YOLO on a diet: 250 MFLOPs and a 1.4 MB footprint

What it does

Yolo-Fastest is a stripped-down YOLO variant built for severely resource-constrained hardware. It swaps accuracy for speed and size, targeting Raspberry Pi 3B boards, low-end ARM mobiles, and embedded Linux devices. The repo provides Darknet-based training configs and points you toward NCNN for actual deployment.

The interesting bit

The authors patched Darknet’s grouped-convolution implementation to fix a 10× inference slowdown on older GPUs like the GTX 1050 Ti. More importantly, they tuned the architecture specifically for NCNN’s ARM inference path—Darknet CPU performance is explicitly labeled as poor, making this a Darknet training pipeline that expects to be deployed through NCNN.

Key highlights

• Base model hits 24.4% COCO mAP at 320×320 using just 0.252B FLOPs and a 1.4 MB weight file; the XL variant reaches 34.3% at 3.7 MB.
• Runs at roughly 16 fps on a Raspberry Pi 3B and under 6 ms per frame on a Snapdragon 888 (NCNN, multi-core).
• Includes a dedicated pedestrian-detection fine-tune.
• Supports NCNN, MNN, ONNX/TensorRT, and OpenCV DNN export.

Caveats

• The authors warn this is only for “simple single object detection” and “simple application scenarios,” so expect plenty of false negatives.
• Darknet CPU inference is explicitly not recommended; you will need to convert models to NCNN or another framework.
• Yolo-FastestV2 exists in a separate repository, suggesting this V1 branch is effectively frozen.

Verdict

Reach for this if you are squeezing vision onto a Pi or an old Android handset and can tolerate modest accuracy. Look elsewhere if you need modern mAP scores or are deploying on anything with a decent CUDA core count.