Increase Bounding Box Accuracy in Object Detection Problem

0
votes

Though this is not a machine learning platform But I feel more comfortable in StackOverflow as a software engineer. My question is how to increase bounding box accuracy in the object detection problem after applying ML, for example, YOLO, Faster RCNN etc. I got one level of accuracy(mAP) after applying YOLOV3. But I need more accuracy( 95%+). Most of the object shapes in my case are rectangle/square. can anyone give me any clue about image processing or anything else?

1
deep-learningobject-detectionbounding-box
without explaining your own dataset and problem, its hard to give accurate advice. If they are mostly rectangle/square, you can use approxPolyDP in opencv. example. If you manage to find 95%+ of your objects that way, you can then train a nn to validate if the shape you think is a rectangle/square is indeed your object of interest. - juvian

1 Answers

1
votes

Try increasing the width and hight of the input image in te config file.Try Different Batch sizes.Use anchor boxes suitable for your dataset .These will help you to increae the map.