NanoDet-Plus

Super fast and high accuracy lightweight anchor-free object detection model. Real-time on mobile devices.

⚡Super lightweight: Model file is only 980KB(INT8) or 1.8MB(FP16).
⚡Super fast: 97fps(10.23ms) on mobile ARM CPU.
👍High accuracy: Up to 34.3 mAP^val@0.5:0.95 and still realtime on CPU.
🤗Training friendly: Much lower GPU memory cost than other models. Batch-size=80 is available on GTX1060 6G.
😎Easy to deploy: Support various backends including ncnn, MNN and OpenVINO. Also provide Android demo based on ncnn inference framework.

Introduction

NanoDet is a FCOS-style one-stage anchor-free object detection model which using Generalized Focal Loss as classification and regression loss.

In NanoDet-Plus, we propose a novel label assignment strategy with a simple assign guidance module (AGM) and a dynamic soft label assigner (DSLA) to solve the optimal label assignment problem in lightweight model training. We also introduce a light feature pyramid called Ghost-PAN to enhance multi-layer feature fusion. These improvements boost previous NanoDet's detection accuracy by 7 mAP on COCO dataset.

NanoDet-Plus 知乎中文介绍

NanoDet 知乎中文介绍

QQ交流群：908606542 (答案：炼丹)

Benchmarks

Model	Resolution	mAP^val 0.5:0.95	CPU Latency^(i7-8700)	ARM Latency^(4xA76)	FLOPS	Params	Model Size
NanoDet-m	320*320	20.6	4.98ms	10.23ms	0.72G	0.95M	1.8MB(FP16) \| 980KB(INT8)
NanoDet-Plus-m	320*320	27.0	5.25ms	11.97ms	0.9G	1.17M	2.3MB(FP16) \| 1.2MB(INT8)
NanoDet-Plus-m	416*416	30.4	8.32ms	19.77ms	1.52G	1.17M	2.3MB(FP16) \| 1.2MB(INT8)
NanoDet-Plus-m-1.5x	320*320	29.9	7.21ms	15.90ms	1.75G	2.44M	4.7MB(FP16) \| 2.3MB(INT8)
NanoDet-Plus-m-1.5x	416*416	34.1	11.50ms	25.49ms	2.97G	2.44M	4.7MB(FP16) \| 2.3MB(INT8)
YOLOv3-Tiny	416*416	16.6	-	37.6ms	5.62G	8.86M	33.7MB
YOLOv4-Tiny	416*416	21.7	-	32.81ms	6.96G	6.06M	23.0MB
YOLOX-Nano	416*416	25.8	-	23.08ms	1.08G	0.91M	1.8MB(FP16)
YOLOv5-n	640*640	28.4	-	44.39ms	4.5G	1.9M	3.8MB(FP16)
FBNetV5	320*640	30.4	-	-	1.8G	-	-
MobileDet	320*320	25.6	-	-	0.9G	-	-

Download pre-trained models and find more models in Model Zoo or in Release Files

Notes (click to expand)

ARM Performance is measured on Kirin 980(4xA76+4xA55) ARM CPU based on ncnn. You can test latency on your phone with ncnn_android_benchmark.
Intel CPU Performance is measured Intel Core-i7-8700 based on OpenVINO.
NanoDet mAP(0.5:0.95) is validated on COCO val2017 dataset with no testing time augmentation.
YOLOv3&YOLOv4 mAP refers from Scaled-YOLOv4: Scaling Cross Stage Partial Network.

NEWS!!!

[2022.08.26] Upgrade to pytorch-lightning-1.7. The minimum PyTorch version is upgraded to 1.9. To use previous version of PyTorch, please install NanoDet <= v1.0.0-alpha-1
[2021.12.25] NanoDet-Plus release! Adding AGM(Assign Guidance Module) & DSLA(Dynamic Soft Label Assigner) to improve 7 mAP with only a little cost.

Find more update notes in Update notes.

Demo

Android demo

Android demo project is in demo_android_ncnn folder. Please refer to Android demo guide.

Here is a better implementation 👉 ncnn-android-nanodet

NCNN C++ demo

C++ demo based on ncnn is in demo_ncnn folder. Please refer to Cpp demo guide.

MNN demo

Inference using Alibaba's MNN framework is in demo_mnn folder. Please refer to MNN demo guide.

OpenVINO demo

Inference using OpenVINO is in demo_openvino folder. Please refer to OpenVINO demo guide.

Web browser demo

https://nihui.github.io/ncnn-webassembly-nanodet/

Pytorch demo

First, install requirements and setup NanoDet following installation guide. Then download COCO pretrain weight from here

👉COCO pretrain checkpoint

The pre-trained weight was trained by the config config/nanodet-plus-m_416.yml.

Inference images

python demo/demo.py image --config CONFIG_PATH --model MODEL_PATH --path IMAGE_PATH

Inference video

python demo/demo.py video --config CONFIG_PATH --model MODEL_PATH --path VIDEO_PATH

Inference webcam

python demo/demo.py webcam --config CONFIG_PATH --model MODEL_PATH --camid YOUR_CAMERA_ID

Besides, We provide a notebook here to demonstrate how to make it work with PyTorch.

Install

Requirements

Linux or MacOS
CUDA >= 10.0
Python >= 3.6
Pytorch >= 1.9
experimental support Windows (Notice: Windows not support distributed training before pytorch1.7)

Step

Create a conda virtual environment and then activate it.

 conda create -n nanodet python=3.8 -y
 conda activate nanodet

Install pytorch

conda install pytorch torchvision cudatoolkit=11.1 -c pytorch -c conda-forge

Clone this repository

git clone https://github.com/RangiLyu/nanodet.git
cd nanodet

Install requirements

pip install -r requirements.txt

Setup NanoDet

python setup.py develop

Model Zoo

NanoDet supports variety of backbones. Go to the config folder to see the sample training config files.

Model	Backbone	Resolution	COCO mAP	FLOPS	Params	Pre-train weight
NanoDet-m	ShuffleNetV2 1.0x	320*320	20.6	0.72G	0.95M	Download
NanoDet-Plus-m-320 (NEW)	ShuffleNetV2 1.0x	320*320	27.0	0.9G	1.17M	Weight \| Checkpoint
NanoDet-Plus-m-416 (NEW)	ShuffleNetV2 1.0x	416*416	30.4	1.52G	1.17M	Weight \| Checkpoint
NanoDet-Plus-m-1.5x-320 (NEW)	ShuffleNetV2 1.5x	320*320	29.9	1.75G	2.44M	Weight \| Checkpoint
NanoDet-Plus-m-1.5x-416 (NEW)	ShuffleNetV2 1.5x	416*416	34.1	2.97G	2.44M	Weight \| Checkpoint

Notice: The difference between Weight and Checkpoint is the weight only provide params in inference time, but the checkpoint contains training time params.

Legacy Model Zoo

Model	Backbone	Resolution	COCO mAP	FLOPS	Params	Pre-train weight
NanoDet-m-416	ShuffleNetV2 1.0x	416*416	23.5	1.2G	0.95M	Download
NanoDet-m-1.5x	ShuffleNetV2 1.5x	320*320	23.5	1.44G	2.08M	Download
NanoDet-m-1.5x-416	ShuffleNetV2 1.5x	416*416	26.8	2.42G	2.08M	Download
NanoDet-m-0.5x	ShuffleNetV2 0.5x	320*320	13.5	0.3G	0.28M	Download
NanoDet-t	ShuffleNetV2 1.0x	320*320	21.7	0.96G	1.36M	Download
NanoDet-g	Custom CSP Net	416*416	22.9	4.2G	3.81M	Download
NanoDet-EfficientLite	EfficientNet-Lite0	320*320	24.7	1.72G	3.11M	Download
NanoDet-EfficientLite	EfficientNet-Lite1	416*416	30.3	4.06G	4.01M	Download
NanoDet-EfficientLite	EfficientNet-Lite2	512*512	32.6	7.12G	4.71M	Download
NanoDet-RepVGG	RepVGG-A0	416*416	27.8	11.3G	6.75M	Download

How to Train

Prepare dataset

If your dataset annotations are pascal voc xml format, refer to config/nanodet_custom_xml_dataset.yml

Or convert your dataset annotations to MS COCO format(COCO annotation format details).
Prepare config file

Copy and modify an example yml config file in config/ folder.

Change save_dir to where you want to save model.

Change num_classes in model->arch->head.

Change image path and annotation path in both data->train and data->val.

Set gpu ids, num workers and batch size in device to fit your device.

Set total_epochs, lr and lr_schedule according to your dataset and batchsize.

If you want to modify network, data augmentation or other things, please refer to Config File Detail
Start training

NanoDet is now using pytorch lightning for training.

For both single-GPU or multiple-GPUs, run:
```
python tools/train.py CONFIG_FILE_PATH
```
Visualize Logs

TensorBoard logs are saved in save_dir which you set in config file.

To visualize tensorboard logs, run:
```
cd <YOUR_SAVE_DIR>
tensorboard --logdir ./
```

How to Deploy

NanoDet provide multi-backend C++ demo including ncnn, OpenVINO and MNN. There is also an Android demo based on ncnn library.

Export model to ONNX

To convert NanoDet pytorch model to ncnn, you can choose this way: pytorch->onnx->ncnn

To export onnx model, run tools/export_onnx.py.

python tools/export_onnx.py --cfg_path ${CONFIG_PATH} --model_path ${PYTORCH_MODEL_PATH}

Run NanoDet in C++ with inference libraries

ncnn

Please refer to demo_ncnn.

OpenVINO

Please refer to demo_openvino.

MNN

Please refer to demo_mnn.

Run NanoDet on Android

Please refer to android_demo.

Citation

If you find this project useful in your research, please consider cite:

@misc{=nanodet,
    title={NanoDet-Plus: Super fast and high accuracy lightweight anchor-free object detection model.},
    author={RangiLyu},
    howpublished = {\url{https://github.com/RangiLyu/nanodet}},
    year={2021}
}