Single Shot MultiBox Detector

---

Keras/Tensorflow implementation of the SSD

The base source code is placed in the ssd dir. Currently SSD300 NN based on VGG16 model is implemented - see ssd300_vgg16.py.

Dataset

The SSD300-VGG16 is trained on Pascal VOC 2007+2012 dataset. Initially Ground-Truth Boxes are fetched from Pascal VOC dataset and stored as a hashtable. The keys of the hashtable are filenames, values are numpy arrays containing normalized bounding boxes, one-hot-encoded classes and difficulty property: [xmin, ymin, xmax, ymax, one-hot-encoded-class, is-difficult]. Ground-Truth Boxes are stored in the following pickle files:

• pascal_voc_2007_test.p (see http://host.robots.ox.ac.uk/pascal/VOC/voc2007/#testdata)
• pascal_voc_2007_trainval.p (see http://host.robots.ox.ac.uk/pascal/VOC/voc2007/#devkit)
• pascal_voc_2012_trainval.p (see http://host.robots.ox.ac.uk/pascal/VOC/voc2012/#devkit)

PriorBoxes (a.k.a. DefaultBoxes and Anchors)

PriorBoxes are generated as in the origin Caffe implementation.

PriorBoxes.ipynb contains samples of how PriorBoxes might look.

Data Augmentation

See DataAugmentation.ipynb for the whole process samples. See Imaging.ipynb for photo-metric distortion samples.

Hard negative mining

Mining hard examples is implemented in the SsdLoss class

Training

SSD300-VGG16.ipynb contains the training process.

Evaluating

This implementation has a lower performance comparing to the original Caffe implementation: overall mAP = 66%

Class	AP (%)
aeroplane	76
bicycle	76
bird	66
boat	63
bottle	39
bus	71
car	80
cat	80
chair	36
cow	59
diningtable	64
dog	70
horse	75
motorbike	72
person	70
pottedplant	44
sheep	56
sofa	71
train	76
tvmonitor	68

Dependencies

See Anaconda env file for dependencies.

To create env use:

conda env create -f anaconda-dl-env.yml

References

1. SSD: Single Shot MultiBox Detector

2. The origin SSD Caffe implementation

3. mAP calculating in py-faster-rcnn project