Practical overview
When it fits
Use the repository when evaluating compact YOLO-based detection models for mobile or embedded inference, or when the documented Nano-to-Caffe sample and ncnn landmark example fit the work. Review its stated GPU caveats and project-reported benchmarks before selecting a route.
Who it is for
- Developers building mobile or embedded computer-vision applications.
- Teams evaluating compact object or face detection models for resource-constrained devices.
- Computer-vision developers adapting pretrained models or exploring facial-landmark inference.
Documented uses
- Add lightweight object detection to a mobile or embedded vision pipeline.
- Select among Lite and Nano variants according to the project’s stated target-resource guidance.
- Add a compact face detector or use the ncnn C sample as a starting point for 106-point facial landmarks.
- Convert the documented Nano VOC model to Caffe and run the included Python detector sample.
Requirements & limits
The project is available as a public GitHub repository and is intended for self-hosted use.
- The README recommends Lite for higher-performance ARM CPU and mobile-GPU use, Nano for limited ARM CPU resources, and SPP for several embedded targets.
- The repository lists ncnn and MNN for mobile inference; its README says the YoloFace-500k-v2 model temporarily does not support MNN.
- Benchmark figures are configuration-specific and reported by the project. MNN timings cover forward inference only; ncnn timings include post-processing.
- The README warns about Darknet group-convolution support on some NVIDIA Pascal GPUs and identifies a PyTorch YOLOv3 route for certain issues.
- Development stopped in 2021; the README points to Yolo-Fastest as its successor.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Convert a Nano model to Caffe and run the documented detector sample
Convert the MobileNetV2-YOLOv3-Nano VOC weights and run the README’s Python detector sample.
Before you start
- Python 2.7
- python-opencv
- Caffe with the added upsample layer, compiled for CPU
- The repository’s darknet2caffe files and Nano VOC config and weights
Steps
From darknet2caffe/, run darknet2caffe.py with the Nano VOC config and weights to produce a prototxt and caffemodel.
Copy the generated prototxt and caffemodel into sample/.
From sample/, run python detector.py.
Expected outcome
The documented sample detector runs using the converted Nano VOC model.
Guide limitations
- The README specifies a CPU Caffe build and Python 2.7 environment.
- The instructions cover this Nano VOC conversion and sample, not a general deployment procedure.
Prepare a Darknet model for MNN conversion
Apply the README’s stated intermediate conversion and layer-editing steps before MNN model conversion.
Before you start
- A Darknet model
- darknet2caffe conversion
- An editable generated prototxt
Steps
Convert the Darknet model to a Caffe model with darknet2caffe.
In the prototxt, replace the documented Upsample layer with an Interp layer and set its height and width.
Continue with the MNN model-conversion process.
Expected outcome
A Caffe-format model and edited prototxt prepared for the MNN conversion stage.
Guide limitations
- The README links to separate MNN conversion instructions; this recipe does not specify their full steps.
- The example gives dimensions for one Nano prototxt and does not establish settings for other models.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 17 October 2026
- Original website title
- dog-qiuqiu/MobileNet-Yolo · Official homepage title · retained in its source language