papers/zoo-mobile-inference/sections/06-zoobench-field-benchmark.tex

\section{ZooBench-Field Benchmark}

\label{sec:benchmark}

\subsection{Dataset Construction}

Existing species recognition benchmarks (iNaturalist, CUB-200, PlantCLEF) are dominated by curated photographs taken under favorable conditions.
We construct ZooBench-Field to evaluate recognition under authentic field conditions:

\begin{table}[H]
\centering
\caption{ZooBench-Field dataset statistics.}
\label{tab:zoobench}
\begin{tabular}{@{}lr@{}}
\toprule
\textbf{Statistic} & \textbf{Value} \\
\midrule
Total images & 287,412 \\
Species & 4,203 \\
Collection sites & 31 \\
Countries & 18 \\
Camera types & 47 (smartphones, camera traps, DSLRs) \\
Expert-verified labels & 100\% \\
Quality annotations & 100\% (4-axis) \\
Condition distribution & Low light 23\%, Motion blur 18\%, Occlusion 31\%, Long distance 28\% \\
\bottomrule
\end{tabular}
\end{table}

\subsection{Benchmark Results}

\begin{table}[H]
\centering
\caption{Species recognition accuracy on ZooBench-Field.}
\label{tab:benchmark_results}
\begin{tabular}{@{}lcccc@{}}
\toprule
\textbf{Model} & \textbf{Size (MB)} & \textbf{Latency} & \textbf{Top-1 (\%)} & \textbf{Top-5 (\%)} \\
\midrule
ViT-Large (cloud) & 1,142 & 2.1s & 91.8 & 98.2 \\
BioCLIP (cloud) & 428 & 890ms & 89.4 & 97.1 \\
EfficientNet-B7 (cloud) & 256 & 680ms & 88.7 & 96.8 \\
MobileNetV3 (standard KD) & 21.4 & 89ms & 83.1 & 94.7 \\
MobileNetV3 (TKD) & 18.7 & 82ms & 85.4 & 95.8 \\
\textbf{ZooMobile (full)} & \textbf{3.4} & \textbf{47ms} & \textbf{86.8} & \textbf{95.9} \\
ZooMobile w/o CAI & 3.4 & 41ms & 84.2 & 94.3 \\
ZooMobile w/o TKD & 3.1 & 47ms & 83.8 & 94.1 \\
\bottomrule
\end{tabular}
\end{table}

ZooMobile achieves 86.8\% top-1 accuracy with a 3.4 MB model and 47ms latency, closing 76\% of the accuracy gap to the 1,142 MB cloud baseline while being 45x faster and 336x smaller.

\subsection{Accuracy by Condition}

\begin{table}[H]
\centering
\caption{ZooMobile accuracy by image condition (top-1 \%).}
\label{tab:condition_accuracy}
\begin{tabular}{@{}lccc@{}}
\toprule
\textbf{Condition} & \textbf{ZooMobile} & \textbf{w/o CAI} & \textbf{Cloud Baseline} \\
\midrule
High quality & 92.4 & 91.8 & 95.1 \\
Low light & 83.7 & 78.4 & 89.2 \\
Motion blur & 81.2 & 76.1 & 87.8 \\
Partial occlusion & 79.8 & 77.3 & 85.4 \\
Long distance & 78.4 & 73.2 & 84.7 \\
\midrule
\textbf{Overall} & \textbf{86.8} & 84.2 & 91.8 \\
\bottomrule
\end{tabular}
\end{table}

CAI provides the largest benefit under low light (+5.3\%) and long distance (+5.2\%) conditions, where adaptive preprocessing (histogram equalization and super-resolution, respectively) substantially improve feature quality.

% ============================================================
% 7. Offline-First Architecture
% ============================================================