YOLOv5 on a crash diet: 900 KB and a Raspberry Pi

What it does

YOLOv5-Lite is a family of slimmed-down object detectors forked from YOLOv5. It swaps in lighter backbones—ShuffleNetV2, PPLcNet, and RepVGG—and trims the detection head to shrink FLOPs and memory. The project ships a model zoo that covers nearly every edge inference runtime imaginable, from ncnn and MNN to TensorRT and AXera NPU.

The interesting bit

The authors treat quantization pain as an architectural bug. They removed the Focus layer and its slice operations because those ops degrade int8 accuracy, then published pre-quantized models as small as 900 KB. It reads less like a research paper and more like a deployment checklist with weights attached.

Key highlights

• Four tiers (v5lite-e through v5lite-g) spanning 0.78M to 5.39M parameters; the smallest variant clocks 0.73G FLOPs at 320×320
• Pre-exported weights for ncnn, MNN, TNN, ONNX Runtime, OpenVINO, TensorRT, TFLite, and AXera NPU
• Benchmarks on COCO claim 35.1 mAP@0.5 for the 1.7 MB v5lite-e model and 57.6 mAP@0.5 for the 10.9 MB v5lite-g model
• Latency tables cover x86 CPUs, Android phones, NVIDIA GPUs, and the Raspberry Pi 4B; the Pi 4B manages 71–84 ms per frame at 320×320 depending on the framework
• Includes a heatmap analysis tool and an auto-annotation helper linked via a separate repo

Caveats

• Documentation leans heavily on Chinese-language resources: Baidu Drive downloads, a Zhihu tutorial for the Pi, and a QQ support group
• The accuracy comparison table omits mAP@.5:.95 for the smallest v5lite-e variant and leaves gaps for other models
• The README is truncated mid-sentence in the heatmap section, suggesting the file is incomplete or maintenance has slowed

Verdict

Grab this if you need to ship object detection on ARM CPUs, low-end Android, or Raspberry Pi hardware and you want the quantization work done for you. Skip it if you need the full accuracy of a large YOLOv5 model or an English-first support experience.