NonMaximumSupression
After cloning the repo, use the following command to run:
g++ -std=c++17 `pkg-config --cflags opencv` detect.cpp `pkg-config --libs opencv` -o out -lstdc++fs -lpthread
Make sure your gcc is updated so that you are use C++17. I your compiler is older than gcc 7.1 use the flag -std=c++1z (experimental) instead.
The inputs are in the following directories :
img- input images showing vehiclesinput- a list of candidate detections per image resulting from our vehicle detection algorithm.
Each item in the list is a 5-tuple that describe a rectangle in the image coordinates, along with its confidence measure, a number between 0.0 and 1.0. For example, this candidate rectangle:
0.816 962 169 153 256
has
- confidence = 0.816
- top-left point: (x=962, y=169)
- dimensions: (width=153, height=256)
function 1
This implements an aggregate function that takes the candidate rectangles list as an argument and returns a list of detection rectangles, one per object.
Think of this as a the last stage in a vehicle detection algorithm. The detection algorithm generates a candidate rectangle everywhere in the image it thinks it's found a vehicle, and this can lead to multiple rectangles of slightly different sizes and shapes for the same vehicle. The aggregate function will get rid of all but the highest-confidence rectangle for each vehicle, such as seen in the figure above.
Pseudocode:
list<Rect> cars = aggregate(detections) {
// your code here
}
The code will read a file from the input/ folder and save the result to a Solution/ folder, in similar format to the files in the truth/ folder. Two files per image
- A text file with a space-separated 4-tuple (x y width height) per detection rectangle, with each detection in a new line
- An image file showing the rectangles from the text file drawn on top of the input image
function 2
In some of the images, the highest ranked candidate is not a good quality detection. For example, in img/2.png and img/3.png more than one vehicle gets grouped into one rectangle. Can you do better?
That brings up the question, how would you quantify the overall quality of your solution? The answer is to use the Intersection over Union (IoU) algorithm to compare your detection rectangles with the "truth" rectangles that we've drawn by hand (in the truth folder). The code should compare each detection rectangle with the best-fitting truth rectangle and generate a value between 0.0 (no overlap) and 1.0 (perfect overlap). It then saves these results to a scoring folder with one file per image, one detection score per line. Then also saves the overall average for all five images to a file named overall_score.