yolo5 Unity3D BarraCUDA

Question

Is there documentation for what yolo-v5's output is?

We (or I) are trying to use YOLOv5 (for now) to detect faces in real time from an HDMI input. Cartoon faces, video game faces, natural faces - all are welcome for our application. I've followed the instructions to retrain it, but, the results seem nonsensical.

I've red the embedding/export instructions - but - don't understand the details of how to interpret the output tensor.

• https://github.com/Hexmagic/ONNX-yolov5/blob/master/src/test.cpp

TLDR; I'm trying to interpret the yolo-v5 output but the examples don't seem to illustrate what's going on to me.

Answer 1

So ... the Unity3D BarraCUDA thing seems to emit YOLOv5 results in column major order. The C++ and Python examples show the results in row-major order.

My transpose function looks something like this ... and it works now

        /// <summary>
        /// 
        /// </summary>
        /// <param name="width">5 + number of classes</param>
        /// <param name="floats">outputTensor.ToReadOnlyArray()</param>
        /// <returns></returns>
        private static IEnumerable<(Rect, float, float)> Transpose(int width, float[] floats)
        {
            Debug.Assert(6 == width); // i only have one class (you'll need to change the return value if you use more)
            int l = floats.Length;
            int count = l / width;


            // pre-compute these
            var heads = Enumerable.Range(0, width).ToArray();


            // process each "row" - you (i) can run this in parallel just fine
            return Enumerable.Range(0, count).Select(i =>
            {
                // compute the entry
                var entry = heads.Select(h => floats[i + (h * count)]).ToArray();


                var rect = new Rect();

                rect.x = entry[0];
                rect.y = entry[1];
                rect.width = entry[2];
                rect.height = entry[3];

                return (rect, entry[4], entry[5]);
            });
        }