yolo-v2-ava-sparse-70-0001

Use Case and High-Level Description

This is a reimplemented and retrained version of the YOLO v2 object detection network trained with the VOC2012 training dataset. Network weight pruning is applied to sparsify convolution layers (70% of network parameters are set to zeros).

Specification

Metric

Value

Mean Average Precision (mAP)

62.9%

GFlops

29.4205

MParams

50.6451

Source framework

TensorFlow*

For Average Precision metric description, see The PASCAL Visual Object Classes (VOC) Challenge. Tested on the VOC 2012 validation dataset.

Inputs

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

  • B - batch size

  • H - image height

  • W - image width

  • C - number of channels

Expected color order is BGR.

Outputs

The net outputs a blob with the shape 1, 21125 which can be reshaped to 5, 25, 13, 13, where each number corresponds to [num_anchors, cls_reg_obj_params, y_loc, x_loc] respectively:

  • num_anchors : number of anchor boxes, each spatial location specified by y_loc and x_loc has five anchors

  • cls_reg_obj_params : parameters for classification and regression. The values are made up of the following:

    • Regression parameters (4)

    • Objectness score (1)

    • Class score (20), mapping to class names provided by <omz_dir>/data/dataset_classes/voc_20cl.txt file.

  • y_loc and x_loc : spatial location of each grid

Demo usage

The model can be used in the following demos provided by the Open Model Zoo to show its capabilities: