How to calculate IOU for multiple bounding boxes in one img?

user_7967059 0
Viewed 2230

Suppose there is a object detection task. There are many cars in the img. What I need to do, is to draw boxes to detect these cars. Someone else is doing the same task. My result is [car1,car2,car5,car6], someone's result is [car1,car3,car5,car5,car7].（carx is the coordinate of #x cat in this img)
Both of us only detect some of the car in the img.
Now I want to find a way evaluate the similarity between our result.
If both of us only find one car, it is easy,

def cal_iou(coordinate_x, coordinate_y):
    polygon_x = geometry.Polygon(coordinate_x)
    polygon_y = geometry.Polygon(coordinate_y)
    intersection = polygon_x.intersection(polygon_y).area
    total_area = polygon_x.area + polygon_y.area - intersection
    if total_area > 0.0:
        return float(intersection / total_area)
    return 0.0

I can compute the IoU between two boxex. What if there are more than one box in our result? I don't know which box in other's result should be compared with car1? Another way I think is to get all area in my result, and compare with other's result. But I have not find a good way to do this. plz give me some suggestion

object-detection computer-vision opencv python
2 Answers

The IoU between box1 to box2 is the same as box2 to box1. What I most of the time do is calculate all IoUs and pick the highest. The 2 boxes how belong to this IoU can be removed and the process can be repeated. Till all boxes are linked or the highest IoU is below the threshold

user_43339820

calc all box with one box and then, pick best box that's the closest box

user_163973430
Related