## Use Case and High-Level Description¶

This is a YOLO V3 network fine-tuned for Person/Vehicle/Bike detection for security surveillance applications. It works in a variety of scenes and weather/lighting conditions.

Yolo V3 is a real-time object detection model implemented with Keras* from this repository and converted to TensorFlow* framework.

This model was pre-trained on Common Objects in Context (COCO) dataset with 80 classes and then fine-tuned for Person/Vehicle/Bike detection.

## Specification¶

Metric

Value

Mean Average Precision (mAP)

48.89%

AP people

58.94%

AP vehicles

62.05%

AP bikes/motorcycles

25.66%

GFlops

65.98

MParams

61.92

Source framework

Keras*

Average Precision (AP) is defined as an area under the precision/recall curve.

Validation dataset consists of 34757 images from various scenes and includes:

Type of object

Number of bounding boxes

Vehicle

229503

Pedestrian

240009

Bike/Motorcycle

62643

Similarly, training dataset has 17084 images with:

Type of object

Number of bounding boxes

Vehicle

121111

Pedestrian

119546

Bike/Motorcycle

30220

## Inputs¶

Image, name: image_input, shape: 1, 3, 416, 416 in the format B, C, H, W, where:

• B - batch size

• C - number of channels

• H - image height

• W - image width

Expected color order: BGR.

## Outputs¶

1. The array of detection summary info, name: conv2d_58/Conv2D/YoloRegion, shape: 1, 255, 13, 13. The anchor values are 116,90, 156,198, 373,326.

2. The array of detection summary info, name: conv2d_66/Conv2D/YoloRegion, shape: 1, 255, 26, 26. The anchor values are 30,61, 62,45, 59,119.

3. The array of detection summary info, name: conv2d_74/Conv2D/YoloRegion, shape: 1, 255, 52, 52. The anchor values are 10,13, 16,30, 33,23.

For each of the arrays the output format is B, N*85, Cx, Cy, where:

• B - batch size

• N - number of detection boxes for cell

• Cx, Cy - cell index

Detection box has format [x, y, h, w, box_score, class_no_1, …, class_no_80], where:

• (x, y) - coordinates of box center relative to the cell

• h, w - raw height and width of box, apply exponential function and multiply them by the corresponding anchors to get the absolute height and width values

• box_score - confidence of detection box in [0, 1] range

• class_no_1, …, class_no_80 - probability distribution over the classes in the [0, 1] range, multiply them by the confidence value box_score to get confidence of each class

Since the model is finetuned on person/vehicle/bike detection dataset, it returns non-zero scores for the following classes:

• person - the first class score

• non-vehicle (bike/motorcycle) - the second class score

• vehicle - the third class score Note that the indexes of these 3 classes are aligned with the indexes of the classes person, bike, and car in the original Common Objects in Context (COCO) dataset. Also note that the model returns class scores for all 80 COCO classes for backward compatibility with the original Yolo V3.