def readDataset(self, datasetFolder, compressionWidth, feature):
"""
This method reads the folder structure of the dataset and initialize
the attributes of the data manager. In general the folder structure is
divided in three parts: i) Videos: contains the raw video sequences,
ii) Positives: Containing the masks of the positive samples, iii)
Negatives: Containing the masks of the negative samples. For
illustrative purposes lets name our dataset as "EV", and lets define
its root folder as "EV/". The folder structure is briefly summarized in
the next table:
.. list-table::
:widths: 10 20 60
:header-rows: 1
* - Path
- Content
- Description
* - <dataset>/Videos
- Full video files
- Original video sequences. Each video could contains positives as
well as negative frames. Each video should be named as
<dataset>_<videoid>.<extension>. For example the full path of a
video in the EV dataset could be "EV/Videos/EV_Video1.MP4".
* - <dataset>/Positives
- Folders
- <dataset>/Positives contains a folder per video that is going to
be used to extract positive frames (with hands). For example,
lets assume that the frame 10, 20 and 30 of
"EV/Videos/EV_Video1.MP4" are going to be used as positive
samples in the training stage, then the positives folder should
contain these files::
"EV/Positives/EV_Video1/mask10.jpg",
"EV/Positives/EV_Video1/mask20.jpg",
"EV/Positives/EV_Video1/mask30.jpg"
respectively. In practice the mask files could be empty files
because they are used only to guide the scanning of the video.
However, as a way to validate the used frames we suggest to use
compressed snapshots of the real frames.
* - <dataset>/Negatives
- Folders
- <dataset>/Negatives contains a folder per video that is going to
be used to extract negative frames (without hands). For example,
lets assume that the frame 30, 100 and 120 of
"EV/Videos/EV_Video2.MP4" are going to be used as negative
samples in the training stage. To do this the negatives folder should
contain these files::
"EV/Positives/EV_Video1/mask30.jpg",
"EV/Positives/EV_Video2/mask100.jpg",
"EV/Positives/EV_Video2/mask120.jpg"
respectively. In practice the mask files could be empty files
because they are only used to guide the scanning of the video.
However, as a way to validate the used frames we suggest to use
compressed snapshots of the real frames.
Finally, following the previous example the folder structure is::
EV/Videos/EV_Video1.MP4
EV/Videos/EV_Video2.MP4
EV/Positives/EV_Video1/mask10.jpg
EV/Positives/EV_Video1/mask20.jpg
EV/Positives/EV_Video1/mask30.jpg
EV/Negatives/EV_Video2/mask30.jpg
EV/Negatives/EV_Video2/mask100.jpg
EV/Negatives/EV_Video2/mask120.jpg
Example 1: How to read the dataset folder from egovision::
from egovision.handDetection import HandDetectionDataManager
from egovision.values.paths import DATASET_PATH
feature = "HOG"
dataset = "UNIGEmin"
datasetFolder = DATASET_PATH.format(dataset)
dm = HandDetectionDataManager()
dm.readDataset(datasetFolder, 200, feature)
"""
from datetime import datetime
from egovision import Video
from egovision.features import FeatureController
self.datasetFolder = datasetFolder
self.compressionWidth = compressionWidth
self.feature = feature
categories = ["Negatives", "Positives"]
featureController = FeatureController(compressionWidth, feature)
for nc, cat in enumerate(categories):
categoryFolder = "".join([datasetFolder, "/", cat, "/"])
videoNames = os.listdir(categoryFolder)
for videoName in videoNames:
masks = os.listdir("".join([categoryFolder, videoName]))
masks.sort(key=lambda x: int(x[4:-4]))
fVideoName = "".join([datasetFolder, "/Videos/", videoName])
fVideoName = fullVideoName(fVideoName)
video = Video(fVideoName)
for mask in masks:
sys.stdout.flush()
fmNumber = int(mask[4:-4])
t0 = datetime.now()
success, frame = video.readFrame(fmNumber)
t1 = datetime.now()
success, desc = featureController.getFeatures(frame)
t2 = datetime.now()
# sysout = "\r{0}: - {1} - {2} - {3}".format(videoName, mask, t2-t1,t1-t0)
# sys.stdout.write(sysout)
self.headers.append("".join(
[fVideoName, "_", str(fmNumber)]))
self.attributes.append(desc.next())
self.categories.append(nc)
self.attributes = numpy.vstack(self.attributes)
# TRAINING THE MULTIMODEL PIXEL BY PIXEL
hs = PixelByPixelMultiHandSegmenter(FEATURE, COMPRESSION_WIDTH, CLASSIFIER,
STEP)
dataManagerList = hs.trainClassifier(allMasks[:N_MODELS], modelRecommender)
# DEFINING THE HAND2HAND OCLUSSION DETECTOR
occDetector = SuperpixelsOcclusionDetector(COMPRESSION_WIDTH)
# occDetector.tuneSuperpixelAlgorithm(SIGMA, dataManagerList, hs)
occDetector.setSuperpixelAlgorithm(
nSegments=250, compactness=7,
sigma=SIGMA) # obtained from the optimization
# LOADING TESTING VIDEO
video = Video(DATASET_VIDEOS_PATH.format(DATASET, TESTINGVIDEO) + ".mp4")
fFrame = 0
video.readFrame(fFrame)
writer = VideoWriter(
RESULTS_PATH.format(DATASET, TESTINGVIDEO) + "_segmented.avi")
writer.setParametersFromVideo(video)
# leftState = []
# leftFilter = EllipseTracker(COMPRESSION_WIDTH, 1/60.0, modelId = 0)
# rightFilter = EllipseTracker(COMPRESSION_WIDTH, 1/60.0, modelId = 1)
# PER MASK
fout = open(RESULTS_PATH.format(DATASET, TESTINGVIDEO) + ".csv", "w")
for frameNumber, frame in enumerate(video):
print fFrame + frameNumber
binaryShape = (frame.matrix.shape[0], frame.matrix.shape[1], 1)
# SEGMENT