Display_D_HP/TeamAAS-VM/Core/DetectionResult.cs

using OpenCvSharp.Dnn;
using OpenCvSharp;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
using System.Threading.Tasks;
using Sdcb.OpenVINO;

namespace TeamAAS_VP.Core
{
    public class DetectionResult
    {
        public int ClassId { get; }
        public string Class { get; }
        public Rect Rect { get; }
        public float Confidence { get; }

        public DetectionResult(int classId, string @class, Rect rect, float confidence)
        {
            ClassId = classId;
            Class = @class;
            Rect = rect;
            Confidence = confidence;
        }

        public static DetectionResult[] FromYolov8DetectionResult(ReadOnlySpan<float> tensorData, Shape shape, Size2f sizeRatio, string[] dicts)
        {
            // tensorData: 1x84x8400=705600xF32
            // shape: 1x84x8400, 84=(x, y, width, height)+80 class confidences, 8400=possible object count(code should for loop 8400 first)
            float[] t = Transpose(tensorData, shape[1], shape[2]);
            List<DetectionResult> detResults = new List<DetectionResult>();

            int objectCount = shape[2];
            int clsRowCount = shape[1];
            if (dicts.Length != clsRowCount - 4) throw new ArgumentException($"dicts length {dicts.Length} does not match shape cls row count{clsRowCount}.");
            for (int i = 0; i < objectCount; i++)
            {
                int startIdx = i * clsRowCount;
                ReadOnlySpan<float> rectData = t.AsSpan().Slice(startIdx, 4);
                ReadOnlySpan<float> confidenceInfo = t.AsSpan().Slice(startIdx + 4, clsRowCount - 4);
                int maxConfidenceClsId = IndexOfMax(confidenceInfo);
                float confidence = confidenceInfo[maxConfidenceClsId];

                int centerX = (int)(rectData[0] * sizeRatio.Width);
                int centerY = (int)(rectData[1] * sizeRatio.Height);
                int width = (int)(rectData[2] * sizeRatio.Width);
                int height = (int)(rectData[3] * sizeRatio.Height);
                detResults.Add(new DetectionResult(
                    maxConfidenceClsId, dicts[maxConfidenceClsId],
                    new Rect(centerX - width / 2, centerY - height / 2, width, height),
                    confidence));
            }

            CvDnn.NMSBoxes(detResults.Select(x => x.Rect).ToList(), detResults.Select(x => x.Confidence).ToList(), scoreThreshold: 0.5f, nmsThreshold: 0.5f, out int[] indices);
            return detResults.Where((x, i) => indices.Contains(i)).ToArray();
        }

        private static int IndexOfMax(ReadOnlySpan<float> data)
        {
            if (data.Length == 0) throw new ArgumentException("The provided data span is null or empty.");

            // 初始化最大值及其索引
            int maxIndex = 0;
            float maxValue = data[0];

            // 遍历跨度查找最大值及其索引
            for (int i = 1; i < data.Length; i++)
            {
                if (data[i] > maxValue)
                {
                    maxValue = data[i];
                    maxIndex = i;
                }
            }

            // 返回最大值及其索引
            return maxIndex;
        }

        private static unsafe float[] Transpose(ReadOnlySpan<float> tensorData, int rows, int cols)
        {
            float[] transposedTensorData = new float[tensorData.Length];

            fixed (float* pTensorData = tensorData)
            {
                fixed (float* pTransposedData = transposedTensorData)
                {
                    for (int i = 0; i < rows; i++)
                    {
                        for (int j = 0; j < cols; j++)
                        {
                            // Index in the original tensor
                            int index = i * cols + j;

                            // Index in the transposed tensor
                            int transposedIndex = j * rows + i;

                            pTransposedData[transposedIndex] = pTensorData[index];
                        }
                    }
                }
            }

            return transposedTensorData;
        }
    }
}