update 2018/8/21
Prepare custom datasets for gluonCV and mxnet
https://gluon-cv.mxnet.io/build/examples_datasets/detection_custom.html
this simple tutorial will introduce how to use im2rec for mx.image.ImageIter and ImageDetIter and how to use im2rec for COCO DataSet ok let's start but prepare your im2rec first https://github.com/apache/incubator-mxnet/blob/master/tools/im2rec.py
step 1. you should make a .lst file run in terminal
python im2rec.py --list testlst Dataset
--list => make .lst file
you should set True when your folder didnt have any .lst file
test => name of your .lst name what you want
DataSet => your image dataset folder
check your folder path correct
use .lst for imageIter
data_iter = mx.image.ImageIter(
batch_size=4,
data_shape=(3,816, 1232),
label_width=1,
path_imglist='test.lst',
path_root='DataSet')
for data in data_iter:
d = data.data[0]
break
img =nd.transpose(d,(0,2,3,1))
print(img.shape)
io.imshow(img[0].asnumpy().astype(np.uint8))
io.show()step 2. make .rec file
python im2rec.py testlst.lst Dataset
testlst => name of your .rec name what you want
Dataset => your image dataset folder
use .rec for imageiter
train_iter = mx.image.ImageIter(
batch_size=32,
data_shape=(3, 816, 1232),
path_imgrec='data.rec',
path_imgidx='data.idx', #help shuffle performance
shuffle=True,
aug_list=[mx.image.HorizontalFlipAug(0.5)]
)
train_iter.reset()
for batch in train_iter:
x = batch.data[0]
break
img =nd.transpose(x,(0,2,3,1))
print(img.shape)
io.imshow(img[0].asnumpy().astype(np.uint8))
io.show()other feature
'--list'
If this is set im2rec will create image list(s) by traversing root folder
and output to .lst
Otherwise im2rec will read .lst and create a database at .rec
'--exts', default=['.jpeg', '.jpg', '.png']
list of acceptable image extensions.
'--train-ratio', default=1.0
Ratio of images to use for training.
'--test-ratio', default=0,
Ratio of images to use for testing.
'--no-shuffle', dest='shuffle', action='store_false',
If this is passed,im2rec will not randomize the image order in <prefix>.lst'
'--resize', type=int, default=0,
resize the shorter edge of image to the newsize, original images will
be packed by default'
'--center-crop',
specify whether to crop the center image to make it rectangular.
'--quality', default=95,
JPEG quality for encoding, 1-100; or PNG compression for encoding, 1-9
'--num-thread',default=1,
number of thread to use for encoding. order of images will be different
'--color', default=1, choices=[-1, 0, 1],
specify the color mode of the loaded image.
1: Loads a color image. Any transparency of image will be neglected.
It is the default flag
0: Loads image in grayscale mode.
-1:Loads image as such including alpha channel.
A B [extra header] [(object0), (object1), ... (objectN)]
- you should write the .lst with your image information
Where A is the width of header (2 + length of extra header), B is the width of each object. Extra header is optional and used for inserting helper information such as (width, height). Each object is usually 5 or 6 numbers describing the object properties, for example: [id, xmin, ymin, xmax, ymax, difficulty] Putting all together, we have a
lstfile for object detection:
ref: https://mxnet.incubator.apache.org/api/python/image/image.html
with open('dataset.lst', 'w+') as f:
for i in range(3):
f.write(
str(i) + '\t' +
# idx
str(4) + '\t' + str(5) + '\t' +
# width of header and width of each object.
str(256) + '\t' + str(256) + '\t' +
# (width, height)
str(1) + '\t' +
# class
str((i / 10)) + '\t' + str((i / 10)) + '\t' + str(((i + 3) / 10)) + '\t' +str(((i + 3) / 10)) + '\t' +
# xmin, ymin, xmax, ymax
str(i) + '.jpg\n'
)
you will see like this format in .lst
idx A B [extra header] [(object0), (object1), ... (objectN)]
0 4 5 256 256 1 0.1 0.1 0.3 0.3 0.jpg
1 4 5 256 256 1 1.1 1.1 1.3 1.3 1.jpg
2 4 5 256 256 1 2.1 2.1 2.3 2.3 2.jpg
must '\t' between data information
python im2rec.py --pack-label dataset.lst Dataset
--pack-label => when you have label
dataset.lst => name for .rec type for your self in .lst
Dataset => DataSet image folder
import mxnet as mx
import mxnet.ndarray as nd
from skimage import io
import numpy as np
train_iter = mx.image.ImageDetIter(
batch_size=3,
data_shape=(3, 256, 256),
path_imgrec='dataset.rec',
path_imgidx='dataset.idx',
shuffle=True,
rand_crop=1,
min_object_covered=0.95,
max_attempts=200
) # you can aug your data in ImageDetIter
train_iter.reset()
for batch in train_iter:
x = batch.data[0]
y = batch.label[0]
break
print(y[0])
img =nd.transpose(x,(0,2,3,1))
print(img.shape)
io.imshow(img[0].asnumpy().astype(np.uint8))
io.show()1. download cococ dataset first!!!
2. make own .json or use COCO.json
3. .json => .lst
4. .lst => .rec
5. .rec for ImageDetIter
import json
with open('annotations/instances_train2017.json', 'r') as f:
DataSets = json.load(f)
print(DataSets['annotations'][0])
>>{'segmentation': [[239.97, 260.24, 222.04, 270.49, 199.84, 253.41, 213.5, 227.79, 259.62, 200.46, 274.13, 202.17, 277.55, 210.71, 249.37, 253.41, 237.41, 264.51, 242.54, 261.95, 228.87, 271.34]],
'area': 2765.1486500000005,
'iscrowd': 0,
'image_id': 558840,
'bbox': [199.84, 200.46, 77.71, 70.88],
'category_id': 58, 'id': 156}as you see, these annotations about image 558840 , boundingbox,class.... in detail you can check coco website
make own dataset (class:bird) from COCO dataset
import json
from mxnet import image
from skimage import io
import os
## load COCO annotations
with open('annotations/instances_train2017.json', 'r') as f:
DataSets = json.load(f)
print(DataSets['annotations'][0])
## save class and own dataset .json
jsonName = 'ownset.json'
directory = 'ownSet/'
data = {}
data['DataSet] = []
with open(jsonName, 'w') as outfile:
if not os.path.exists(directory):
os.makedirs(directory)
for DataSet in DataSets['annotations']:
box = DataSet['bbox']
default_name = "000000000000"
img_id = str(DataSet['image_id'])
img_name = default_name[:len(default_name) - len(img_id)] + str(img_id) + '.jpg'
coco_name = 'train2017/' + img_name
if DataSet['category_id'] == 16: #bird
with open(coco_name, 'rb') as f:
img = image.imdecode(f.read())
height = img.shape[0]
width = img.shape[1]
box[0] = box[0]/width #normalize
box[2] = box[2]/width
box[1] = box[1]/height
box[3] = box[3]/height
io.imsave(directory + img_name, img.asnumpy())
data['DataSet'].append({
'img_name': img_name,
'height': height,
'width': width,
'bbox': box,
'class':DataSet['category_id']
})
json.dump(data, outfile)
print('save ok')
with open(jsonName, 'r') as f:
Sets = json.load(f)
print(Sets['DataSet'][0]){'img_name': '000000202273.jpg', 'height': 640, 'width': 480, 'bbox': [0.6530625000000001, 0.089296875, 0.33064583333333336, 0.075390625], 'class': 16}
as you see this is your own dataset annotations
1.How to use own own dataset.json and make own dataset .lst
import json
import mxnet as mx
from skimage import io
jsonName = 'ownset.json'
directory = 'ownSet/'
with open(jsonName, 'r') as f:
DataSet = json.load(f)
print(DataSet['DataSet'][0]['img_name'])
img_idx = 0
with open('ownSet.lst', 'w+') as f:
for Data in DataSet['DataSet']:
x_min = Data['bbox'][0]
y_min = Data['bbox'][1]
x_max = Data['bbox'][0]+ Data['bbox'][2]
y_max = Data['bbox'][1]+ Data['bbox'][3]
f.write(
str(img_idx) + '\t' + # idx
str(4) + '\t' + str(5) + '\t' + # width of header and width of each object.
str(int(Data['height'])) + '\t' + str(Data['width']) + '\t' + # (width, height)
str(1) + '\t' + # class
str(x_min) + '\t' + str(y_min) + '\t' + str(x_max) + '\t' + str(y_max) + '\t' + # xmin, ymin, xmax, ymax
str(Data['img_name'])+'\n')
img_idx += 1
you will see your .lst like this format
this step just use 'bird' class (Previous) to show you, .lst => .rec
python im2rec.py --pack-label ownSet.lst ownSet
ownSet.lst => last step you make
ownSet => image folder
import mxnet as mx
shape = 800
train_iter = mx.image.ImageDetIter(
batch_size=32,
data_shape=(3, shape, shape),
path_imgrec='ownSet.rec',
path_imgidx='ownSet.idx',
shuffle=False,
) # you can aug your data in ImageDetIter
import matplotlib.pyplot as plt
def box_to_rect(box, color, linewidth=3):
box = box.asnumpy()
print((box[0], box[1]), box[2] - box[0], box[3]-box[1])
return plt.Rectangle(
(box[0], box[1]), box[2] - box[0], box[3]-box[1],
fill=False, edgecolor=color, linewidth=linewidth
)
train_iter.reset()
batch = train_iter.next()
img, labels = batch.data[0], batch.label[0]
print(labels.shape)
img = img.transpose((0,2,3,1))
img = img.clip(0,255).asnumpy()/255
for i in range(32):
_, fig = plt.subplots()
plt.imshow(img[i])
rect = box_to_rect(labels[i][0][1:5]*shape,'red',2)
fig.add_patch(rect)
fig.axes.get_xaxis().set_visible(False)
fig.axes.get_yaxis().set_visible(False)
plt.show()
result
will generate dataset.lst
with open('dataset.lst', 'w+') as f:
for i in range(12):
f.write(
str(i) + '\t' +
# idx
str(4) + '\t' + str(5) + '\t' +
# width of header and width of each object.
str(256) + '\t' + str(256) + '\t' +
# (width, height)
str(1) + '\t' +
# class
str((i / 15)) + '\t' + str((i / 15)) + '\t' + str(((i + 3) / 15)) + '\t' +str(((i + 3) / 15)) + '\t' +
str(2) + '\t' +
# class
str((i / 50)) + '\t' + str((i / 50)) + '\t' + str(((i + 3) / 50)) + '\t' + str(((i + 3) / 50)) + '\t' +
str(3) + '\t' +
# class
str((i / 100)) + '\t' + str((i / 100)) + '\t' + str(((i + 3) / 100)) + '\t' + str(((i + 3) / 100)) + '\t' +
# xmin, ymin, xmax, ymax
str(i) + '.jpg\n'
)check dataset.lst
python im2rec.py --pack-label dataset.lst data
import mxnet as mx
shape = 800
train_iter = mx.image.ImageDetIter(
batch_size=32,
data_shape=(3, shape, shape),
path_imgrec='dataset.rec',
path_imgidx='dataset.idx',
shuffle=False,
) # you can aug your data in ImageDetIter
import matplotlib.pyplot as plt
def box_to_rect(box, color, linewidth=3):
box = box.asnumpy()
print((box[0], box[1]), box[2] - box[0], box[3]-box[1])
return plt.Rectangle(
(box[0], box[1]), box[2] - box[0], box[3]-box[1],
fill=False, edgecolor=color, linewidth=linewidth
)
train_iter.reset()
batch = train_iter.next()
img, labels = batch.data[0], batch.label[0]
print(labels.shape)
img = img.transpose((0,2,3,1))
img = img.clip(0,255).asnumpy()/255
for i in range(12):
_, fig = plt.subplots()
plt.imshow(img[i])
color_list = ['red','blue','black']
for k in range(labels[i].shape[0]): # how many object in your label
rect = box_to_rect(labels[i][k][1:5]*shape,color_list[k],2)
fig.add_patch(rect)
fig.axes.get_xaxis().set_visible(False)
if you feel useful pls give a star!!!