Getting LOST can be good..
Posted on
PlantVillage AI engineers Rohit Gangupantulu and Peter McCloskey, have successfully utilized open access resources on GitHub to design and develop an image annotation software. The LOST (Label Objects and Save Time) tool allows our experts (located globally) and undergraduates at Penn State annotate images in batches, online, all within the web browser. This erases the need for annotators to: 1) be in our lab facility 2) need extra resources (hardware/software) 3) annotate one image at a time. Rohit explains why PlantVillage needed this tool, “In today’s technical landscape, we have immense access to cloud resources that render on-premise or local solutions pointless. I have witnessed software issues in annotations that we had done in the past, and that along with the fact that scalability is quite critical in serving more and more users, made us look for a solution. How can we improve this part of the machine learning pipeline and expand its reach? An online, OS-agnostic, cloud-based solution was always going to be our answer. We wanted many people to be able to annotate, and for it to seamlessly work into adding to our training data.”
Figure 1. The pipeline for the annotation process.
LOST is an open-source tool available as a repository on GitHub. It offers single image and multi-image annotation through the web browser. According to Rohit, “We have a lot of ability to truly make it our own in terms of the pipelines built, especially on a custom level. It works well with other technologies and frameworks we currently have in place, which allows us to create unique preprocessing workflows for machine learning. Its user management and UI allows for anyone to annotate at any time, just by logging in as one would do for logging into Facebook to chat." This tool is now a critical part of our model training pipeline for Pete. “The annotation process is consistently the rate-limiting step in developing highly accurate machine learning models. Thanks to LOST, we are able to reduce the time it takes to label our data by weeks or even months. Allowing us to quickly get our AI tools into the hands of small-holder farmers around the world who face an uncertain future full of climate catastrophes, invasive trans-boundary pests, and devastating diseases”.
Figure 2. The annotation labels for our locust annotation.
It is easy to use and PlantVillage has integrated this tool for our locust work, with undergraduates annotating thousands of images within days. We are now looking to implement this tool in our other projects which would help annotators from Peru to Kenya!
Figure 3. An example of a task an annotator would receive.
We are hoping we can share this tool with the public to crowdsource annotations from willing members of the community. In a world with declining experts, we need to increase the ease of access to obtaining their knowledge.
Written By: Annalyse Kehs, Operations Director
Contact: Peter McCloskey (pzm44@psu.edu)
LOST Credit (Name _ GitHub):
Jonas Jäger @jaeger-j Gereon Reus @gereonreus Dennis Weiershäuser @cartok Tobias Kwant @tkwant
Citation:
@article{jaeger2019lost,
title={{LOST}: A flexible framework for semi-automatic image annotation},
author={Jonas J\"ager and Gereon Reus and Joachim Denzler and Viviane Wolff and Klaus Fricke-Neuderth},
year={2019},
Journal = {arXiv preprint arXiv:1910.07486},
eprint={1910.07486},
archivePrefix={arXiv},
primaryClass={cs.CV}