Skip to content

Commit b8b17bb

Browse files
sovrasovAlexanderDokuchaev
sovrasov
authored and
AlexanderDokuchaev
committed
Add instructions for face detection (openvinotoolkit#97)
* Add instructions for face detection * Update fr readme and init script * Update mmdetection submodule * Update mmdetection submodule
1 parent 797588f commit b8b17bb

File tree

4 files changed

+117
-4
lines changed

4 files changed

+117
-4
lines changed

‎external/mmdetection

‎pytorch_toolkit/object_detection/README_face_detection.md

+111
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,111 @@
1+
# Real time face detector
2+
3+
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.
4+
5+
## Requirements
6+
7+
* Ubuntu 16.04
8+
* Python 3.5 or 3.6 (3.6 is preferable)
9+
* PyTorch 1.1
10+
11+
## Prerequisites
12+
13+
1. Download mmdetection submodule `git submodule update --init --recommend-shallow external/mmdetection`
14+
2. Download the [WIDER Face](http://shuoyang1213.me/WIDERFACE/) and unpack it to `data` folder.
15+
3. Annotation in the VOC format
16+
can be found in this [repo](https://github.com/sovrasov/wider-face-pascal-voc-annotations.git). Move the annotation files from `WIDER_train_annotations` and `WIDER_val_annotations` folders
17+
to the `Annotation` folders inside the corresponding directories `WIDER_train` and `WIDER_val`.
18+
Also annotation lists `val.txt` and `train.txt` should be copied to `data/WIDERFace` from `WIDER_train_annotations` and `WIDER_val_annotations`.
19+
The directory should be like this:
20+
21+
```
22+
object_detection
23+
├── tools
24+
├── data
25+
│ ├── WIDERFace
26+
│ │ ├── WIDER_train
27+
│ | │ ├──0--Parade
28+
│ | │ ├── ...
29+
│ | │ ├── Annotations
30+
│ │ ├── WIDER_val
31+
│ | │ ├──0--Parade
32+
│ | │ ├── ...
33+
│ | │ ├── Annotations
34+
│ │ ├── val.txt
35+
│ │ ├── train.txt
36+
37+
```
38+
4. Create virtual environment `bash init_venv.sh`
39+
40+
## Training
41+
42+
1. Download pre-trained MobileNetV2 weights `mobilenet_v2.pth.tar` from: [https://github.com/tonylins/pytorch-mobilenet-v2](https://github.com/tonylins/pytorch-mobilenet-v2). Move the file with weights to the folder `snapshots`.
43+
2. Run in terminal:
44+
```bash
45+
python3 ../../external/mmdetection/tools/train.py \
46+
../../external/mmdetection/configs/wider_face/mobilenetv2_tiny_ssd300_wider_face.py
47+
```
48+
to train the detector on a single GPU.
49+
50+
## Validation
51+
52+
1. Run in terminal
53+
```bash
54+
python3 ../../external/mmdetection/tools/test.py \
55+
../../external/mmdetection/configs/wider_face/mobilenetv2_tiny_ssd300_wider_face.py \
56+
<CHECKPOINT> \
57+
--out result.pkl
58+
```
59+
to dump detections.
60+
2. Then run
61+
```bash
62+
python3 ../../external/mmdetection/tools/voc_eval.py \
63+
result.pkl \
64+
../../external/mmdetection/configs/wider_face/mobilenetv2_tiny_ssd300_wider_face.py
65+
```
66+
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`.
67+
68+
## Conversion to OpenVINO format
69+
70+
1. Convert PyTorch model to ONNX format: run script in terminal
71+
```bash
72+
python3 tools/onnx_export.py \
73+
../../external/mmdetection/configs/wider_face/mobilenetv2_tiny_ssd300_wider_face.py
74+
<CHECKPOINT> \
75+
face_detector.onnx
76+
```
77+
This command produces `face_detector.onnx`.
78+
2. Convert ONNX model to OpenVINO format with Model Optimizer: run in terminal
79+
```bash
80+
python3 <OpenVINO_INSTALL_DIR>/deployment_tools/model_optimizer/mo.py \
81+
--input_model face_detector.onnx \
82+
--scale 255 \
83+
--reverse_input_channels
84+
```
85+
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.
86+
87+
88+
## Python demo
89+
90+
To run the demo connect a webcam end execute command:
91+
```bash
92+
python3 tools/detection_live_demo.py \
93+
../../external/mmdetection/configs/wider_face/mobilenetv2_tiny_ssd300_wider_face.py \
94+
<CHECKPOINT> \
95+
--cam_id 0
96+
```
97+
98+
99+
## Estimate theoretical computational complexity
100+
101+
To get per layer computational complexity estimations run the following command:
102+
```bash
103+
python3 tools/count_flops.py \
104+
../../external/mmdetection/configs/wider_face/mobilenetv2_tiny_ssd300_wider_face.py
105+
```
106+
107+
108+
## Fine-tuning
109+
110+
* The annotations have to be in [PASCAL VOC](http://host.robots.ox.ac.uk/pascal/VOC/voc2012/index.html#data) format.
111+
* Fine-tuning steps are the same as step 2 for training, just edit the initial checkpoint and data paths in detector config.

‎pytorch_toolkit/object_detection/init_venv.sh

+5
Original file line numberDiff line numberDiff line change
@@ -22,6 +22,11 @@ else
2222
echo "Model optimizer requirements were not installed. Please install the OpenVino toolkit to use one."
2323
fi
2424

25+
. venv/bin/activate
26+
cd ../../external/mmdetection/
27+
bash compile.sh
28+
python setup.py develop
29+
deactivate
2530

2631
echo
2732
echo "===================================================="

‎pytorch_toolkit/object_detection/prepare_mmdet.sh

-3
This file was deleted.

0 commit comments

Comments
 (0)