Label data for ML/DL methods with ease

[Not finished yet!]

Some existing tools

• ViTBAT: Video Tracking and Behavior Annotation Tool

  • I found this a little late, it seems did all the things I needed, however, mine is more adjustable -- with source code to tweak with.

• Mathworks, Define Ground Truth for Image Collections

• Labelbox

• FastAnnotationTool, seems have lots of methods

• LabelMeAnnotationTool, this is huge

• label-V, something I would like to do, maybe can make use of it

• wkentaro/labelme, seems nice, no special operations for different platforms

• dbolkensteyn/vatic.js, one js tool, really impressive

Blogs

A Definitive Guide To Build Training Data For Computer Vision

This one

Design Philosophy

Semi-automatic labeling, label aided by algorithms.

Pipeline

Video ->
image frames (png for potential transparent part) ->
distinguish the blurred ->
extracted very N frame for labeling ->
use existing tracking algorithms to track, both forward and backward ->
combine the tracked BBox -> refine BBox (optional) ->
train with ML detection algorithms

NOTES

• the object is out of the scene for tracking
• data format transfer coco2pascal.py

TODO

It seems that, we need a UI, really.

Possible lib:

• jdoepfert/roipoly.py, Select a polygonal region of interest (ROI) with python and matplotlib.

Dependencies

• dlib

• labelme

• scipy

• preprocess

• tqdm

• scikit-image

• json_tricks

• imutils