TensorRT raptor classifier runs at 313 FPS on Jetson Orin Nano

A new computer vision system for real-time raptor identification on edge devices promises to reduce bird collisions with wind turbines. Researcher Takeshi Nishikawa trained a lightweight image classifier using knowledge distillation from a large DINOv2-L teacher model (304M parameters) into three smaller students: MobileNetV4, ViT-Small, and EfficientNet-B0. A key innovation was dataset expansion via video-frame extraction, boosting Steller's Sea Eagle images from 463 to 2,050 to reduce confusion with the similar White-tailed Eagle.

The ensemble of three lightweight models achieved 0.935 macro recall on a strict group split (0.955 on conventional splits), retaining 97.5% of the teacher's accuracy with roughly one-eighth the parameters. White-tailed Eagle recall improved by up to 38.6 percentage points, and misclassification as Steller's Sea Eagle dropped from 61% to 15% of errors. Deploying EfficientNet-B0 with TensorRT FP16 on an NVIDIA Jetson Orin Nano reached 3.19 ms per image (313 images/s) with 99.95% FP32 agreement. The study notes that the primary performance gains came from dataset expansion and teacher fine-tuning, not from the distillation or teacher model choice.