Skip to content

Latest commit

 

History

History
108 lines (85 loc) · 3.79 KB

face_detection.md

File metadata and controls

108 lines (85 loc) · 3.79 KB
Raw

Real time face detector

This repository contains training scripts for lightweight SSD-based face detector. The detector is based on the MobileNetV2 backbone and has single SSD head with manually designed anchors. As a result it has computational complexity 0.51 GMACs and 1.03 M of parameters.

Prerequisites

  1. Download the WIDER Face and unpack it to data folder.
  2. Annotation in the VOC format can be found in this repo. Move the annotation files from WIDER_train_annotations and WIDER_val_annotations folders to the Annotation folders inside the corresponding directories WIDER_train and WIDER_val. Also annotation lists val.txt and train.txt should be copied to data/WIDERFace from WIDER_train_annotations and WIDER_val_annotations. The directory should be like this:
data
└── WIDERFace
    ├── WIDER_train
    │   ├──0--Parade
    │   ├── ...
    │   └── Annotations
    ├── WIDER_val
    │   ├──0--Parade
    │   ├── ...
    │   └── Annotations
    ├── val.txt
    └── train.txt

Training

  1. Download pre-trained MobileNetV2 weights mobilenet_v2.pth.tar from: https://github.com/tonylins/pytorch-mobilenet-v2. Move the file with weights to the folder snapshots. Or checkpoint that was trained on Wider.
  2. To train the detector on a single GPU run in terminal:
python3 ../../external/mmdetection/tools/train.py \
configs/mobilenetv2_tiny_ssd300_wider_face.py

Validation

  1. To dumb detection of your model:
python3 ../../external/mmdetection/tools/test.py    \
  configs/mobilenetv2_tiny_ssd300_wider_face.py   \
  <CHECKPOINT>   \
  --out result.pkl
  1. Then run
python3 ../../external/mmdetection/tools/voc_eval.py    \
  result.pkl    \
  configs/mobilenetv2_tiny_ssd300_wider_face.py

One should observe 0.305 AP on validation set. For more detailed results and comparison with vanilla SSD300 see ../../external/mmdetection/configs/wider_face/README.md.

Conversion to OpenVINO format

  1. Convert PyTorch model to ONNX format: run script in terminal
python3 tools/onnx_export.py \
      configs/mobilenetv2_tiny_ssd300_wider_face.py
      <CHECKPOINT> \
      face_detector.onnx
  1. Convert ONNX model to OpenVINO format with Model Optimizer: run in terminal
mo.py --input_model face_detector.onnx \
  --scale 255 \
  --reverse_input_channels \
  --output_dir=./IR \
  --data_type=FP32

This produces model face_detector.xml and weights face_detector.bin in single-precision floating-point format (FP32). The obtained model expects normalized image in planar BGR format.

Python demo

To run the demo connect a webcam end execute command:

python3 tools/detection_live_demo.py  \
  configs/mobilenetv2_tiny_ssd300_wider_face.py \
  <CHECKPOINT> \
  --cam_id 0

Estimate theoretical computational complexity

To get per layer computational complexity estimations run the following command:

python3 tools/count_flops.py configs/mobilenetv2_tiny_ssd300_wider_face.py

Fine-tuning

  • Dataset should have the same data layout as WIDER Face in VOC format described in this instruction.
  • Fine-tuning steps are the same as step 2 for training, but some adjustments in config are needed:
    • specify initial checkpoint containing a valid detector in load_from field of config configs/mobilenetv2_tiny_ssd300_wider_face.py
    • edit data section of config to pass a custom dataset.