Object Detection CNN

Object detection using convolutional neural networks with already trained YOLO (You Only Look Once) model.

Notes

Trained model

• Download trained model (195 MB) from here.
• Save it in /model_data folder

Input image

For every new image:

• Update size:

# car_detection_yolo.py
96  image_shape = (1080., 1440.) # image_shape = (Height, Width)

• Update name:

# car_detection_yolo.py
126  out_scores, out_boxes, out_classes = predict(sess, "sh_taxi.jpg") # name = "sh_taxi.jpg"

How does it work?

1. CNN output

The input image goes through a CNN, resulting in a (19,19,5,85) dimensional output. Run summary() to see whole framework architecture:

# car_detection_yolo.py
99  yolo_model.summary()

2. Output processing

After flattening the last two dimensions, the output is a volume of shape (19, 19, 425):

• Each cell in a 19x19 grid over the input image gives 425 numbers.
• 425 = 5 x 85 because each cell contains predictions for 5 boxes, corresponding to 5 anchor boxes, as seen in lecture.
• 85 = 5 + 80 where 5 is because (pc,bx,by,bh,bw)(pc,bx,by,bh,bw) has 5 numbers, and and 80 is the number of classes we'd like to detect

3. Box selection

You then select only few boxes based on:

• Score-thresholding: throw away boxes that have detected a class with a score less than the threshold

# car_detection_yolo.py
def yolo_filter_boxes(box_confidence, boxes, box_class_probs, threshold = .6)

• Non-max suppression: Compute the Intersection over Union and avoid selecting overlapping boxes

# car_detection_yolo.py
def iou(box1, box2)
def yolo_non_max_suppression(scores, boxes, classes, max_boxes = 10, iou_threshold = 0.5)

4. Final output

This gives you YOLO's final output: