------ SegmentationObject
To prepare data for segmentation task, use DataPreprocess.py
in utils
, for example
convert_segmentations_CUB()
make_dataset_CUB()
convert_segmentations_VOC()
make_dataset_VOC
After that, there are train
and test
directory in above mentioned datasets
Run the code
To run the code, you should see the example in run.py
.
import keras_segmentation
from keras.models import load_model
import tensorflow as tf
import os
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"]="2, 3"
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.8)
DATA_NAME = "VOC"
EPOCH = [5, 10]
CHECKOUTPOINT_PATH = "./tmp/voc_5_10"
if DATA_NAME == "VOC":
train_images_path = "./Datasets/VOC/train/imgs/"
train_segs_path = "./Datasets/VOC/train/segs/"
test_images_path = "./Datasets/VOC/test/imgs/"
test_segs_path = "./Datasets/VOC/test/segs"
class_num = 21
if DATA_NAME == "CUB":
train_images_path = "./Datasets/CUB_200_2011/train/imgs/"
train_segs_path = "./Datasets/CUB_200_2011/train/segs/"
test_images_path = "./Datasets/CUB_200_2011/test/imgs/"
test_segs_path = "./Datasets/CUB_200_2011/test/segs"
class_num = 2
'''
Change model name
'''
model = keras_segmentation.models.unet.resnet50_unet(n_classes=class_num, input_height=416, input_width=608)
# model.load_weights("./tmp/cub_psspnet_vgg_pspnet.9")
for i in range(EPOCH[0]):
'''
Train
'''
model.train(
train_images = train_images_path,
train_annotations = train_segs_path,
checkpoints_path = CHECKOUTPOINT_PATH , epochs=EPOCH[1], verify_dataset = False
)
'''
Output
'''
# for (i, image_dir) in enumerate(os.listdir(test_images_path)):
# if i%50 == 0:
# out = model.predict_segmentation(
# inp= os.path.join(test_images_path, image_dir),
# out_fname= os.path.join("./Output/", image_dir)
# )
# # import matplotlib.pyplot as plt
# # plt.imshow(out)
'''
Test mIoU
'''
test_image_list = [os.path.join(test_images_path, i) for i in os.listdir(test_images_path)]
test_segs_list = [os.path.join(test_segs_path, i) for i in os.listdir(test_segs_path)]
model.evaluate_segmentation(
inp_images = test_image_list,
annotations = test_segs_list
)
To get output images, you should see example in test.py
:
import keras_segmentation
import numpy as np
import cv2
import os
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"]="2,3"
DATA_NAME = "VOC"
if DATA_NAME == "VOC":
train_images_path = "./Datasets/VOC/train/imgs/"
train_segs_path = "./Datasets/VOC/train/segs/"
test_images_path = "./Datasets/VOC/test/imgs/"
test_segs_path = "./Datasets/VOC/test/segs"
class_num = 21
if DATA_NAME == "CUB":
train_images_path = "./Datasets/CUB_200_2011/train/imgs/"
train_segs_path = "./Datasets/CUB_200_2011/train/segs/"
test_images_path = "./Datasets/CUB_200_2011/test/imgs/"
test_segs_path = "./Datasets/CUB_200_2011/test/segs"
class_num = 2
'''
Change model name
'''
model = keras_segmentation.models.unet.resnet50_unet(n_classes=class_num, input_height=416, input_width=608)
model.load_weights("./tmp/voc_5_10.9")
# load any of the 3 pretrained models
for (i, image_dir) in enumerate(os.listdir(test_images_path)):
if image_dir == "2011_002114.jpg":
out = model.predict_segmentation(
inp= os.path.join(test_images_path, image_dir),
out_fname= os.path.join("./Output_VOC/", image_dir))
Change model names, checkpoint names to run the corresponding models.
Contact
If you have any problem, please contact me through: lidongyue AT sjtu DOT edu DOT cn