furiosa-ai/smart-cctv-demo

Setup

sudo apt install cmake
sudo apt install \
  furiosa-libhal-warboy=0.9.0-2+nightly-220810 \
  libonnxruntime=1.11.1-2 \
  furiosa-libnux=0.8.0-2+nightly-220810 \
  furiosa-libcompiler=0.8.0-2+nightly-220803
conda env create -n "cctv" python=3.8
conda activate cctv
conda install cython
conda env update --file environment.yml
bash build.sh

Datasets

• Download CrowdHuman from here and extract to "data/CrowdHuman/"
• Download PRW from here and extract to "data/"
• Download WIDERFACE from here and extract to "data/"
• Download MSM1-RetinaFace from here and extract to "data/"
• market1501 dataset will be downloaded automatically at progam execution

Test files (face recognition)

• Download video and move to "data/tom_cruise_test.mp4"
• Download query and move to "data/tom_cruise_profile1.jpeg"

Weights

• Download reid checkpoint from here and move to weights/
• Download remaining checkpoints from here and extract to weights/

Final directory structure

smart-cctv-demo$ tree -L 2 data weights
data
├── CrowdHuman
│   └── Images
├── PRW-v16.04.20
│   ├── ID_test.mat
│   ├── ID_train.mat
│   ├── annotations
│   ├── frame_test.mat
│   ├── frame_train.mat
│   ├── frames
│   ├── generate_query.m
│   ├── query_box
│   ├── query_info.txt
│   └── readme.txt
├── WIDER_val
│   └── images
├── market1501
│   ├── Market-1501-v15.09.15
│   └── Market-1501-v15.09.15.zip
├── ms1m-retinaface-t1
│   ├── agedb_30.bin
│   ├── cfp_fp.bin
│   ├── lfw.bin
│   ├── property
│   ├── train.idx
│   ├── train.lst
│   └── train.rec
├── tom_cruise_profile1.jpeg
└── tom_cruise_test.mp4
weights
├── ms1mv3_r50_leakyrelu_1.pt
├── resnet50_market_xent.pth.tar
├── yolov5_face_relu.pt
└── yolov5m_warboy.pt

Query demo

# Optionally clear cache
rm *.onnx  # f32 model cache
rm *.dfg  # i8 model cache
rm -rf galleries/*  # gallery cache
rm -rf results/*  # delete previous results

# Face recognition (results will be written to "results/face")
cd insightface
python recognition/arcface_torch/query_video.py \
  --gallery ../data/tom_cruise_test.mp4 \
  --query ../data/tom_cruise_profile1.jpeg \
  --device furiosa

# ReId (results will be written to "results/reid")
cd torchreid
python query_video.py \
  --gallery ../data/PRW-v16.04.20/frames \
  --query ../data/PRW-v16.04.20/query_box/479_c1s3_016471.jpg \
  --device furiosa

GUI

• Local inference

  python demo.py

• Server inference

  Add to ~/.ssh/config

  Host vis
    HostName 210.109.63.237
    User ubuntu
    IdentityFile ~/.ssh/vision-intelligence.pem
    IdentitiesOnly yes
  

  Run

  python demo.py --server cfg/server_vis.yaml

Demo

Face recognition

Person recognition