What Input Resolution Is Required for Bird Species Identification, and What Is Its Latency Cost on an Edge Device? A Study of 14 Input Resolutions and Six Architectures with On-Device Measurements

Takeshi Nishikawa

cs.CVSep 13, 2026

What Input Resolution Is Required for Bird Species Identification, and What Is Its Latency Cost on an Edge Device? A Study of 14 Input Resolutions and Six Architectures with On-Device Measurements

Authors: Takeshi Nishikawa

Organizations: Foundation for Computational Science (FOCUS), Kobe, Japan

Abstract

Bird-strike mitigation at wind farms requires identifying distant birds that span only tens of pixels, so the classifier's input resolution N is a design variable, not a fixed specification. We study it with a factorial design over 14 side lengths N (16 to 224), six architectures, two training and evaluation regimes and 30 random seeds -- 2,520 checkpoints and 5,040 evaluations -- plus latency measured on an NVIDIA Jetson Orin Nano. Four results. (1) The selected N depends on the target: 0.90 is met on validation by ResNet50 at N=112 in an estimated 1.85 ms (0.8980 on test) and 0.95 by DINOv2-L at N=144 in 12.70 ms; changing the model buys more accuracy than raising N (+5.93 versus +2.33 points at N=112). (2) The benefit of lowering N depends on the assumed preprocessing path: N=224 -> 80 saves 13.5% when each individual is decoded from its own file but 46.2% when the detector decodes the 4K frame once; the Pareto set grows from 20 to 23 configurations. (3) Accuracy must be measured on the deployed engine: half precision costs ViT-S/16 alone 4 to 7 points at N>=96 while the CNNs stay within 0.1 points, and with selection held at validation the choice differs at 26 of 176 targets. A broken FP16 engine can run faster than a correct one, undetectable from latency; admitting 14 ViT-S/16 FP32 configurations moves the recommendation over the 0.931-0.938 band and under the 10 ms budget. (4) ViT-L-scale models fit this device, but activations exceed the FP16 range; splitting the graph at transformer-block boundaries confines FP32 to the affected segments, making the deployed DINOv2-L chain 1.85x faster than the single-engine build. We also quantify how the regime-difference sign stabilises with seed count; a sensitivity split removing some forms of group sharing preserves all 14 non-trivial signs at the selection boundary.

Figures & tables

Figure 1: Accuracy vs. resolution for the six architectures. Left: regime A (degraded input to a 224-trained model). Right: regime B (resolution-native retraining). Bands show the standard deviation over 30 seeds.

Model	$N{=}16$	32	48	96
MobileNetV4	+38.0	+33.5	+18.0	+2.4
EfficientNet-B0	+37.2	+27.1	+15.2	+1.4
ResNet50	+47.8	+33.2	+19.6	+3.4
ViT-S/16	+25.4	+3.3	$-$ 3.1	$-$ 2.3
DINOv2-L	+11.3	$-$ 4.2	$-$ 3.8	$-$ 0.9
DINOv3-L	+7.0	$-$ 10.7	$-$ 5.8	$-$ 0.9

Table 1: Gain of regime B over regime A (points, mean of 30 seeds).

Figure 2: Minimum required input side length for each model. Retraining lowers the requirement for CNNs and raises it for the ViT models.

Model	$N{=}16$	$N{=}112$	$N{=}224$	$r224/r16$
MobileNetV4	0.966	1.034	1.426	1.48
EfficientNet-B0	1.284	1.427	2.113	1.65
ResNet50	0.996	1.164	1.858	1.86
ViT-S/16	1.211	1.310	2.359	1.95

Table 2: Orin Nano FP16 inference time (

\mathrm{ms}

; median of 30 runs).

	Preprocessing [ms]		MobileNetV4 classification [ms]		Saving
Path	$N{=}224$	$N{=}80$	$N{=}224$	$N{=}80$	$224\rightarrow 80$
Offline evaluation	5.98	5.43	7.40	6.41	13.5%
Deployment	1.22	0.45	2.65	1.42	46.2%

Table 3: Effect of the assumed preprocessing path (Orin Nano,

n=200

). The classification columns are estimated preprocessing plus GPU compute for one individual; GPU time is identical in both rows, so only the preprocessing model differs. The last column is the saving from

N{=}224

N{=}80

within each path, not a per-resolution quantity.

Figure 3: Latency against input resolution for all six models. Left: measured latency. Right: normalised to $N{=}16$ , so the right-hand edge is $t_{224}/t_{16}$ .

DINOv2-L, 5-way split, FP32 only on the last part; $t_{224}/t_{16}=2.64$
$N^{\prime}$	Orin ( $\mathrm{ms}$ )	H200 ( $\mathrm{ms}$ )	edge/server	accuracy	agreement
14	10.240	1.099	9.3	0.7343	99.79%
28	8.657	1.100	7.9	0.8443	99.79%
42	12.596	1.075	11.7	0.8895	97.93%
70	9.056	1.076	8.4	0.9304	98.94%
84	13.889	1.167	11.9	0.9501	99.15%

Table 4: ViT-L on the edge at all 14 resolutions.

N^{\prime}

is the side length actually fed to the model: DINOv2-L is patch-14, so the requested

N\in\{16,32,\dots,224\}

is rounded to a multiple of 14, whereas DINOv3-L is patch-16 and

N^{\prime}=N

. Rows of the two blocks correspond to the same requested

N

. Accuracy is top-1 measured on the device; agreement is argmax agreement with server FP32. Every configuration was run on all

n=1{,}882

test images and verified to depend on its input. The accuracy and agreement columns come from a single checkpoint (seed 42), not from a multi-seed mean; the 29-seed means are given in § 7.4 .

Figure 4: Left: edge/server latency ratio, all at $N{=}32$ . Right: per-stage breakdown of the deployed split chains at $N{=}32$ .

Target	Model	$N$	$N^{\prime}$	Classification [ms]	Validation	Test accuracy
0.80	ResNet50	32	32	1.16 (6.17)	0.8134	0.8049
0.85	ResNet50	64	64	1.38 (6.38)	0.8694	0.8614
0.90	ResNet50	112	112	1.85 (6.84)	0.9051	0.8980 †
0.93	ResNet50	224	224	3.08 (7.84)	0.9302	0.9213 †
0.95	DINOv2-L	144	140	12.70 (17.68)	0.9653	0.9630
0.96	DINOv2-L	144	140	12.70 (17.68)	0.9653	0.9630

Table 5: Fastest configuration meeting each accuracy target, among the 84 base candidates. The time column is the estimated classification time (preprocessing plus GPU compute for one individual) on the deployment path, with the offline-evaluation figure in parentheses; it excludes detection, tracking and frame decoding and is not a measured whole-system time.

N^{\prime}

is the side length actually fed to the model; DINOv2-L rounds to a multiple of 14. Accuracy is measured on the deployed Orin engine (30 seeds; 29 for the ViT-L-scale models). Configurations are selected on the validation accuracy shown and the test accuracy is reported alongside; the two rows marked

\dagger

meet the target on validation but fall slightly short on test.

Location	max. abs. value	vs. FP16 ceiling
DINOv2-L, inside blocks 18–20	87	fits
DINOv2-L, inside blocks 21–23 + head	356,409	$5.4\times$ over
DINOv3-L, stage boundaries (all resolutions)	155,118	$2.4\times$ over

Table 6: Maximum absolute activation, against the FP16 ceiling of 65,504.

Figure 5: Accuracy versus measured latency on Orin Nano. Large markers are the Pareto set selected on validation over all 84 candidates, the same decision as the tables; because the vertical axis is test accuracy, a selected point can appear dominated there. The small models use FP16; DINOv2-L uses mixed FP16/FP32 and DINOv3-L uses FP32 throughout. The vertical axis is accuracy measured on the deployed Orin engines (30 seeds; 29 for the ViT-L-scale models).

Figure 6: Argmax agreement between the deployed Orin configurations and server FP32. The CNNs and ViT-S/16 run in FP16; DINOv3-L is FP32 throughout. For DINOv2-L both the deployed mixed-precision configuration (solid) and the all-FP32 control (dashed) are shown. The control increases agreement at five of the six tested resolutions, the largest increase being 1.06 percentage points at requested $N{=}64$ ; this indicates a small but measurable contribution from the precision configuration and does not establish low resolution as the cause of the remaining disagreement. ViT-L is measured at the six requested levels $N{=}16,32,64,112,128,224$ ; the horizontal axis is the side length actually fed to the model, so DINOv2-L appears at 14, 28, 70, 112, 126 and 224.

$N$	distilled	CE only	difference [pp] (sign agreement)
16	0.2965	0.2779	$+1.86$ (mixed)
32	0.4851	0.4853	$\mathbf{-0.02}$ (mixed)
64	0.7558	0.7561	$-0.04$ (mixed)
96	0.8574	0.8585	$-0.11$ (mixed)
128	0.8978	0.8927	$+0.51$ (mixed)
176	0.9145	0.9100	$+0.44$ ( 3/3 )

Table 7: Degradation robustness with and without distillation at selected resolutions (EfficientNet-B0, three seeds; accuracy on a 0–1 scale, differences in percentage points).

Figure 7: $15\text{\,}\mathrm{W}$ mode against MAXN_SUPER for the four smaller models (56 configurations); the ViT-L split chains are summarised in the text. Left: latency in the two power modes. Right: energy per inference.

On-Device Inference

On-device inference runs machine-learning models locally on phones, wearables, microcontrollers, and edge hardware to reduce latency, energy use, and data sharing.

See all 190 papers

Jul 28, 2026

Lightweight Image Classification of Raptor Species for Edge Devices: Rare-Species Dataset Expansion via Video Frame Extraction, Knowledge Distillation, and TensorRT Deployment

We investigate lightweight raptor-species classification for real-time edge deployment in wind-turbine collision mitigation. Using DINOv2-L (304M parameters) as a teacher, we distilled three lightweight students (MobileNetV4, ViT-Small, and EfficientNet-B0). To reduce confusion between closely related species, we expanded the dataset to 12,519 images, including an increase in Steller's Sea Eagle images from 463 to 2,050 via video-frame extraction. Under a group split that separates samples at the video- and source-image level to mitigate source leakage at that granularity, the three-student ensemble achieved a macro recall of 0.935 +/- 0.004 over five distillation seeds (0.955 on a conventional image-level split, retaining 97.5% of the teacher's macro recall) with roughly one-eighth as many parameters. On a subset of 1,258 images disjoint from the former training images, White-tailed Eagle recall improved by up to 38.6 percentage points, while the rate at which it was misclassified as the Steller's Sea Eagle decreased from 61% to 15% of errors. TensorRT FP16 deployment of EfficientNet-B0 on an NVIDIA Jetson Orin Nano achieved 3.19 ms/image including host-device transfer (313 images/s), with 99.95% argmax agreement with FP32. In five-seed controlled comparisons, neither distillation (versus CE-only) nor the change from a DINOv2-L to a DINOv3-L teacher yielded a clear ensemble-level improvement; the primary gains stem from the dataset expansion and teacher re-fine-tuning.

Takeshi Nishikawa

Foundation for Computational Science (FOCUS), Computational Science Center Building 1F, 7–1–28 Minatojima-Minamimachi, Chuo-ku, Kobe 650–0047, Japan

Edge Inference Fine-Grained Image Classification

Ask PDF

Jul 22, 2026

Ultra-Compact CNN Architectures for Tropical Bird Audio Detection on Microcontrollers

Passive acoustic monitoring of tropical biodiversity is bottlenecked by the storage and battery cost of continuously recording soundscapes in which bird vocalisations typically occupy less than 10% of the audio. Autonomous recording units built on low-power microcontrollers (typically ARM Cortex-M with

\leq

256 kB of RAM) address this by triggering only on likely-positive segments, but the on-device options are unsatisfying: coarse frequency-energy triggers such as Goertzel filters flood SD cards with false positives at

\sim

71% precision, whereas neural detectors developed for temperate single-species tasks are either too large to deploy or transfer poorly to species-rich tropical settings. We present DrongoNet, a family of three INT8 CNN detectors sized for this envelope and validated on a 50,000-clip, 1,677-species Southeast Asian tropical dataset (SEABAD). The headline model, DrongoNet-Micro (919 parameters, 6.26 kB, 0.9810 AUC, 98.3% mean recall at τ = 0.35), is a drop-in replacement for the Goertzel trigger used in commodity field recorders: at α = 0.10 tropical prevalence it captures 8 pp more bird vocalisations than Goertzel and extends a 32 GB card from

\sim

28 to

\sim

45 days of monitoring. DrongoNet-Nano (5.09 kB) bounds the ultra-low-flash extreme; DrongoNet-Edge (33.06 kB, 0.9991 AUC) targets Linux SBCs. On SEABAD, Micro matches a retrained TinyChirp CNN-Mel baseline within 0.1 pp AUC at 28

\times

fewer parameters, confirming that the family is deployment-agnostic across mel-spectrogram bird corpora but requires per-environment retraining. Full INT8 quantisation costs

<

0.12% AUC across all three variants.

Muhammad Mun'im Ahmad Zabidi, Mohd Yamani Idna Idris, Norisma Idris

Faculty of Computer Science and Information Technology, Universiti Malaya, Kuala Lumpur, Malaysia

Convolutional Neural Networks Wildlife Monitoring Sound Event Detection Tiny ML Bioacoustics Passive Acoustic Monitoring+1

Ask PDF

May 26, 2026

JetViT: Efficient High-Resolution Vision Transformer with Post-Training Attention Search

We introduce JetViT, a novel family of hybrid-architecture Vision Transformer (ViT) models that match the accuracy of state-of-the-art full-attention vision foundation models while achieving substantially higher inference efficiency on high-resolution images. At the core of our approach is Post-Training Attention Search, a post-training acceleration framework that converts pre-trained full-attention ViTs into efficient hybrid-attention variants by identifying and replacing redundant full-attention blocks with linear or window-attention blocks. By inheriting the MLP and attention weights from the base model, Post-Training Attention Search efficiently explores the architectural design space through three key steps: (1) optimizing the linear-attention block design; (2) finding the best combination of linear-attention and window-attention blocks; and (3) identifying and preserving critical full-attention blocks. We evaluate JetViT on two representative high-resolution vision foundation models, DINOv3 and DepthAnythingV2. On the NVIDIA H100 GPU, JetViT achieves up to 1.79x higher throughput and up to 44.81% lower latency without sacrificing accuracy. We will release our code and accelerated ViT models soon.

Dongyun Zou, Zhuoyang Zhang, Junyu Chen +8

MIT · NVIDIA · University of Pennsylvania +1

Efficient Transformer Inference Vision Foundation Models Transformer Attention Efficient ViTs Neural Architecture Search

Ask PDF

Abstract

On-Device Inference

Explore similar work

Lightweight Image Classification of Raptor Species for Edge Devices: Rare-Species Dataset Expansion via Video Frame Extraction, Knowledge Distillation, and TensorRT Deployment

Ultra-Compact CNN Architectures for Tropical Bird Audio Detection on Microcontrollers

JetViT: Efficient High-Resolution Vision Transformer with Post-Training Attention Search