papers/zoo-mobile-inference/zoo-mobile-inference.tex

\documentclass[11pt,a4paper]{article}
\usepackage{shared/zoocover}

% ============================================================
% Preamble
% ============================================================
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{lmodern}
% geometry loaded by zoocover
\usepackage{amsmath,amssymb,amsfonts}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{hyperref}
\usepackage{cleveref}
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{enumitem}
\usepackage{xcolor}
\usepackage{tikz}
\usetikzlibrary{shapes,arrows,positioning,fit}
\usepackage{float}
\usepackage{subcaption}
\usepackage{multirow}
\usepackage{tabularx}
\usepackage{fancyhdr}
\usepackage{authblk}
\usepackage{abstract}
\usepackage{natbib}
\bibliographystyle{plainnat}

\hypersetup{
    colorlinks=true,
    linkcolor=blue!70!black,
    citecolor=green!50!black,
    urlcolor=blue!80!black
}

\pagestyle{fancy}
\fancyhf{}
\rhead{Antje Worring, Zoo Labs Foundation Research}
\lhead{Mobile Inference for Field Conservation}
\rfoot{Page \thepage}

% ============================================================
\title{\textbf{Mobile Inference for Field Conservation: On-Device Species Recognition} \\[0.5em]
\large Model Compression, Quantization, and Offline-First Deployment for Edge Biodiversity Monitoring}

\author[1]{Antje Worring, Zoo Labs Foundation Research}
\affil[1]{Antje Worring, Zoo Labs Foundation Research, Research Division \\ \texttt{[email protected]}}

\date{December 2024}

\begin{document}

\zoocoverpage

% ============================================================
% Abstract
% ============================================================
\begin{abstract}
\noindent
Automated species recognition has achieved remarkable accuracy in laboratory settings, but deploying these models to the field---where connectivity is absent, hardware is constrained, and real-time identification drives conservation decisions---remains an open challenge.
We present ZooMobile, a comprehensive framework for deploying species recognition models on consumer mobile devices and embedded sensors in conservation field settings.
ZooMobile addresses three technical challenges: reducing model size by 94\% through a novel \emph{Taxonomic Knowledge Distillation} (TKD) pipeline that exploits the hierarchical structure of biological classification; achieving sub-100ms inference latency on mid-range smartphones through mixed-precision quantization with species-specific calibration; and maintaining recognition accuracy under degraded field conditions (low light, motion blur, partial occlusion) through a \emph{Condition-Aware Inference} (CAI) module that adapts preprocessing and confidence thresholds to estimated image quality.
We evaluate ZooMobile on a new benchmark, ZooBench-Field, comprising 287,000 images of 4,200 species captured under authentic field conditions across 31 conservation sites.
On this benchmark, ZooMobile achieves 87.3\% top-1 accuracy and 96.1\% top-5 accuracy with a 12 MB model running at 47ms per inference on a Qualcomm Snapdragon 7 Gen 1 processor, compared to 91.8\% top-1 accuracy for a 203 MB cloud-based baseline.
In a 6-month field deployment supporting 142 rangers across 8 national parks, ZooMobile enabled 23,400 real-time species identifications with 84.7\% field-verified accuracy, directly informing 47 anti-poaching patrol decisions and 12 human-wildlife conflict interventions.
\end{abstract}

\vspace{0.5em}
\noindent\textbf{Keywords:} species recognition, model compression, quantization, edge inference, mobile deployment, conservation technology

% ============================================================
% 1. Introduction
% ============================================================

\input{sections/01-introduction}
\input{sections/02-related-work}
\input{sections/03-taxonomic-knowledge-distillation}
\input{sections/04-species-specific-mixed-precision-quantiz}
\input{sections/05-condition-aware-inference}
\input{sections/06-zoobench-field-benchmark}
\input{sections/07-offline-first-architecture}
\input{sections/08-field-deployment}
\input{sections/09-discussion}
\input{sections/99-conclusion}

\end{document}