Skip to content

Latest commit

 

History

History
126 lines (74 loc) · 5.27 KB

README.md

File metadata and controls

126 lines (74 loc) · 5.27 KB
Raw

euclid labeller, and euclidaug augment engine for friction-less Deep Learning

Euclid object labeller for object detection training purposes based on Python. Tested on Linux, Windows, and Mac.

  • Supports Kitti format
  • Supports Yolo annotation format used in labelling, in the darknet framework (Generates bounding boxes, as well as training list file)
  • Supports PascalVOC format (XML tree)
  • Autolabelling (euclidaug is the augmentation tool, refer section below)

Dynamic selection of format is now supported.

Optimised box selection using keyboard shortcuts.

Example output from the IEEE paper "On the Applicability of Deep Learning for Road Signal Recognition", by Vinicios R. Soares, Mathias R. Luz, Afonso H. Ribeiro, Rafael S. Rbeiro, Marcella S. R. Martins, Max M. D. Santos, and Joaquim de Mira https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=8627071

alt text

Typical object labelling workflow using Euclid:

Create a folder containing the images

  • Ensure images are of uniform (not too big) shape using a command for resize like below. (Linux)

Resize all jpg images to 256x256, rename to sz-256-.jpg

for file in *.jpg; do convert $file -resize 256x256 sz-256-$file; done

  • $python euclid.py

  • Select the folder containing the images

  • Euclid will show the first image in the folder.

  • Select the class ID, and start labelling. Once done for this image, move to the next image, till all imagea are done.

  • Euclid also generates a supplementary file "train.txt", containing the class ID and full path of training file. This can be used in YOLO format training.

YOLO training and detection.

Refer below link for YOLO training and detection on Linux and Windows.

https://github.com/prabindh/darknet

Prebuilt binaries for Yolo are provided at https://github.com/prabindh/yolo-bins

Euclidaug image augmentation engine (frictionless Deep Learning)

Euclidaug automates the annotation task, so that developers can focus on gathering insights and improving accuracy, than spending time on manual labelling. Combined with AI generated artifical images (like https://generated.photos/), training becomes easy.

alt text

Lack of training images reduces training accuracy, when manual annotation is not an option. euclidaug serves to improve availability of training images, specially 2D.

Euclidaug can generate 1000s of precisely annotated/labelled training images in minutes, without manual work.

Example application of Euclidaug (auto-annotation)

For an example of auto-annotation of industrial capacitors to detect in a PCB using euclidaug, refer to https://github.com/prabindh/yolo-bins/tree/master/capacito

Installing euclidaug

cd euclidaug

pip install -r requirements.txt

Augmentation operations:

Random scaling, random packing, and random blending - are the operations performed on the images

Purpose of euclidaug:

To significantly augment availability of annotated data (# of images), specially 2D shapes for object recognition. Targeted for Yolo, SqueezeDet, Tensorflow (and others that support KITTI and Yolo formats)

How to run for example animal images (part of the tree):

(The images were taken from https://www.freeiconspng.com/img/1075, https://www.freeiconspng.com/img/1065, https://www.freeiconspng.com/img/1052)

Assuming euclid is cloned at c:\xxx, run the below command:

python euclidaug.py C:\xxx\euclid\euclidaug\sample-objects C:\xxx\euclid\euclidaug\sample-background C:\xxx\euclid\euclidaug\train-sample.txt

After seeing below, the augmented outputs will be generated in out_images and out_labels.

Info: Added [3] object images, Max obj/class of [1] Info: Added [1] base images Info: Beginning [11] images @ 0.421875 (sec) ........... Info: Completed @ 0.265625 (sec)

How to run adding new classes:

  • Place all object png images in a directory organised by IDs (for 3 classes, folders should be named as 0,1,2)

  • Place all background png images in a folder (can be any name)

  • Update cfgWidth, cfgHeight, numClasses, and Alpha/scale configurations - in the script, to match the framework requirements

  • Invoke this script as "python <script> "

  • output image files will be written to 'output-images' and 'output-labels' and training list file will be written

  • Note: The labels are in Yolo format, and output image files will be of same resolution as that of the "background" images.

Dependencies

Python 2.7 pip install pillow pip install image Python 3 Python 3 + Pillow on Ubuntu, do the below sudo apt-get install python-imaging-tk sudo apt-get install python3-pil.imagetk

Converting to TensorFlow format

After labelling the images, the labels can be read and converted to TFRecord using Python scripts available in Tensorflow, using tf.train.Example and tf.train.Features. Note: Yolo and TF share the same bounding box notations (normalised).

Writing to KITTI Format

Use the function "write2Kitti" instead of "write2Yolo"

Who uses euclidaug (Thank you!)

https://hackaday.io/project/161581-ai-equiped-wasp-and-asian-hornet-sentry-gun/log/155557-processing-the-images

https://github.com/dpogosov/yolo_kfm

https://github.com/suji104

https://github.com/VanitarNordic

https://github.com/lesterlo (DetectNet with Nvidia)

among others