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

Takeshi Nishikawa

REVIEW 3 major objections 4 minor 39 references

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

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A three-student ensemble reproduces 97.5% of a 304M-parameter teacher's raptor classification recall with one-eighth the parameters, after expanding the rare species' training images; the same model runs in FP16 at 313 images/s on an edge G

desk verdict A solid, honest applied ML paper whose headline group-split recall is probably a bit optimistic because the split doesn't close leakage at iNaturalist-observation or non-V6-video-frame level; still, the core claims and negative results are useful. read the letter →

arxiv 2607.26238 v1 pith:XPBD7ACG submitted 2026-07-28 cs.CV cs.LGeess.IV

Takeshi Nishikawa This is my paper · ORCID

classification cs.CVcs.LGeess.IV

keywords raptorspeciesclassificationknowledgedistillationdatasetexpansionvideoframeextractionedgeinferenceTensorRTFP16DINOv2birdstrikemitigation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a lightweight three-student ensemble can carry fine-grained six-species raptor identification onto an edge GPU without losing much of a much larger teacher's accuracy—provided the scarce related species gets enough training images. The authors report a macro recall of 0.935±0.004 on a group split that keeps video/source-image samples out of the same partition (0.955 on a conventional image-level split), reproducing about 97.5% of the re-fine-tuned teacher's recall with roughly one-eighth the parameters. Their controlled five-seed comparisons are the distinctive part: knowledge distillation versus hard-label training shows no clear ensemble-level gain, and a newer teacher generation shows no clear gain either; the accuracy improvements come from expanding the Steller's Sea Eagle class from 463 to 2,050 images (video-frame extraction) and re-fine-tuning the teacher. The same deployed student runs in half precision on an edge GPU at about 313 images/s with virtually no accuracy loss, putting real-time use within reach for wind-turbine collision mitigation.

What carries the argument

The carrying mechanism is a two-stage transfer pipeline. A large self-supervised Vision Transformer teacher (DINOv2-L, 304M parameters) is re-fine-tuned on the expanded training set, then three small students are trained with a loss that combines temperature-scaled KL divergence toward the teacher's soft outputs with hard-label cross-entropy; at inference the students' softmax outputs are averaged. The other load-bearing piece is the group split: samples are grouped by video ID for video frames and by source image for detector-crop series, and each group is kept entirely within one split, so the headline number is intended as a leakage-mitigated generalization estimate rather than a same-vid

What would settle it

Take the saved manifest metadata, reconstruct observation-level and recording-session-level group IDs for every image, and re-run the same group-split five-seed distillation; if ensemble macro recall falls well below 0.935±0.004, the reported unseen-data performance is inflated by source leakage. A simpler check is a cross-split pHash scan of all images for near-duplicate pairs spanning training and test.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that data expansion plus teacher re-fine-tuning—not the distillation mechanism itself—produced the gains. After expanding the six-species set to 12,519 images, the large self-supervised teacher was re-fine-tuned for 30 epochs (new-test accuracy 0.8985→0.9777), and three students (MobileNetV4, ViT-Small, EfficientNet-B0) were distilled from it. The ensemble's macro recall is 0.935±0.004 over five distillation seeds under a group split that prevents video frames and source-image crops from straddling training and test, and 0.955 on the image-level split. The related-species confusion metric improved sharply: White-tailed Eagle misclassified as Steller's Sea E

Load-bearing premise

The headline generalization figure depends on the group split keeping every near-duplicate image of the same source observation or video out of both training and test, and the paper itself states that grouping at the observation level or at the frame level of earlier video sources cannot be recovered from the file names.

Editorial extensions

If this is right

Under the leakage-mitigating group split, the three-student ensemble (34.13M total parameters) reaches macro recall 0.935±0.004 over five distillation seeds, or about 97.5% of the re-fine-tuned teacher's level.
Expanding Steller's Sea Eagle training images from 324 to 1,432 (with the teacher re-fine-tuned) cuts White-tailed Eagle-to-Steller's misclassification from 61% to 15% of errors and lifts White-tailed Eagle recall by up to 38.6 points on the disjoint subset.
Because five-seed comparisons show distillation versus CE-only differences within seed variation (paired −0.0034±0.0046 at ensemble level), the paper claims the practical accuracy gains come from data expansion and teacher re-fine-tuning, not from distillation itself under this setting.
Similarly, switching the teacher generation from DINOv2-L to DINOv3-L yields no clear ensemble-level improvement (paired difference −0.0021±0.0063), supporting the paper's conclusion not to replace a proven teacher on generational grounds alone.
Half-precision (FP16) deployment of EfficientNet-B0 on a Jetson Orin Nano runs at 3.19 ms/image including transfer (≈313 images/s), a ~9.4× margin against the 30 ms operational budget, with 99.95% argmax agreement with FP32; no INT8 route beat FP16 in both speed and accuracy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper's own leakage caveat implies the reported 0.935 is an upper bound until splits are grouped at observation level and by every video's frame set; re-deriving those groups from the saved manifests is the direct check.
The 'expand the related rare class first' recipe suggests a genus/subfamily-level class-balancing policy: before adding more images of a species that already has enough data, add images of its morphologically closest confusable relative; this is testable on other fine-grained datasets.
Because OOD rejection by confidence threshold is weak (AUROC 0.699, FPR95 77.9%), a deployable collision-avoidance system should treat this six-species classifier as one component behind an upstream detector and ahead of a calibrated open-set/uncertainty gate, not as a standalone decision module.
The finding that distillation gives no clear gain at this scale and task does not generalize automatically; on a larger species set (60–100 classes) or with fewer labels, the teacher's soft labels may matter more, and that is an empirical question.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Dear colleague,

This is a solid applied ML paper, more careful than most in the edge-deployment niche. What's new: a 12,519-image raptor dataset with rare-class expansion via video frames, plus controlled five-seed comparisons showing that knowledge distillation itself and a DINOv2-to-DINOv3 teacher swap both give null ensemble-level results. Those negative findings are genuinely useful, and the authors report them plainly. The TensorRT FP16 latency on a Jetson Orin Nano (3.19 ms/image including transfer) is a concrete, reproducible-in-principle deployment number.

The evaluation discipline is the paper's real strength. Five-seed means and ranges, a group split at video/source-image level, a disjoint subset for old-vs-new comparisons, and explicit statements about what is causal and what isn't. The reduction in white-tailed/Steller's confusion (61% to 15% of errors) is a decision-relevant result, even if it's not a controlled isolation of any single factor.

The main soft spot is exactly what the authors disclose in §4.7: the group split groups only Steller's video IDs and V6 source images. It does not group by iNaturalist observation, and non-V6 video frames aren't grouped either. Combined with a single split seed, the headline 0.935±0.004 could be inflated by near-duplicates straddling train and test. I don't think this is fatal — the core claims (data expansion helps, teacher re-fine-tuning helps, FP16 deployment is fast) survive even if the exact number drops a few points — but the paper should not present 0.935 as the clean unseen-generalization estimate without acknowledging that residual leakage is possible. It does acknowledge it, but the abstract still leads with the number.

The other limitation: no code, data, or weights released. The manifest is described but not shared, so nobody can independently check the leak or the results.

Who gets value: researchers working on fine-grained classification, video-derived datasets, or conservation applications; also anyone thinking about leakage in evaluation splits. It deserves a serious referee. My recommendation: send it to review, and require the authors to release group IDs (or the manifest) and run multiple split seeds. If the number holds under that, it's a credible engineering contribution.

Referee Report

3 major / 4 minor

Summary. This empirical study builds a lightweight six-class raptor classifier for edge deployment. A DINOv2-L teacher is distilled into three students (MobileNetV4, ViT-Small, EfficientNet-B0); the Steller's Sea Eagle class is expanded from 463 to 2,050 images via video-frame extraction, and the teacher is re-fine-tuned on the enlarged set. The paper reports a three-student ensemble macro recall of 0.955 on a conventional image-level split and 0.935 +/- 0.004 under a group split intended to mitigate video/source-image leakage (five distillation seeds, single group split), corresponding to about 97.5% of the teacher's macro recall with roughly 1/8.9 of the teacher's parameters. On a 1,258-image subset disjoint from the former training images, White-tailed Eagle recall improves by up to 38.6 percentage points and misclassification as Steller's Sea Eagle drops from 61% to 15% of errors. TensorRT FP16 deployment of EfficientNet-B0 on an NVIDIA Jetson Orin Nano achieves 3.19 ms/image including host-device transfer, with 99.95% argmax agreement with FP32. Five-seed controlled comparisons find no clear ensemble-level improvement from distillation versus CE-only training, nor from switching the teacher from DINOv2-L to DINOv3-L; the authors attribute the primary gains to data expansion and teacher re-fine-tuning.

Significance. If the reported numbers hold, the paper is a useful applied contribution: it demonstrates a realistic edge-deployment path for fine-grained raptor identification, quantifies hardware latency rather than relying on FLOPs, and reports null results honestly. The evaluation is more careful than is typical in this area, with five-seed variance, a group-split check, a held-out subset disjoint from prior training images, and FP16-versus-FP32 agreement on real hardware. The main limitation is that the primary generalization figure rests on an incomplete group split that the paper itself concedes cannot separate all same-observation or same-video near-duplicates; this makes the headline 0.935 +/- 0.004 an upper-bound-style estimate rather than a verified measure of unseen-data generalization. The causal attribution of gains to data expansion plus teacher re-fine-tuning is also not fully decomposed. These issues are fixable and do not invalidate the practical contribution, but they need to be addressed before the central claims can be accepted.

major comments (3)

[§4.7, Table 10] The primary generalization estimate is not yet established. The group split used for the headline 0.935 +/- 0.004 separates only (i) the 33 Steller's Sea Eagle video IDs and (ii) V6 source-image crops; §4.7 explicitly states that grouping down to the iNaturalist observation unit or to the frame unit of non-V6 videos cannot be recovered from the file names. Since the expansion added thousands of iNaturalist stills and frames from non-V6 videos (Table 3), near-duplicates from the same observation or same video can appear on both sides of the split. §3.1 also notes that cross-route pHash duplicate removal was not performed. The reported +/- 0.004 is only distillation-seed variance; the split itself is a single construction (split seed 42). I therefore do not think the abstract's primary generalization claim is yet supported. I request either (a) a stricter group split at observation/video l
[§4.5, §4.6, §5.1] The causal attribution that the primary gains stem from dataset expansion and teacher re-fine-tuning is not isolated. Table 9 compares the old model (old training data, old teacher) with the new model (new data, re-fine-tuned teacher); both the data and the teacher change simultaneously. The only controlled teacher comparison (Table 6, v2 versus v3) holds the data constant and shows that teacher re-fine-tuning alone can contribute up to +3.24 points for ViT-Small. There is no experiment holding the teacher fixed and varying only the training data. Consequently, the specific claim that data expansion, rather than teacher re-fine-tuning, is responsible for the White-tailed Eagle recall improvement and the 61%-to-15% error change is underdetermined. An ablation with the same new teacher trained on the old and new data would directly support the stated conclusion.
[§4.5] The subset used to claim 'White-tailed Eagle recall improved by up to 38.6 percentage points' is defined by exact match of image relative paths with the old training split, and the paper acknowledges that perceptually similar but different-path near-duplicates are not removed. Given the same-source leakage concern, this 38.6-point gain could be partially inflated by near-duplicate images from the same observation or video appearing on both sides. The paper should at least quantify near-duplicate overlap on this subset (e.g., pHash) or report the result after stricter deduplication. This is closely related to the group-split issue but directly affects a headline contribution.

minor comments (4)

[§4.7] The text says 'All six species maintained an ensemble recall of at least 0.915 (Steller's Sea Eagle 0.919...)' for the single group split, but later in the same section reports a five-seed per-species Steller's Sea Eagle recall of 0.909. Please report per-species ranges across the five seeds so readers can see the variance around the single-seed claim.
[§3.1] The paper states that an image manifest is saved for rights handling and reproducibility, but no manifest or code is provided. Given the leakage concerns, including the manifest (or a redacted version) would be essential for independent verification.
[Abstract and §4.3] 'Roughly one-eighth as many parameters' is acceptable, but 34.13M/304M is 1/8.9; the text could state the exact ratio for consistency.
[§2/§4.4] The phrase 'no clear improvement' in the distillation and teacher-generation comparisons is appropriately cautious. However, with n=5 seeds the comparisons have low power; the paper should avoid implying equivalence and could state explicitly that larger seed counts would be needed to detect small effects.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the headline accuracy is a measured held-out result, and the only self-references are non-load-bearing data-sourcing and system-positioning mentions.

full rationale

This is an empirical evaluation paper rather than a derivation, so the circularity patterns mostly do not apply. The headline macro recall (0.935±0.004 group split, 0.955 image-level) is produced by training and evaluating models on held-out splits, with hyperparameters (T, α, early stopping, checkpoints) selected on validation only: §3.4 states 'T and α were selected for each student by a preliminary sweep on the validation set (the test set was used only for final evaluation)' and §3.5 repeats this. No fitted parameter is renamed as a prediction, and no claim is forced by definition. The only self-referential material is (i) the 'V6' bird-image classifier used as a source of pre-existing video frames ('adding to the set of video frames already prepared for training our existing bird-image classifier ... hereafter V6', §3.1) and (ii) the author's parallel under-review IUCN-detection study used for system positioning (§2.3–2.4). Neither enters the computation of the reported accuracies; V6 is data provenance and the parallel study is contextual. The acknowledged group-split limitation (§4.7: grouping 'down to the iNaturalist observation unit ... or to the frame unit of non-V6 videos cannot be recovered from the file names') is an external-validity caveat about residual source leakage, not a circular step in the paper's own derivation chain. Likewise the 'single group split (split seed 42)' caveat concerns robustness of the estimate, not circularity. There is no quoted equation or construction that reduces a claimed prediction to an input or to an unverified self-citation, so no circular step is identified.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper's outputs are empirical measurements rather than derivations. The central assumptions are about data quality (labels from video creators plus an identification guide) and leakage control (group split at video/source-image granularity is adequate). Free parameters are distillation hyperparameters and preprocessing thresholds tuned on a validation set.

free parameters (4)

Distillation temperature T (MNV4/ViT-Small = 4.0; EffNet-B0 = 2.0) = 4.0 / 4.0 / 2.0
Selected by validation sweep (§3.4, Table 4); affect the soft-target distribution that students learn, and thus a core free choice in the method.
Distillation weight α (MNV4/ViT-Small = 0.7; EffNet-B0 = 0.5) = 0.7 / 0.7 / 0.5
Chosen by validation sweep; balances KD vs CE loss (Eq. 1).
Frame-extraction confidence threshold = 0.7
Detector confidence for accepting video frames (§3.2); hand-chosen, not justified by data.
pHash duplicate threshold = Hamming distance < 8
Used for duplicate removal within video frames and Steller's Sea Eagle augmentation candidates (§3.1, §3.2); hand-chosen.

assumptions (3)

domain assumption Ground-truth labels are determined by the video creator's label plus morphological visual confirmation against Maki's identification guide; discrepancies are discarded.
Assumed correct species identity for all training/test images; label noise would bias all metrics (§3.1, §3.2).
domain assumption Group split at video/source-image granularity is sufficient to reveal the main source leakage.
The paper assumes residual leakage (iNaturalist observation-level, non-V6 video frames) does not substantially inflate the reported 0.935 macro recall (§4.7).
domain assumption Closed-set classification into six species is the relevant task; OOD rejection is explicitly out of scope.
All main metrics assume test images belong to one of the six species; MSP-based OOD rejection is weak (AUROC 0.699) and deferred (§4.11).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Lightweight Image Classification of Raptor Species for Edge Devices: Rare-Species Dataset Expansion via Video Frame Extraction, Knowledge Distillation, and TensorRT Deployment." pith.science (2026). https://pith.science/paper/XPBD7ACG

@misc{pith2026260726238,
  author       = {Pith},
  title        = {Pith review of: Lightweight Image Classification of Raptor Species for Edge Devices: Rare-Species Dataset Expansion via Video Frame Extraction, Knowledge Distillation, and TensorRT Deployment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XPBD7ACG}},
  note         = {Machine review of arXiv:2607.26238}
}

read the original abstract

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.

Figures

Figures reproduced from arXiv: 2607.26238 by the authors.

**Figure 2.** Performance change of the DINOv2-L teacher under dataset expansion and recovery by re-finetuning. 3.3 Re-fine-tuning of the teacher DINOv2-L For the teacher model, we used DINOv2 ViT-L/14 [6, 16]: after loading pre-trained weights via timm [14], we added a classification head for the six-class problem. We used 224 × 224 pixel input, AdamW (backbone learning rate 1×10−5 , head learning rate 1×10−3 , weight decay 0.0… view at source ↗

**Figure 3.** Per-class recall on the subset disjoint from the former training images (old left, new right). are not removed), the condition is unseen for both the old and the new models. The comparison of the three students under this fair condition is shown in [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Confusion matrices of EffNet-B0 on the subset disjoint from the former training images (old left, new right). Red boxes indicate misclassification of White-tailed Eagle as Steller’s Sea Eagle. ID for video frames (33 Steller’s Sea Eagle videos), (ii) source-image level…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 2 canonical work pages

[1]

et al.: Bird collisions at wind turbines in a mountainous area related to bird movement intensities measured by radar,Biol

Aschwanden, J., Stark, H., Peter, D. et al.: Bird collisions at wind turbines in a mountainous area related to bird movement intensities measured by radar,Biol. Conserv., Vol. 220, pp. 228– 236 (2018)

2018
[2]

T., Batalha, H., Rodrigues, S

Marques, A. T., Batalha, H., Rodrigues, S. et al.: Understanding bird collisions at wind farms: An updated review on the causes and possible mitigation strategies,Biol. Conserv., Vol. 179, pp. 40–52 (2014)

2014
[3]

et al.: Mitigating wind-turbine induced avian mortality: Sensory, aerodynamic and cognitive constraints and options,Renew

May, R., Reitan, O., Bevanger, K. et al.: Mitigating wind-turbine induced avian mortality: Sensory, aerodynamic and cognitive constraints and options,Renew. Sustain. Energy Rev., Vol. 42, pp. 170–181 (2015)

2015
[4]

Shiraki, S.: Current status of collisions of White-tailed Eagles (Haliaeetus albicilla) with wind turbines in Hokkaido,Japanese Journal of Conservation Ecology, Vol. 17, No. 1, pp. 85–96 (2012) (in Japanese)

2012
[5]

and Dean, J.: Distilling the knowledge in a neural network, arXiv: 1503.02531 (2015)

Hinton, G., Vinyals, O. and Dean, J.: Distilling the knowledge in a neural network, arXiv: 1503.02531 (2015)

arXiv 2015
[6]

et al.: DINOv2: Learning robust visual features without supervision,Trans

Oquab, M., Darcet, T., Moutakanni, T. et al.: DINOv2: Learning robust visual features without supervision,Trans. Mach. Learn. Res., January (2024)

2024
[7]

and Takada, R.: Differences in flight behavior between White-tailed Eagles and Steller’s Sea Eagles,Bird Research, Vol

Ueda, M., Fukuda, Y. and Takada, R.: Differences in flight behavior between White-tailed Eagles and Steller’s Sea Eagles,Bird Research, Vol. 6, pp. A43–A52 (2010) (in Japanese)

2010
[8]

Maki, K.: Identification Guide to Eagles, Hawks and Falcons, Heibonsha, Tokyo (2012) (in Japanese)

2012

Show all 39 references

[9]

V., Seitzer, M

Sim´ eoni, O., Vo, H. V., Seitzer, M. et al.: DINOv3,Trans. Mach. Learn. Res.(2026), arXiv:2508.10104

2026 arXiv
[10]

Gou, J., Yu, B., Maybank, S. J. and Tao, D.: Knowledge distillation: A survey,Int. J. Comput. Vis., Vol. 129, pp. 1789–1819 (2021). 19

2021
[11]

et al.: Training data-efficient image transformers & distil- lation through attention,Proc

Touvron, H., Cord, M., Douze, M. et al.: Training data-efficient image transformers & distil- lation through attention,Proc. Int. Conf. Mach. Learn. (ICML), PMLR 139, pp. 10347–10357 (2021)

2021
[12]

and Belongie, S.: The Caltech-UCSD Birds-200- 2011 dataset, Technical Report CNS-TR-2011-001, California Institute of Technology (2011)

Wah, C., Branson, S., Welinder, P., Perona, P. and Belongie, S.: The Caltech-UCSD Birds-200- 2011 dataset, Technical Report CNS-TR-2011-001, California Institute of Technology (2011)

2011
[13]

et al.: The iNaturalist species classification and detection dataset,Proc

Van Horn, G., Mac Aodha, O., Song, Y. et al.: The iNaturalist species classification and detection dataset,Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 8769–8778 (2018)

2018
[14]

Wightman, R.: PyTorch Image Models (timm) (online), available athttps://github.com/h uggingface/pytorch-image-models(accessed 2026-05-05)

2026
[15]

et al.: MobileNetV4 – Universal models for the mobile ecosystem,Proc

Qin, D., Leichner, C., Delakis, M. et al.: MobileNetV4 – Universal models for the mobile ecosystem,Proc. Eur. Conf. Comput. Vis. (ECCV), LNCS, Vol. 15098, pp. 78–96 (2024), DOI:10.1007/978-3-031-73661-2 5

2024 doi
[16]

et al.: An image is worth 16x16 words: Transformers for image recognition at scale,Proc

Dosovitskiy, A., Beyer, L., Kolesnikov, A. et al.: An image is worth 16x16 words: Transformers for image recognition at scale,Proc. Int. Conf. Learn. Represent. (ICLR)(2021)

2021
[17]

and Le, Q.: EfficientNet: Rethinking model scaling for convolutional neural networks, Proc

Tan, M. and Le, Q.: EfficientNet: Rethinking model scaling for convolutional neural networks, Proc. Int. Conf. Mach. Learn. (ICML), pp. 6105–6114 (2019)

2019
[18]

NVIDIA Corporation: NVIDIA TensorRT (online), available athttps://developer.nvidia .com/tensorrt(accessed 2026-07-10)

2026
[19]

et al.: Quantization and training of neural networks for efficient integer-arithmetic-only inference,Proc

Jacob, B., Kligys, S., Chen, B. et al.: Quantization and training of neural networks for efficient integer-arithmetic-only inference,Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 2704–2713 (2018)

2018
[20]

et al.: A survey of quantization methods for efficient neural network inference, arXiv: 2103.13630 (2021)

Gholami, A., Kim, S., Dong, Z. et al.: A survey of quantization methods for efficient neural network inference, arXiv: 2103.13630 (2021)

2021 arXiv
[21]

and Sun, G.: Squeeze-and-Excitation networks,Proc

Hu, J., Shen, L. and Sun, G.: Squeeze-and-Excitation networks,Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 7132–7141 (2018)

2018
[22]

Ministry of the Environment, Nature Conservation Bureau: Guidelines for the appropriate siting of wind-power facilities with respect to birds and other wildlife, Ministry of the Envi- ronment, Tokyo (2011) (in Japanese)

2011
[23]

Ministry of the Environment, Nature Conservation Bureau: Guidelines for examining and implementing bird-strike prevention measures for sea eagles at wind-power facilities (revised August 2022), Ministry of the Environment, Tokyo (2022) (in Japanese)

2022
[24]

and Naemura, T.: Construction of a bird image dataset for ecological investigations,Proc

Yoshihashi, R., Kawakami, R., Iida, M. and Naemura, T.: Construction of a bird image dataset for ecological investigations,Proc. IEEE International Conference on Image Processing (ICIP), pp. 4248–4252 (2015)

2015
[25]

and Naemura, T.: Bird detection and species clas- sification with time-lapse images around a wind farm: Dataset construction and evaluation, Wind Energy, Vol

Yoshihashi, R., Kawakami, R., Iida, M. and Naemura, T.: Bird detection and species clas- sification with time-lapse images around a wind farm: Dataset construction and evaluation, Wind Energy, Vol. 20, No. 12, pp. 1983–1995 (2017). 20

1983
[26]

McClure, C. J. W., Rolek, B. W., Dunn, L. et al.: Eagle fatalities are reduced by automated curtailment of wind turbines,J. Appl. Ecol., Vol. 58, No. 3, pp. 446–452 (2021)

2021
[27]

McClure, C. J. W., Rolek, B. W., Dunn, L. et al.: Confirmation that eagle fatalities can be reduced by automated curtailment of wind turbines,Ecol. Solut. Evidence, Vol. 3, No. 3, e12173 (2022)

2022
[28]

IUCN: The IUCN Red List of Threatened Species (online), available athttps://www.iucnre dlist.org/(accessed 2026-07-10)

2026
[29]

iNaturalist contributors, iNaturalist: iNaturalist Research-grade Observations, GBIF Occur- rence dataset (online), available athttps://doi.org/10.15468/ab3s5x(accessed 2026-05- 04)

2026 doi
[30]

gbif.org/(accessed 2026-05-04)

GBIF.org: The Global Biodiversity Information Facility (online), available athttps://www. gbif.org/(accessed 2026-05-04)

2026
[31]

Google: Gemini API (online), available athttps://ai.google.dev/gemini-api/docs (accessed 2026-02-15)

2026
[32]

et al.: PyTorch distributed: experiences on accelerating data parallel training,Proc

Li, S., Zhao, Y., Varma, R. et al.: PyTorch distributed: experiences on accelerating data parallel training,Proc. VLDB Endow., Vol. 13, pp. 3005–3018 (2020)

2020
[33]

et al.: ONNX: Open Neural Network Exchange (online), available athttps://github.com/onnx/onnx(accessed 2026-05-05)

Bai, J., Lu, F., Zhang, K. et al.: ONNX: Open Neural Network Exchange (online), available athttps://github.com/onnx/onnx(accessed 2026-05-05)

2026
[34]

NVIDIA Corporation: NVIDIA TensorRT Model Optimizer (online), available athttps: //github.com/NVIDIA/TensorRT-Model-Optimizer(accessed 2026-07-10)

2026
[35]

Foundation for Computational Science: FOCUS Supercomputer User Guide (online), available athttps://www.j-focus.jp/user_guide/(accessed 2026-07-10) (in Japanese)

2026
[36]

nvidia.com/en-us/data-center/h200/(accessed 2026-07-10)

NVIDIA Corporation: NVIDIA H200 Tensor Core GPU (online), available athttps://www. nvidia.com/en-us/data-center/h200/(accessed 2026-07-10)

2026
[37]

NVIDIA Corporation: NVIDIA Jetson Orin Nano (online), available athttps://www.nv idia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/(accessed 2026-07-10)

2026
[38]

Apple Inc.: MacBook Pro (14-inch, 2023) Technical Specifications (online), available athttps: //support.apple.com/en-us/111340(accessed 2026-07-10)

2023
[39]

Ministry of the Environment: List of Nationally Endangered Species of Wild Fauna and Flora (online), available athttps://www.env.go.jp/nature/kisho/domestic/list.html(ac- cessed 2026-07-25) (in Japanese). 21

2026

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

et al.: Bird collisions at wind turbines in a mountainous area related to bird movement intensities measured by radar,Biol

Aschwanden, J., Stark, H., Peter, D. et al.: Bird collisions at wind turbines in a mountainous area related to bird movement intensities measured by radar,Biol. Conserv., Vol. 220, pp. 228– 236 (2018)

2018

[2] [2]

T., Batalha, H., Rodrigues, S

Marques, A. T., Batalha, H., Rodrigues, S. et al.: Understanding bird collisions at wind farms: An updated review on the causes and possible mitigation strategies,Biol. Conserv., Vol. 179, pp. 40–52 (2014)

2014

[3] [3]

et al.: Mitigating wind-turbine induced avian mortality: Sensory, aerodynamic and cognitive constraints and options,Renew

May, R., Reitan, O., Bevanger, K. et al.: Mitigating wind-turbine induced avian mortality: Sensory, aerodynamic and cognitive constraints and options,Renew. Sustain. Energy Rev., Vol. 42, pp. 170–181 (2015)

2015

[4] [4]

Shiraki, S.: Current status of collisions of White-tailed Eagles (Haliaeetus albicilla) with wind turbines in Hokkaido,Japanese Journal of Conservation Ecology, Vol. 17, No. 1, pp. 85–96 (2012) (in Japanese)

2012

[5] [5]

and Dean, J.: Distilling the knowledge in a neural network, arXiv: 1503.02531 (2015)

Hinton, G., Vinyals, O. and Dean, J.: Distilling the knowledge in a neural network, arXiv: 1503.02531 (2015)

arXiv 2015

[6] [6]

et al.: DINOv2: Learning robust visual features without supervision,Trans

Oquab, M., Darcet, T., Moutakanni, T. et al.: DINOv2: Learning robust visual features without supervision,Trans. Mach. Learn. Res., January (2024)

2024

[7] [7]

and Takada, R.: Differences in flight behavior between White-tailed Eagles and Steller’s Sea Eagles,Bird Research, Vol

Ueda, M., Fukuda, Y. and Takada, R.: Differences in flight behavior between White-tailed Eagles and Steller’s Sea Eagles,Bird Research, Vol. 6, pp. A43–A52 (2010) (in Japanese)

2010

[8] [8]

Maki, K.: Identification Guide to Eagles, Hawks and Falcons, Heibonsha, Tokyo (2012) (in Japanese)

2012