Single Shot MultiBox Detector implemented by Keras

Introduction

SSD(Single Shot MultiBox Detector) is a state-of-art object detection algorithm, brought by Wei Liu and other wonderful guys, see SSD: Single Shot MultiBox Detector @ arxiv, recommended to read for better understanding.

Also, SSD currently performs good at PASCAL VOC Challenge, see http://host.robots.ox.ac.uk:8080/leaderboard/displaylb.php?challengeid=11&compid=3

SSD Architecture

Below find the architectural defference between Yolo and SSD.

Guides

The code structures looks like below:

1. Aseets - Prior boxes  (prior_boxes_ssd300.pkl is the model pre-defined static prior boxes)
2. Data-set - VOC-2007
3. SSD Model - The training and the test scripts 

  - ssd_v2.py # main model architecture using Keras
	- ssd_layers.py # Normalize and PriorBox defenition
	- ssd_training.py # MultiboxLoss Definition
	- ssd_utils.py # Utilities including encode,decode,assign_boxes
  
4.  data-generator  # customrized generator, which return proper training data structure
				            # including image and assigned boxes(similar to input boxex)
  - get_data_from_XML.py # parse Annotations of PASCAL VOC, helper of generator
  

Walk-through

The multibox loss is consist of L1 smooth loss and softmax loss. Let's see how they llok like

Arguments y_true: Ground truth bounding boxes, tensor of shape (?, num_boxes, 4). y_pred: Predicted bounding boxes, tensor of shape (?, num_boxes, 4).

Returns l1_loss: L1-smooth loss, tensor of shape (?, num_boxes).

References - https://arxiv.org/abs/1504.08083

def _l1_smooth_loss(self, y_true, y_pred):
	abs_loss = tf.abs(y_true - y_pred)
	sq_loss = 0.5 * (y_true - y_pred)**2
	l1_loss = tf.where(tf.less(abs_loss, 1.0), sq_loss, abs_loss - 0.5)
	return tf.reduce_sum(l1_loss, -1)

Now let's walk through the softmax loss

Arguments y_true: Ground truth targets, tensor of shape (?, num_boxes, num_classes). y_pred: Predicted logits, tensor of shape (?, num_boxes, num_classes).

Returns softmax_loss: Softmax loss, tensor of shape (?, num_boxes).

def _softmax_loss(self, y_true, y_pred):
	y_pred = tf.maximum(tf.minimum(y_pred, 1 - 1e-15), 1e-15)
	softmax_loss = -tf.reduce_sum(y_true * tf.log(y_pred),
				      axis=-1)
	return softmax_loss

Resources

dataset can be downloaded from [http://host.robots.ox.ac.uk/pascal/VOC/, use The VOC2007 Challenge in this example

Weights can be downloaded at https://drive.google.com/file/d/0B5o_TPhUdyJWWEl5WG1lcUxCZzQ/view?usp=sharing

Hint

The folder called Data-set has just 5-10 samples of the VOC-2007. to train your own model please download the entire data set by clicking on the link above. Thanks

References

My work is just playing with this fantastic algorithm, and see the detection result of my own. Many many thanks goes to the author of the SSD paper