Use Case and High-Level Description

This is a reimplemented and retrained version of the tiny YOLO v2 object detection network trained with the VOC2012 training dataset. Network weight pruning is applied to sparsify convolution layers (60% of network parameters are set to zeros).

Example

Specification

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

Performance

Inputs

Name: input, shape: [1x3x416x416] - An input image in the format [BxCxHxW], where:

• B - batch size
• C - number of channels
• H - image height
• W - image width 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)
• y_loc and x_loc: spatial location of each grid

Legal Information

[*] Same as the original implementation.

[**] Other names and brands may be claimed as the property of others.