admin
a0b115d954
- 标注：AI 预标注直写 labels_json（去候选确认两阶段）；重叠去重（minIoU）；全量标注按钮 - 训练：脚本迁移入 server/training/（Go 化 prepare_yolo/analyze_rfdetr，保留 train_server.py）；tflite 产物自检并入训练流程（check_tflite） - 数据目录/权重不进 git；.gitignore 迁移至仓库根
520 lines
19 KiB
Dart
520 lines
19 KiB
Dart
import 'dart:typed_data';
|
||
|
||
import 'package:flutter/foundation.dart' show debugPrint;
|
||
import 'package:tflite_flutter/tflite_flutter.dart';
|
||
|
||
import 'detection_result.dart';
|
||
import 'nms.dart';
|
||
|
||
/// YOLOv8n 端侧推理实现（对应 Kotlin TFLiteDetector）。
|
||
/// 模型输出布局（ultralytics litert 导出）：[1, 4 + nc, anchors]，
|
||
/// cx/cy/w/h 已归一化，类别得分已过 sigmoid；按 out[dim][anchor] 索引。
|
||
/// 输入为 NCHW [1, 3, 704, 704]（litert 导出保留 torch 布局）。
|
||
class TfliteDetector {
|
||
// 输入尺寸取自模型本身（ultralytics litert 导出 NCHW [1,3,H,W]，各数据集
|
||
// 训练 imgsz 可不同），默认 704 兜底
|
||
static const int defaultInputSize = 704;
|
||
// 野鸡数据置信度普遍偏低（0.1~0.2 量级），保留低分池供运动检测提升；
|
||
// 可运行时调整（设置页滑块），默认 0.10
|
||
double minScore = 0.10;
|
||
static const double iouThreshold = 0.45;
|
||
static const int maxDetections = 20;
|
||
static const String modelAsset = 'assets/model.tflite';
|
||
static const String labelsAsset = 'assets/labels.txt';
|
||
|
||
final Interpreter _interpreter;
|
||
final List<String> _labels;
|
||
final int _numClasses;
|
||
final int _numAnchors;
|
||
final int inputSize;
|
||
|
||
/// 模型身份（多模型并行推理区分来源；内置资产模型为 -1/空）
|
||
final int modelId;
|
||
final String modelName;
|
||
|
||
late final Float32List _input;
|
||
|
||
/// 输出按模型形状 [1, 4+nc, anchors] 的嵌套 List 组织，
|
||
/// run() 要求输出对象形状与模型完全一致（扁平 List 会被拒）。
|
||
final List<List<List<double>>> _output;
|
||
|
||
TfliteDetector._(this._interpreter, this._labels, this._numClasses,
|
||
this._numAnchors, this._output, this.inputSize, this.modelId,
|
||
this.modelName) {
|
||
_input = Float32List(1 * inputSize * inputSize * 3);
|
||
}
|
||
|
||
/// 模型缺失或加载失败返回 null（App 降级为仅预览）。
|
||
/// 在后台 isolate 内调用（模型字节由主 isolate 读取后传入）。
|
||
static Future<TfliteDetector?> fromBuffer(
|
||
Uint8List bytes,
|
||
List<String> labels, {
|
||
int modelId = -1,
|
||
String modelName = '',
|
||
}) async {
|
||
try {
|
||
final interpreter = Interpreter.fromBuffer(
|
||
bytes,
|
||
options: InterpreterOptions()..threads = 4,
|
||
);
|
||
return TfliteDetector._fromModel(
|
||
interpreter, labels, modelId, modelName);
|
||
} catch (_) {
|
||
return null;
|
||
}
|
||
}
|
||
|
||
/// 输出布局 [1, 4+nc, anchors] 取自模型本身，类别数不与 labels 文件长度耦合。
|
||
factory TfliteDetector._fromModel(Interpreter interpreter,
|
||
List<String> labels, int modelId, String modelName) {
|
||
final shape = interpreter.getOutputTensor(0).shape;
|
||
final numClasses =
|
||
shape.length >= 3 && shape[1] > 4 ? shape[1] - 4 : labels.length;
|
||
final numAnchors = shape.length >= 3 && shape[2] > 0 ? shape[2] : 2100;
|
||
final inputShape = interpreter.getInputTensor(0).shape;
|
||
final inputSize = inputShape.length >= 4
|
||
? inputShape[3]
|
||
: defaultInputSize;
|
||
final output = List.generate(
|
||
1,
|
||
(_) => List.generate(
|
||
numClasses + 4,
|
||
(_) => List<double>.filled(numAnchors, 0),
|
||
),
|
||
);
|
||
return TfliteDetector._(interpreter, labels, numClasses, numAnchors,
|
||
output, inputSize, modelId, modelName);
|
||
}
|
||
|
||
/// 原始数据接口（后台 isolate 用，不依赖 CameraImage）。
|
||
/// 输出坐标统一反算为原图归一化空间（与 MotionDetector 一致），
|
||
/// 否则 CENTER_CROP 裁剪偏移会让检测框系统性偏移。
|
||
List<DetectionResult> detectRaw({
|
||
required List<Uint8List> planes,
|
||
required List<int> strides,
|
||
required int width,
|
||
required int height,
|
||
required bool isBgra,
|
||
bool rgbaOrder = false,
|
||
}) {
|
||
preprocess(
|
||
planes: planes,
|
||
strides: strides,
|
||
width: width,
|
||
height: height,
|
||
isBgra: isBgra,
|
||
rgbaOrder: rgbaOrder);
|
||
// 传原始字节视图而非 Float32List：tflite_flutter 会对非 ByteBuffer/Uint8List
|
||
// 输入调用 resizeInputTensor（1 维 [1486848]），使 node 0 TRANSPOSE prepare 失败
|
||
_interpreter.run(_input.buffer.asUint8List(), _output);
|
||
final dets = postprocess();
|
||
// 反算与 preprocess 的 scale/dx/dy 公式一致（704 输入空间 → 原图归一化）
|
||
final scale = inputSize / width < inputSize / height
|
||
? inputSize / width
|
||
: inputSize / height;
|
||
final dx = (inputSize - width * scale) / 2;
|
||
final dy = (inputSize - height * scale) / 2;
|
||
if (dx == 0 && dy == 0) return dets;
|
||
return dets
|
||
.map((r) => r.copyWith(
|
||
left: (r.left * inputSize - dx) / (width * scale),
|
||
right: (r.right * inputSize - dx) / (width * scale),
|
||
top: (r.top * inputSize - dy) / (height * scale),
|
||
bottom: (r.bottom * inputSize - dy) / (height * scale),
|
||
))
|
||
.toList();
|
||
}
|
||
|
||
/// 按像素格式分派：单平面 RGBA/BGRA（iOS bgra8888 / Android 实验） / yuv420 多平面。
|
||
void preprocess({
|
||
required List<Uint8List> planes,
|
||
required List<int> strides,
|
||
required int width,
|
||
required int height,
|
||
required bool isBgra,
|
||
bool rgbaOrder = false,
|
||
}) {
|
||
if (isBgra) {
|
||
_preprocessBgra(planes[0], strides[0], width, height, rgbaOrder);
|
||
} else {
|
||
_preprocessYuv(planes, strides, width, height);
|
||
}
|
||
}
|
||
|
||
/// 单平面 8888（iOS bgra8888 = [b,g,r,a]；Android 实验 RGBA_8888 = [r,g,b,a]）：
|
||
/// 每像素 4 字节，双线性采样，letterbox（等比缩到长边 704，短边黑边补 0）。
|
||
void _preprocessBgra(
|
||
Uint8List src, int stride, int srcW, int srcH, bool rgbaOrder) {
|
||
final plane = inputSize * inputSize;
|
||
final scale = inputSize / srcW < inputSize / srcH
|
||
? inputSize / srcW
|
||
: inputSize / srcH;
|
||
final dx = (inputSize - srcW * scale) / 2;
|
||
final dy = (inputSize - srcH * scale) / 2;
|
||
// rgbaOrder=false（iOS BGRA）: +0 B、+1 G、+2 R、+3 A；
|
||
// rgbaOrder=true（Android RGBA）: +0 R、+1 G、+2 B、+3 A
|
||
final rOff = rgbaOrder ? 0 : 2;
|
||
final bOff = rgbaOrder ? 2 : 0;
|
||
|
||
for (var oy = 0; oy < inputSize; oy++) {
|
||
final syf = (oy - dy) / scale;
|
||
if (syf < 0 || syf >= srcH) {
|
||
for (var ox = 0; ox < inputSize; ox++) {
|
||
final p = oy * inputSize + ox;
|
||
_input[p] = 0;
|
||
_input[p + plane] = 0;
|
||
_input[p + 2 * plane] = 0;
|
||
}
|
||
continue;
|
||
}
|
||
for (var ox = 0; ox < inputSize; ox++) {
|
||
final p = oy * inputSize + ox;
|
||
final sxf = (ox - dx) / scale;
|
||
if (sxf < 0 || sxf >= srcW) {
|
||
_input[p] = 0;
|
||
_input[p + plane] = 0;
|
||
_input[p + 2 * plane] = 0;
|
||
continue;
|
||
}
|
||
final x0 = sxf.floor(), y0 = syf.floor();
|
||
final x1 = x0 < srcW - 1 ? x0 + 1 : x0;
|
||
final y1 = y0 < srcH - 1 ? y0 + 1 : y0;
|
||
final fx = sxf - x0, fy = syf - y0;
|
||
|
||
final i00 = y0 * stride + x0 * 4;
|
||
final i10 = y0 * stride + x1 * 4;
|
||
final i01 = y1 * stride + x0 * 4;
|
||
final i11 = y1 * stride + x1 * 4;
|
||
final r00 = src[i00 + rOff].toDouble();
|
||
final g00 = src[i00 + 1].toDouble();
|
||
final b00 = src[i00 + bOff].toDouble();
|
||
final r10 = src[i10 + rOff].toDouble();
|
||
final g10 = src[i10 + 1].toDouble();
|
||
final b10 = src[i10 + bOff].toDouble();
|
||
final r01 = src[i01 + rOff].toDouble();
|
||
final g01 = src[i01 + 1].toDouble();
|
||
final b01 = src[i01 + bOff].toDouble();
|
||
final r11 = src[i11 + rOff].toDouble();
|
||
final g11 = src[i11 + 1].toDouble();
|
||
final b11 = src[i11 + bOff].toDouble();
|
||
|
||
_input[p] = _bl(r00, r10, r01, r11, fx, fy) / 255.0;
|
||
_input[p + plane] = _bl(g00, g10, g01, g11, fx, fy) / 255.0;
|
||
_input[p + 2 * plane] = _bl(b00, b10, b01, b11, fx, fy) / 255.0;
|
||
}
|
||
}
|
||
}
|
||
|
||
/// letterbox 缩放 + YUV → RGB 归一化 0~1（NCHW），双线性采样。
|
||
/// 兼容 NV12（双平面，UV 交错）与 I420（三平面）。
|
||
/// 首帧自适应：Y 值域（full/limited）与色序（U 先/V 先）因设备而异，
|
||
/// 静态假设会在部分机型上产生偏色 → 检测退化。
|
||
void _preprocessYuv(
|
||
List<Uint8List> planes, List<int> strides, int srcW, int srcH) {
|
||
final plane = inputSize * inputSize;
|
||
final y = planes[0];
|
||
final nv12 = planes.length == 2;
|
||
final uv = nv12 ? planes[1] : null;
|
||
final yStride = strides[0];
|
||
final uvStride = strides[1];
|
||
final vStride =
|
||
nv12 ? uvStride : (strides.length > 2 ? strides[2] : strides[1]);
|
||
|
||
// 色序修正后的 U/V 采样（nv12：偶位 U 奇位 V，NV21 相反；i420：平面 1/2 对调）
|
||
double uAt(int x, int y) => nv12
|
||
? uv![y * uvStride + (_yuvSwapChroma ? x * 2 + 1 : x * 2)] - 128.0
|
||
: planes[_yuvSwapChroma ? 2 : 1][y * uvStride + x] - 128.0;
|
||
double vAt(int x, int y) => nv12
|
||
? uv![y * uvStride + (_yuvSwapChroma ? x * 2 : x * 2 + 1)] - 128.0
|
||
: planes[_yuvSwapChroma ? 1 : 2][y * vStride + x] - 128.0;
|
||
|
||
final scale = inputSize / srcW < inputSize / srcH
|
||
? inputSize / srcW
|
||
: inputSize / srcH;
|
||
final dx = (inputSize - srcW * scale) / 2;
|
||
final dy = (inputSize - srcH * scale) / 2;
|
||
|
||
for (var oy = 0; oy < inputSize; oy++) {
|
||
final syf = (oy - dy) / scale;
|
||
if (syf < 0 || syf >= srcH) {
|
||
for (var ox = 0; ox < inputSize; ox++) {
|
||
final p = oy * inputSize + ox;
|
||
_input[p] = 0;
|
||
_input[p + plane] = 0;
|
||
_input[p + 2 * plane] = 0;
|
||
}
|
||
continue;
|
||
}
|
||
for (var ox = 0; ox < inputSize; ox++) {
|
||
final p = oy * inputSize + ox;
|
||
final sxf = (ox - dx) / scale;
|
||
if (sxf < 0 || sxf >= srcW) {
|
||
_input[p] = 0;
|
||
_input[p + plane] = 0;
|
||
_input[p + 2 * plane] = 0;
|
||
continue;
|
||
}
|
||
final x0 = sxf.floor(), y0 = syf.floor();
|
||
final x1 = x0 < srcW - 1 ? x0 + 1 : x0;
|
||
final y1 = y0 < srcH - 1 ? y0 + 1 : y0;
|
||
final fx = sxf - x0, fy = syf - y0;
|
||
|
||
// Y 双线性
|
||
final y00 = y[y0 * yStride + x0].toDouble();
|
||
final y10 = y[y0 * yStride + x1].toDouble();
|
||
final y01 = y[y1 * yStride + x0].toDouble();
|
||
final y11 = y[y1 * yStride + x1].toDouble();
|
||
final yy = _bl(y00, y10, y01, y11, fx, fy);
|
||
|
||
// U/V 双线性（4:2:0 半分辨率，按像素坐标定位后除 2）
|
||
final maxUx = srcW ~/ 2 - 1;
|
||
final maxUy = srcH ~/ 2 - 1;
|
||
final ux0 = (x0 ~/ 2).clamp(0, maxUx).toInt();
|
||
final uy0 = (y0 ~/ 2).clamp(0, maxUy).toInt();
|
||
final ux1 = (x1 ~/ 2).clamp(0, maxUx).toInt();
|
||
final uy1 = (y1 ~/ 2).clamp(0, maxUy).toInt();
|
||
final u00 = uAt(ux0, uy0);
|
||
final u10 = uAt(ux1, uy0);
|
||
final u01 = uAt(ux0, uy1);
|
||
final u11 = uAt(ux1, uy1);
|
||
final uu = _bl(u00, u10, u01, u11, fx, fy);
|
||
|
||
final v00 = vAt(ux0, uy0);
|
||
final v10 = vAt(ux1, uy0);
|
||
final v01 = vAt(ux0, uy1);
|
||
final v11 = vAt(ux1, uy1);
|
||
final vv = _bl(v00, v10, v01, v11, fx, fy);
|
||
|
||
// 值域展开：有限范围 VideoRange（Y 16~235，Cb/Cr 16~240）需线性拉伸；
|
||
// 全值域相机直接使用原始值（与 iOS bgra 一致）
|
||
final yr = _yuvFullRange ? yy : (yy - 16.0) * (255.0 / 219.0);
|
||
final un = _yuvFullRange ? uu : uu * (255.0 / 224.0);
|
||
final vn = _yuvFullRange ? vv : vv * (255.0 / 224.0);
|
||
|
||
// NCHW：r/g/b 分平面存储
|
||
_input[p] = (yr + 1.402 * vn) / 255.0;
|
||
_input[p + plane] = (yr - 0.344136 * un - 0.714136 * vn) / 255.0;
|
||
_input[p + 2 * plane] = (yr + 1.772 * un) / 255.0;
|
||
}
|
||
}
|
||
}
|
||
|
||
/// 首帧自适应判定 YUV 模式，后续帧复用（相机重启后由 worker 复位重判）。
|
||
/// - 值域：有限范围黑电平恒为 16，低于 12 只可能是全值域。
|
||
/// - 色序：以模型本身为 oracle——同一帧按两种色序各推理一次，
|
||
/// 检测数/最高分/总分更高者为真；两序均无检测时退回亮区色相计数启发
|
||
/// （户外最亮区域为天空应偏蓝，若按默认 U 先序解出偏红则为 V 先序）。
|
||
/// - 首帧可能曝光未收敛（过暗/全黑），此时 oracle 与启发式都不可信，
|
||
/// 保持未判定状态等下一帧，避免在垃圾帧上锁死错误色序（真机零检测根因）。
|
||
bool _yuvFullRange = false;
|
||
bool _yuvSwapChroma = false;
|
||
bool _yuvModeKnown = false;
|
||
bool _yuvRetried = false;
|
||
int _yuvDecisionMs = 0;
|
||
String _yuvDiag = '';
|
||
|
||
bool get yuvModeKnown => _yuvModeKnown;
|
||
bool get yuvRetried => _yuvRetried;
|
||
int get yuvDecisionMs => _yuvDecisionMs;
|
||
String get yuvDiag => _yuvDiag;
|
||
|
||
void decideYuvChroma({
|
||
required List<Uint8List> planes,
|
||
required List<int> strides,
|
||
required int width,
|
||
required int height,
|
||
bool force = false,
|
||
}) {
|
||
if (_yuvModeKnown && !force) return;
|
||
final (yMin, yMax, yMean) = _yStats(planes[0], strides[0], width, height);
|
||
_yuvFullRange = yMin < 12;
|
||
if (yMean < 30 || yMax < 170) {
|
||
// 曝光未稳定：保持未判定，下一帧重试；始终昏暗则维持默认（同旧版）
|
||
if (!_yuvModeKnown) {
|
||
_yuvDiag = '等稳定帧 mean=${yMean.toStringAsFixed(0)} max=$yMax';
|
||
}
|
||
return;
|
||
}
|
||
_yuvModeKnown = true;
|
||
_yuvDecisionMs = DateTime.now().millisecondsSinceEpoch;
|
||
final a = _runWithSwap(planes, strides, width, height, false);
|
||
final b = _runWithSwap(planes, strides, width, height, true);
|
||
var swap = false;
|
||
if (a.$1 != b.$1) {
|
||
swap = b.$1 > a.$1;
|
||
} else if (a.$2 != b.$2) {
|
||
swap = b.$2 > a.$2;
|
||
} else if (a.$3 != b.$3) {
|
||
swap = b.$3 > a.$3;
|
||
} else {
|
||
swap = _brightRegionLeansRed(planes, strides, width, height, yMax);
|
||
}
|
||
_yuvSwapChroma = swap;
|
||
_yuvDiag = 'full=$_yuvFullRange swap=$_yuvSwapChroma'
|
||
' cA=${a.$1} sA=${a.$2.toStringAsFixed(3)}'
|
||
' cB=${b.$1} sB=${b.$2.toStringAsFixed(3)}';
|
||
|
||
debugPrint('[yuv] $_yuvDiag mean=${yMean.toStringAsFixed(0)} max=$yMax');
|
||
}
|
||
|
||
/// 判定后持续无检测的自愈：用实时帧重跑完整判定（仅一次）。
|
||
/// 首帧模糊/暗帧导致启发式猜错时，等画面稳定后 oracle 即可分胜负。
|
||
/// 返回是否执行了重判（随后应重跑 detectRaw 取新结果）。
|
||
bool retryDecision({
|
||
required List<Uint8List> planes,
|
||
required List<int> strides,
|
||
required int width,
|
||
required int height,
|
||
}) {
|
||
if (_yuvRetried || !_yuvModeKnown) return false;
|
||
final (_, yMax, yMean) = _yStats(planes[0], strides[0], width, height);
|
||
if (yMean < 30 || yMax < 170) return false; // 帧仍不可用
|
||
_yuvRetried = true;
|
||
decideYuvChroma(
|
||
planes: planes, strides: strides, width: width, height: height,
|
||
force: true);
|
||
return true;
|
||
}
|
||
|
||
/// 相机（重新）启动后复位，首帧重新判定
|
||
void resetYuvMode() {
|
||
_yuvModeKnown = false;
|
||
_yuvRetried = false;
|
||
_yuvDiag = '';
|
||
}
|
||
|
||
/// 采样统计 Y 值域：(min, max, mean)，步长 16px 约 3600 样本
|
||
(int, int, double) _yStats(Uint8List y, int yStride, int w, int h) {
|
||
var yMin = 255, yMax = 0;
|
||
var sum = 0, n = 0;
|
||
for (var j = 0; j < h; j += 16) {
|
||
final row = j * yStride;
|
||
for (var i = 0; i < w; i += 16) {
|
||
final v = y[row + i];
|
||
if (v < yMin) yMin = v;
|
||
if (v > yMax) yMax = v;
|
||
sum += v;
|
||
n++;
|
||
}
|
||
}
|
||
return (yMin, yMax, sum / n);
|
||
}
|
||
|
||
/// 按指定色序推理一次，返回 (检测数, 最高分, 总分)
|
||
(int, double, double) _runWithSwap(
|
||
List<Uint8List> planes, List<int> strides, int w, int h, bool swap) {
|
||
_yuvSwapChroma = swap;
|
||
preprocess(planes: planes, strides: strides, width: w, height: h,
|
||
isBgra: false);
|
||
_interpreter.run(_input.buffer.asUint8List(), _output);
|
||
final dets = postprocess();
|
||
var maxScore = 0.0, sumScore = 0.0;
|
||
for (final d in dets) {
|
||
sumScore += d.score;
|
||
if (d.score > maxScore) maxScore = d.score;
|
||
}
|
||
return (dets.length, maxScore, sumScore);
|
||
}
|
||
|
||
bool _brightRegionLeansRed(
|
||
List<Uint8List> planes, List<int> strides, int w, int h, int yMax) {
|
||
final y = planes[0];
|
||
final yStride = strides[0];
|
||
final uvStride = strides[1];
|
||
final nv12 = planes.length == 2;
|
||
final uv = nv12 ? planes[1] : null;
|
||
// 最亮带（maxY-40 以上），整体偏暗的场景也能拿到足量样本
|
||
final brightMin = yMax - 40;
|
||
var blue = 0, red = 0;
|
||
for (var j = 0; j < h; j += 8) {
|
||
final yrow = j * yStride;
|
||
for (var i = 0; i < w; i += 8) {
|
||
if (y[yrow + i] < brightMin) continue;
|
||
final cj = j ~/ 2, ci = i ~/ 2;
|
||
if (nv12) {
|
||
final c = cj * uvStride + ci * 2;
|
||
if (c + 1 >= uv!.length) continue;
|
||
if (uv[c] > 150) blue++;
|
||
if (uv[c + 1] > 150) red++;
|
||
} else {
|
||
final c = cj * uvStride + ci;
|
||
if (c >= planes[1].length || c >= planes[2].length) continue;
|
||
if (planes[1][c] > 150) blue++;
|
||
if (planes[2][c] > 150) red++;
|
||
}
|
||
}
|
||
}
|
||
// 亮区偏红多于偏蓝 → 当前 U/V 假设反了
|
||
return red > blue;
|
||
}
|
||
|
||
static double _bl(double a, double b, double c, double d, double fx,
|
||
double fy) =>
|
||
(1 - fx) * (1 - fy) * a + fx * (1 - fy) * b +
|
||
(1 - fx) * fy * c + fx * fy * d;
|
||
|
||
List<DetectionResult> postprocess() {
|
||
final out = _output[0];
|
||
final boxes = <DetectionResult>[];
|
||
for (var a = 0; a < _numAnchors; a++) {
|
||
final cx = out[0][a];
|
||
final cy = out[1][a];
|
||
final w = out[2][a];
|
||
final h = out[3][a];
|
||
var bestCls = 0;
|
||
var bestScore = 0.0;
|
||
for (var c = 0; c < _numClasses; c++) {
|
||
final s = out[4 + c][a];
|
||
if (s > bestScore) {
|
||
bestScore = s;
|
||
bestCls = c;
|
||
}
|
||
}
|
||
final label =
|
||
bestCls < _labels.length ? _labels[bestCls] : 'unknown';
|
||
// 低分池保留，供运动检测提升显示
|
||
if (bestScore < minScore) continue;
|
||
boxes.add(DetectionResult(
|
||
label: label,
|
||
score: bestScore,
|
||
left: (cx - w / 2).clamp(0.0, 1.0),
|
||
top: (cy - h / 2).clamp(0.0, 1.0),
|
||
right: (cx + w / 2).clamp(0.0, 1.0),
|
||
bottom: (cy + h / 2).clamp(0.0, 1.0),
|
||
modelId: modelId,
|
||
modelName: modelName,
|
||
));
|
||
}
|
||
final kept = nms(boxes, iouThreshold);
|
||
return kept.take(maxDetections).toList();
|
||
}
|
||
|
||
/// 诊断：按指定字节序推理一次，返回 (检测数, 最高分)。
|
||
/// 用于对比 BGRA/RGBA 两种顺序在同一帧上的检测差异（验证字节序与场景可达性）。
|
||
(int, double) diagnoseOrder({
|
||
required List<Uint8List> planes,
|
||
required List<int> strides,
|
||
required int width,
|
||
required int height,
|
||
required bool isBgra,
|
||
required bool rgbaOrder,
|
||
}) {
|
||
preprocess(
|
||
planes: planes,
|
||
strides: strides,
|
||
width: width,
|
||
height: height,
|
||
isBgra: isBgra,
|
||
rgbaOrder: rgbaOrder);
|
||
_interpreter.run(_input.buffer.asUint8List(), _output);
|
||
final dets = postprocess();
|
||
var maxScore = 0.0;
|
||
for (final d in dets) {
|
||
if (d.score > maxScore) maxScore = d.score;
|
||
}
|
||
return (dets.length, maxScore);
|
||
}
|
||
|
||
void dispose() => _interpreter.close();
|
||
}
|