analiza aiqbench
This study demonstrates effective techniques for deploying lightweight computer vision models on edge devices to solve a specific environmental problem: preventing bird collisions with wind turbines. By combining video frame extraction, knowledge distillation from DINOv2-L, and TensorRT optimization, the researchers achieved high recall rates while significantly reducing model parameter counts compared to standard approaches. The work highlights practical trade-offs where dataset expansion proved more critical than switching teacher models or using advanced distillation methods for this particular application.
Lightweight Image Classification of Raptor Species for Edge Devices: Rare-Species Dataset Expansion via Video Frame Extraction, Knowledge Distillation, and TensorRT Deployment
arXiv:2607.26238v1 Announce Type: cross 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.
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.
21 pages, 4 figures, 14 tables. English translation of a paper submitted to IPSJ (in Japanese); the Japanese version is the source of record
Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG); Image and Video Processing (eess.IV)
I.4.9; I.5.4; I.2.6
arXiv:2607.26238 [cs.CV]
(or arXiv:2607.26238v1 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2607.26238 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 345 cuvinte
Subiecte conexe
EcoFair: Energy-Efficient Inference Routing for Edge AI under Data Degradation
arXiv · 9 sept. 2026
Radio-Frequency Convolutional Neural Networks
arXiv · 18 sept. 2026
Build a Physical AI model factory with NVIDIA Cosmos 3 on SageMaker HyperPod
AWS · 4 sept. 2026
Beyond Benchmarks: Continuous Edge Inference for Fine-Grained Roadside Perception
arXiv · 17 iun. 2026