Skip to content

Latest commit

 

History

History
174 lines (153 loc) · 7.23 KB

readme.md

File metadata and controls

174 lines (153 loc) · 7.23 KB
Raw

Felidae Information Flows

Felidae AI/ML image classifier

This repository is a work plan and source code product for Felidae information services. http://www.felidaefund.org/

A critical aspect of Felidae research is natural habitat image capture, by motion sensing video and still cameras. Mountain lions, bobcats, owls, possum, deer, skunk, rats, rabbits and humans are captured by cameras in remote locations, in addition to leaves rustling in the breeze and similar false triggers. Ratios of activities, through image categorization, noninvasively measures ecology health and vitality. But, categorization of retrieved media is a time intensive task, currently a critical bottleneck for the research.

"using deep-learning technology can save a tremendous amount of time for biology researchers and the human volunteers that help them by labeling images" (p 9, http://www.pnas.org/cgi/doi/10.1073/pnas.1719367115 Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning

https://news.mongabay.com/wildtech/2018/06/species-recognition-shifts-into-auto-with-neural-networks/ Species recognition shifts into auto with neural networks

This data flow refresh scales and improves the existing Felidae camera trap to data report process. Automation is used to standardize formats and accelerate the pipeline wherever possible. An infrastructure of modular, discrete components is designed to meet today's needs, and remain flexable for future scale and changing needs.

The key enhancements driving this information flow update are:

  • Integration of TensorFlow machine learning to auto categorize the images.
  • Metadata, schema, and flow standardization to facilitate R Project statistical analysis.
  • Data query and access simplification, to facilitate 3rd party research collaboration.
  • Improved backup pipeline, for disaster recovery.

Refresh Overview

Images and video are recorded in strategically placed camera traps. Periodically the memory cards are retrieved and needful field data is noted. A Raspberry Pi class wifi device receives the media and packages it with notebook annotations entered through a web interface. The media is then categorized within the package by an AI tool. A web interface presents pre-categorized media for manual corrections and/or upload to master storage (and duplication to offsite backup). Additional corrections and/or AI training takes place from the master, until packages are exported as CSV and media files. Stastical reports are generated from the exported data and uploaded back to the master data store. Any number of Raspberry Pi class devices may be used with an expected to cost of $20-$50 per site, or $15 each for an upload only version (NodeMCU class). This overview is depicted in the a Felidae-Flows pdf graphic.

Design

Glossary

Major
Minor
Unit
Risk
Plan
Execution
Check
Early
Middle
Final

Process

Since this product is implemented and consumed primarily by volunteers, with transient time availability and potential for high turn over, a special focus on planning documentation, automated operations and deployment robustness will be applied.

The general implementation pattern consists of four phases feeding each other in a continuous loop cycle: Risk Analysis, Planning, Execution and Quality Check. Three degrees of detail compose each phase: Major, Meta and Unit components. Likewise, the degrees of development are expected to mature in Early, Middle and Final stages of the project refinement.

This process allows for each of the four phases to be developed asynchronously, yet becoming tightly coupled as development matures. It is loosly based on the "Delivery Controls" and "Objective Improvement" slides in this operations deck https://github.com/georgalis/pub/blob/master/know/Operations.pdf

Delivery Controls

Objective Improvement

Design Outline

  • Major - broad goals
    • Viability (Risk) - key concerns
    • Twin (Design) - overall storyboard
    • Administer (Execute) - implementation report
    • Performance (Check) - utility validation
  • Meta - Major to Unit translation
    • Compliance (Risk) - plan implementation
    • Primitive (Design) - building blocks
    • Registration (Execute) - resource provisioning
    • Function (Check) - design/built alignment
  • Unit
    • Improve (Risk) - identify gaps
    • Specification (Design) - configuration and automation
    • Implementation (Execute) - setup
    • Test (Check) - confirm specification
  • Product

Major

New remote device connects to domestic internet and receives camera trap media, integrates with data entered through web interface and uploads to master. Master device receives remote device data, initiates backup, supports TensorFlow training, facilitates query and export.

Viability (Risk)

  • Hardware Disaster Recovery
  • Data Disaster Recovery
  • Operational Diagnostic Tools
  • Future Flexability
  • Reliability
  • Ease of use
  • Implementation cost
  • Ongoing cost

Twin (Design)

  • Discrete Modular Components
    • New remote device connects to domestic internet and receives camera trap media.
      • Remote device auto transfers media data to local storage.
      • Remote device presents field volunteer with web interface for media meta data entry.
      • Remote device packages exif and meta data and auto categorizes media.
      • Remote device presents media media by category or meta data query.
      • Remote device accepts manual media categorization corrections and annotates media data.
      • Remote device synchronizes data to master.
      • Remote device reports workflow status.
    • New master device
      • Master receives new media packages.
      • Master presents data administrator with web interface for audit and corrections of field volunteer data entry
      • Master presents data administrator with web interface for deep learning training.
      • Master invokes local and offsite backup.
      • Master provides 3rd party A&A for exports.
      • Master provides training configuration for remote devices.
  • Robust commodity hardware and software
  • Instantiation
  • Offsite Backup
  • Data
  • Hardware autoconfiguration, minimal setup
  • Web interface - low barrier use, zero install
    • Workflows
    • Markup
  • event logs, repair table

Administer (Execute)

Performance (Check)

  • User validation

Meta

Alignment of fundimental algorhythum and physical inventory.

Compliance (Risk)

Primitive (Design)

  • Components : Archive for archive, S3 for simple storage share, ESB for elastic, ingest new data to S3 and drop duplicates.

Registration (Execute)

  • Requirement Traceability Matrix

Function (Check)

Unit

Improve (Risk)

Specification (design)

Implementation (Execute)

  • Vendors
  • Approvals
  • Procurement
  • Bootstrap
  • Configuration Management
  • Implementation Report

Test (Check)

Product

,,,^..^,,,~