OBJECT DETECTION MADE EASY

These is a series on object detection which include various algorithms and codes.

Important points to consider:

1. Incase of faster rcnn every anchor box generated by the regional proposal network will be compare against every ground truth box and loss is calculated. But the loss is calculated and weights are updated through back propogation over only positive boxes (which has the good amount of overlap with atleast one of the gorund truth box).
2. Incase of faster rcnn total four losses will be present

• two classification losses one for region proposal (object or background) and other for object (based on the objects present in the image)
• two regression losses , one for the region proposals that are getting generated from the anchors , another for the bounding boxes for the objects in the image.

3. Incase of region proposals only if the object is present then we calculate the bounding box loss and adjust the coordinates.

4. For the classification object is presnt or not the ground truth taken will be based on the IOU score. If the IOU is greater than the 0.7 for the proposed anchor box then the object is present other wise its considerd as backgorund .

5.The stride values that are considerd for generation of the anchor boxes can be configured in the config file of the object detection api.