UnsubscribesSuppressionLists_UploadDataFile.png' alt='Suppression Files Txt' title='Suppression Files Txt' />Last updated April 17, 2017 Now you can disable your Caps Lock key in Windows. Is there a oneliner that will zipunzip files. PowerShell This document describes and defines the ondisk database file format used by all releases of SQLite since version 3. The Database File. Object detection using Fast R CNNTable of Contents. Summary. This tutorial describes how to use Fast R CNN in the CNTK Python API. Fast R CNN using Brain. Script and cnkt. exe is described here. The above are examples images and object annotations for the grocery data set left and the Pascal VOC data set right used in this tutorial. Fast R CNN is an object detection algorithm proposed by Ross Girshick in 2. The paper is accepted to ICCV 2. Fast R CNN builds on previous work to efficiently classify object proposals using deep convolutional networks. Compared to previous work, Fast R CNN employs a region of interest pooling scheme that allows to reuse the computations from the convolutional layers. Setup. To run the code in this example, you need a CNTK Python environment see here for setup help. Please install the following additional packages in your cntk Python environmentpip install opencv python easydict pyyaml dlib. Pre compiled binaries for bounding box regression and non maximum suppression. The folder ExamplesImageDetectionutilscythonmodules contains pre compiled binaries that are required for running Fast R CNN. The versions that are currently contained in the repository are Python 3. Windows and Python 3. Linux, all 6. 4 bit. If you need a different version you can compile it following the steps described at. Copy the generated cythonbbox and cpunms andor gpunms binaries from FRCNROOTlibutils to CNTKROOTExamplesImageDetectionutilscythonmodules. Example data and baseline model. We use a pre trained Alex. Net model as the basis for Fast R CNN training for VGG or other base models see Using a different base model. Both the example dataset and the pre trained Alex. Net model can be downloaded by running the following Python command from the Fast. RCNN folder python installdataandmodel. Run the toy example. To train and evaluate Fast R CNN run python runfastrcnn. The results for training with 2. ROIs on Grocery using Alex. Net as the base model should look similar to these AP for gerkin 1. AP for butter 1. AP for joghurt 1. AP for egg. Box 1. AP for mustard 1. AP for champagne 1. AP for orange 1. AP for water 0. AP for avocado 1. AP for tomato 1. AP for pepper 1. AP for tabasco 1. AP for onion 1. AP for milk 1. AP for ketchup 0. AP for orange. Juice 1. Mean AP 0. 9. 47. To visualize the predicted bounding boxes and labels on the images open Fast. RCNNconfig. py from the Fast. RCNN folder and set C. VISUALIZERESULTS True. The images will be saved into the Fast. RCNNOutputGrocery folder if you run python runfastrcnn. Train on Pascal VOCTo download the Pascal data and create the annotation files for Pascal in CNTK format run the following scripts python ExamplesImageData. SetsPascalinstallpascalvoc. ExamplesImageData. SetsPascalmappingscreatemappings. Change the datasetcfg in the getconfiguration method of runfastrcnn. Pascalconfig import cfg as datasetcfg. Now youre set to train on the Pascal VOC 2. Beware that training might take a while. Train on your own data. Prepare a custom dataset. Option 1 Visual Object Tagging Tool RecommendedThe Visual Object Tagging Tool VOTT is a cross platform annotation tool for tagging video and image assets. VOTT provides the following features Computer assisted tagging and tracking of objects in videos using the Camshift tracking algorithm. Exporting tags and assets to CNTK Fast RCNN format for training an object detection model. Running and validating a trained CNTK object detection model on new videos to generate stronger models. How to annotate with VOTT Download the latest Release. Follow the Readme to run a tagging job. After tagging Export tags to the dataset directory. Option 2 Using Annotation Scripts. To train a CNTK Fast R CNN model on your own data set we provide two scripts to annotate rectangular regions on images and assign labels to these regions. The scripts will store the annotations in the correct format as required by the first step of running Fast R CNN A1Generate. Input. ROIs. py. First, store your images in the following folder structurelt yourimagefolder negative images used for training that dont contain any objectslt yourimagefolder positive images used for training that do contain objectslt yourimagefolder test. Images images used for testing that do contain objects. For the negative images you do not need to create any annotations. For the other two folders use the provided scripts Run C1Draw. Bboxes. On. Images. In the script set img. Dir lt yourimagefolder positive or test. Images before running. Add annotations using the mouse cursor. Once all objects in an image are annotated. Run C2Assign. Labels. To. Bboxes. py to assign labels to the bounding boxes. In the script set img. Dir lt yourimagefolder positive or test. Images before running. The script loads these manually annotated rectangles for each image, displays them one by one. Ground truth annotations marked as either undecided or exclude are fully excluded from further processing. Train on custom dataset. After storing your images in the described folder structure and annotating them please runpython ExamplesImageDetectionutilsannotationsannotationshelper. Finally, create a My. Data. Setconfig. C. CNTK. DATASET Your. Data. Set. C. CNTK. MAPFILEPATH. Data. Intel Atom Processor Lan Driver. SetsYour. Data. Set. C. CNTK. CLASSMAPFILE classmap. C. CNTK. TRAINMAPFILE trainimgfile. C. CNTK. TESTMAPFILE testimgfile. C. CNTK. TRAINROIFILE trainroifile. C. CNTK. TESTROIFILE testroifile. C. CNTK. NUMTRAINIMAGES 5. C. CNTK. NUMTESTIMAGES 2. C. CNTK. PROPOSALLAYERSCALES 8, 1. Note that C. CNTK. PROPOSALLAYERSCALES is not used for Fast R CNN, only for Faster R CNN. To train and evaluate Fast R CNN on your data change the datasetcfg in the getconfiguration method of runfastrcnn. My. Data. Setconfig import cfg as datasetcfg. Technical details. The Fast R CNN algorithm is explained in the Algorithm details section together with a high level overview of how it is implemented in the CNTK Python API. This section focuses on configuring Fast R CNN and how to you use different base models. Parameters. The parameters are grouped into three parts Detector parameters see Fast. RCNNFast. RCNNconfig. Data set parameters see for example utilsconfigsGroceryconfig. Base model parameters see for example utilsconfigsAlex. Netconfig. pyThe three parts are loaded and merged in the getconfiguration method in runfastrcnn. In this section well cover the detector parameters. Data set parameters are described here, base model parameters here. In the following we go through the most important parameters in Fast. RCNNconfig. py. All parameters are also commented in the file. The configuration uses the Easy. Dict package that allows easy access to nested dictionaries. Number of regions of interest ROIs proposals. C. NUMROIPROPOSALS 2. Io. U overlap of a proposal to qualify for training regression targets. C. BBOXTHRESH 0. Maximum number of ground truth annotations per image. C. INPUTROISPERIMAGE 5. C. IMAGEWIDTH 8. C. IMAGEHEIGHT 8. Use horizontally flipped images during training C. TRAIN. USEFLIPPED True. If set to True conv layers weights from the base model will be trained, too. C. TRAINCONVLAYERS True.