Skip to content

csvance/keras-mobile-detectnet

master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Keras MobileDetectNet

Example

MobileDetectNet is an object detector which uses MobileNet CNN to predict bounding boxes. It was designed to be computationally efficient for deployment on embedded systems and easy to train with limited data. It was inspired by the simple yet effective design of DetectNet and enhanced with the anchor system from Faster R-CNN.

Network Arcitecture

Example

Training

python train.py --help

Label Format

MobileDetectNet uses the KITTI label format and directory structure. See here for more details

Preprocessing

Images are scaled between -1 and 1 to take advantage of transfer learning from pretrained MobileNet.

Anchors

MobileNet outputs a 7x7x256 from its last layer with a 224x224x3 input. In each of the 7x7 squares we place 9 anchors with combinations of the following settings:

  • Scale 1, 2, and 3
  • Aspect Ratio 1, 4/3, and 3/4

We set the anchor to 1 if a rectangle has > 0.3 IoU with the anchor. The bounding box generated is given to the box with the highest IoU over 0.3.

Due to the smaller network receptive size and low spacial dimension output of MobileNet, anchors partially outside the image can be used.

Augmentation

Training is done with imgaug utilizing Keras Sequences for multicore preprocessing and online data augmentation:

return iaa.Sequential([
    iaa.Fliplr(0.5),
    iaa.CropAndPad(px=(0, 112), sample_independently=False),
    iaa.Affine(translate_percent={"x": (-0.4, 0.4), "y": (-0.4, 0.4)}),
    iaa.SomeOf((0, 3), [
        iaa.AddToHueAndSaturation((-10, 10)),
        iaa.Affine(scale={"x": (0.9, 1.1), "y": (0.9, 1.1)}),
        iaa.GaussianBlur(sigma=(0, 1.0)),
        iaa.AdditiveGaussianNoise(scale=0.05 * 255)
    ])
])

Data augmentation is also used for validation for the purpose of making sure smaller objects are detected.

return iaa.Sequential([
    iaa.CropAndPad(px=(0, 112), sample_independently=False),
    iaa.Affine(translate_percent={"x": (-0.4, 0.4), "y": (-0.4, 0.4)}),
])

If a dataset contains many smaller bounding boxes or detecting smaller objects is not a concern, this should be adjusted for both train and validation augmentation.

Loss

Standard loss functions are used for everything other than the bounding box regression, which uses 10*class_true_(ij)*|y_pred_(ij) - y_true_(ij)| in order to not penalize the network for bounding box predictions without an object present and to normalize the loss against class loss. Class loss is binary crossentropy and region loss is mean absolute error.

Optimization

Nadam is the recomended optimizer. A base lr of 0.001 is used, and ReduceLROnPlateau callback reduces it during training. Generally the model should converge to an optimal solution within 50 epochs, depending on the amount of training data used.

Inference

python inference.py --help

TensorRT

A TF-TRT helper function has been intergrated into the model which allows for easy inference acceleration on the nVidia Jetson platform. In model.py MobileDetectNet.tftrt_engine() will create a TensorRT accelerated Tensorflow graph. An example of how to use it is included in inference.py.

Performance

Using an FP16 TF-TRT graph the model runs at ~55 FPS on the Jetson Nano in mode 1 (5W). The performance doesn't seem to be effected running it in mode 0 (10W).

Releases

No releases published

Packages

No packages published