Verdict

Only one model reaches 30 FPS on a bare Raspberry Pi 5: DEIMv2-Atto, at 30.3 FPS and 27.5 mAP with a 320 px input. At 10 FPS you can run DEIMv2-Femto, which gets 34.5 mAP at 16.7 FPS. Past that, accuracy costs real time: DEIM-N reaches 46.8 mAP but runs at 4.3 FPS. All numbers are ONNX Runtime FP32 on the four CPU cores, mAP@50-95 in percent.

Fastest models on the Pi 5

#ModelmAP@50-95FPSms/imageInput (px)Params (M)License
1DEIMv2-Atto27.530.333.023200.5Apache-2.0
2YOLOX-Nano27.721.346.984160.9Apache-2.0
3DEIMv2-Femto34.516.759.794161.0Apache-2.0
4PicoDet-S28.115.066.773201.0Apache-2.0
5YOLOX-Tiny34.39.4106.694165.0Apache-2.0
6YOLO26-N40.17.1140.636402.4AGPL-3.0
7YOLOv5nu33.86.4156.646402.6AGPL-3.0
8PicoDet-M36.56.2160.154162.1Apache-2.0
9DEIMv2-Pico42.26.2161.536401.5Apache-2.0
10YOLO11-N38.76.2162.226402.6AGPL-3.0
11YOLOv9-T40.75.9169.976402.0MIT
12YOLOv8-N36.75.9170.366403.2AGPL-3.0
13DEIMv2-N46.74.4226.266403.6Apache-2.0
14DEIM-N46.84.3229.716403.8Apache-2.0
15D-FINE-N45.84.3229.686403.8Apache-2.0
Fastest detectors on Raspberry Pi 5, ONNX Runtime FP32, batch 1. Check the input column: the top rows run on smaller images.

Check the input size before you pick from the top of this table. DEIMv2-Atto and PicoDet-S run at 320 px. YOLOX-Nano, YOLOX-Tiny, DEIMv2-Femto and PicoDet-M run at 416 px. RF-DETR runs at its own sizes, from 384 to 704 px. The rest run at 640 px. Smaller inputs are part of why those models are fast, and they lose small objects. If your targets are small in the frame, compare at 640 px: there YOLO26-N is fastest at 7.1 FPS and 40.1 mAP (AGPL-3.0), and DEIMv2-Pico gets 42.2 mAP at 6.2 FPS under Apache-2.0.

Try NCNN before you settle

These rows use ONNX Runtime. On the Pi 5, NCNN runs the same models faster, and the gap grows with model size. Compare both in NCNN vs ONNX Runtime on Raspberry Pi 5. If you need more than 30 FPS at useful accuracy, you need an accelerator: see Raspberry Pi 5 with a Hailo-8.

Most accurate models

#ModelmAP@50-95FPSms/imageInput (px)Params (M)License
1D-FINE-X61.40.42638.7564061.7Apache-2.0
2DEIMv2-X61.30.33453.0364050.3Apache-2.0
3ECDet-X61.10.33331.4264049.0Apache-2.0
4ECDet-L60.10.42681.8064032.7Apache-2.0
5RT-DETRv4-X60.00.42685.6864062.6Apache-2.0
6D-FINE-L60.00.71402.9164030.8Apache-2.0
7DEIM-X59.60.42635.5664061.7Apache-2.0
8RF-DETR-L58.60.51924.5070430.3Apache-2.0
9DEIMv2-L58.60.42678.9864032.2Apache-2.0
10ECDet-M58.40.61708.1964019.2Apache-2.0
Most accurate detectors on Raspberry Pi 5, ONNX Runtime FP32, batch 1.

D-FINE-X tops this table at 61.4 mAP and takes 2639 ms per image (0.4 FPS). The top ten all run below 1 FPS. They fit offline jobs, like labelling a folder of images overnight, not a camera feed.

Every number on this page comes from the verified dataset: same 500-image COCO val2017 slice, conf 0.001, IoU 0.6, max 300 detections, pycocotools mAP, identical protocol across all hardware and runtimes. The full protocol is on the methodology page. To rerun this comparison with your own filters, open compare. Accuracy is measured on LibreYOLO retrained checkpoints; other weight sources can yield different values.