about yolov8

[aferust]

I dont have your email, so i write here. I have made some progress and done enough today. I continue another day. Maybe you make some progress. We have to figure out outSlice to extract box coordinates and class indices. The codebase is very rough and not performance-oriiented at the moment. You have to use dcv github repo. Maybe instead of video you can reorganize the codebase to process only an image.

some resouces:
https://github.com/ultralytics/ultralytics/blob/main/examples/YOLOv8-OpenCV-ONNX-Python/main.py
https://github.com/ultralytics/ultralytics/tree/main/examples/YOLOv8-ONNXRuntime-CPP
https://netron.app/ --> upload onnx file and observe inputs and outputs

dcv-yolov8.zip

[aferust]

Hey dont bother it. I was using wrong dims to extract boxes. Here is the correct structure. Now i have some issue with locating the exact positions. I will solve that soon.

// structure: 1 84 8400 is (batch,xywh+80class_score,boxes_num )

auto extractBoxCoordinates(S)(auto ref S outSlice, float confidenceThreshold) {
import mir.algorithm.iteration : minIndex, maxIndex;
//writeln("ws ", wscale, " hs ", hscale);
float[5][] boxes; // Assuming maximum possible number of bounding boxes

foreach (i; 0 .. outSlice.shape[2]) {
    auto classProbabilities = outSlice[0, 4 .. $, i];

    auto minClassLoc = classProbabilities.minIndex[0];
    auto maxClassLoc = classProbabilities.maxIndex[0];
    auto minScore = classProbabilities[minClassLoc];
    auto maxScore = classProbabilities[maxClassLoc];
    //classProbabilities.length.writeln;

    if (maxScore > confidenceThreshold) {
        // Object detected with confidence higher than the threshold
        // Extract bounding box coordinates
        auto width = outSlice[0, 2, i]  ;
        auto height = outSlice[0, 3, i] ;

        auto x = outSlice[0, 0, i];
        auto y = outSlice[0, 1, i];
        
        // Store the bounding box coordinates

        
        boxes ~= [x, y, width, height, maxClassLoc];
    }
}
return boxes;

}

[aferust]

app.d.zip

I corrected the box coords with this version