🚀SparseInst🚀 NCNN Deployment

Introduction

This repo is based on SparseInst. SparseInst is a fast and high precise instance segmemation alogrithm.

Preview

Usage

Compile NCNN

You can reference ncnn_docs. There is for linux backend.

1. clone repo

git clone https://github.com/Tencent/ncnn.git
cd ncnn
git submodule update --init

2. install dependencies

sudo apt install build-essential git cmake libprotobuf-dev protobuf-compiler libvulkan-dev vulkan-utils libopencv-dev

3. build

mkdir -p build && cd build
cmake -DNCNN_VULKAN=ON -DNCNN_BUILD_EXAMPLES=ON -DNCNN_PYTHON=ON -DNCNN_BUILD_TESTS=ON ..
make -j$(proc)
make install

Compile Repo

git clone https://github.com/DDGRCF/sparseinst_ncnn_demo.git
cd sparseinst_ncnn_demo
mkdir -p build && cd build
cmake .. -DNCNN_DIR=/path/to/ncnn/build/install/lib/cmake/ncnn (option)[-DNCNN_PROFILING=ON]
make -j$(nproc)
# Open NCNN_PROFILING option will output time cost and detect image information.

Download Model

This repo provide sparseinst_r50_giam_soft.yaml and sparseinst_inst_cspdarknet53_giam.yaml ncnn converted model. For Resnet50, you can download by following commands:

latest_version=v1.0.0
wget https://github.com/DDGRCF/sparseinst_ncnn_demo/releases/download/${latest_version}/sparseinst-resnet-sim-opt.param
wget https://github.com/DDGRCF/sparseinst_ncnn_demo/releases/download/${latest_version}/sparseinst-resnet-sim-opt.bin

Run Repo

param_path=/path/to/your/param
bin_path=/path/to/your/bin
image_path=/path/to/your/image
save_path=/path/to/your/save_image
./sparseinst_ncnn_demo ${param_path} ${bin_path} ${image_path} ${save_path}
# NOTE: image_path can be dir or file.

About Performance

I compare the ncnn model with other frame (onnxruntime, mnn), results as following:

backend	inference time	remark
ncnn	0.5501 s	not starting vulkan
mnn	0.5492 s	none
onnxruntime	0.9103 s	not using gpu version

Device: CPU: 20 12th Gen Intel(R) Core(TM) i7-12700H.

Extra: Test run 100 times(avg) and did't include data preprocess and data postprocess. For convenience, both of them used python api.

License

This repo is under MIT LICENSE

Thanks

• ncnn
• sparseinst