Skip to content

Latest commit

 

History

History
159 lines (113 loc) · 7.67 KB

face_detection.md

File metadata and controls

159 lines (113 loc) · 7.67 KB
Raw

Face Detection

Models that are able to detect faces on given images.

Model Name Complexity (GFLOPs) Size (Mp) AP for faces > 64x64 WiderFace(Easy/Medium/Hard) Links GPU_NUM
face-detection-retail-0005 0.982 1.021 84.52% 78.97% / 70.80% / 37.32% snapshot, configuration file 2
face-detection-0100 0.785 1.828 86.16% 82.50% / 75.77% / 40.83% snapshot, configuration file 2
face-detection-0102 1.767 1.842 91.47% 88.90% / 83.51% / 49.58% snapshot, configuration file 2
face-detection-0104 2.405 1.851 92.69% 90.15% / 84.80% / 51.61% snapshot, configuration file 4
face-detection-0105 2.853 2.392 93.34% 91.61% / 85.94% / 52.93% snapshot, configuration file 4
face-detection-0106 339.597 69.920 94.49% 94.51% / 93.20% / 83.09% snapshot, configuration file 8

Training pipeline

1. Select a training configuration file and get pre-trained snapshot if available. Please see the table above.

export MODEL_NAME=face-detection-0100
export CONFIGURATION_FILE=./configs/$MODEL_NAME.py

2. Collect dataset

Download the WIDER Face and unpack it to the data folder.

3. Prepare annotation

Convert downloaded and extracted annotation to MSCOCO format with face as the only one class.

  • Training annotation

    python tools/task_specific/face_detection/wider_to_coco.py \
         data/wider_face_split/wider_face_train_bbx_gt.txt \
         data/WIDER_train/images/ \
         data/train.json

  • Validation annotation

    python tools/task_specific/face_detection/wider_to_coco.py \
         data/wider_face_split/wider_face_val_bbx_gt.txt \
         data/WIDER_val/images/ \
         data/val.json

4. Training and Fine-tuning

Try both following variants and select the best one:

  • Training from scratch or pre-trained weights. Only if you have a lot of data, let's say tens of thousands or even more images. This variant assumes long training process starting from big values of learning rate and eventually decreasing it according to a training schedule.
  • Fine-tuning from pre-trained weights. If the dataset is not big enough, then the model tends to overfit quickly, forgetting about the data that was used for pre-training and reducing the generalization ability of the final model. Hence, small starting learning rate and short training schedule are recommended.

If you would like to start training from pre-trained weights do not forget to modify load_from path inside configuration file.

If you would like to start fine-tuning from pre-trained weights do not forget to modify resume_from path inside configuration file as well as increase total_epochs. Otherwise training will be ended immideately.

  • To train the detector on a single GPU, run in your terminal:

    python ../../external/mmdetection/tools/train.py \
         $CONFIGURATION_FILE

  • To train the detector on multiple GPUs, run in your terminal:

    ../../external/mmdetection/tools/dist_train.sh \
         $CONFIGURATION_FILE \
         <GPU_NUM>

5. Validation

  • To dump detection of your model as well as compute MS-COCO metrics run:

    python ../../external/mmdetection/tools/test.py \
         $CONFIGURATION_FILE \
         <CHECKPOINT> \
         --out result.pkl \
         --eval bbox

  • You can also measure WiderFace quality metrics:

    1. Convert result.pkl obtained from previous step to WiderFace-friendly output:

      python tools/test_out_to_wider_predictions.py \
         $CONFIGURATION_FILE \
         result.pkl \
         <OUTPUT_FOLDER>

    2. Run WiderFace validation either with

6. Export PyTorch* model to the OpenVINO™ format

To convert PyTorch* model to the OpenVINO™ IR format run the export.py script:

python ../../external/mmdetection/tools/export.py \
      $CONFIGURATION_FILE \
      <CHECKPOINT> \
      <EXPORT_FOLDER> \
      openvino

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

For SSD networks an alternative OpenVINO™ representation is possible. To opt for it use extra --alt_ssd_export key to the export.py script. SSD model exported in such way will produce a bit different results (non-significant in most cases), but it also might be faster than the default one.

7. Validation of IR

Instead of running test.py you need to run test_exported.py and then repeat steps listed in Validation paragraph.

python ../../external/mmdetection/tools/test_exported.py  \
      $CONFIGURATION_FILE \
      <EXPORT_FOLDER>/$MODEL_NAME.xml \
      --out results.pkl \
      --eval bbox

8. Demo

To see how the converted model works using OpenVINO you need to run test_exported.py with --show option.

python ../../external/mmdetection/tools/test_exported.py  \
      $CONFIGURATION_FILE \
      <EXPORT_FOLDER>/$MODEL_NAME.xml \
      --show

Other

Theoretical computational complexity estimation

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

python ../../external/mmdetection/tools/get_flops.py \
       $CONFIGURATION_FILE