使用SwiftUI展示VNRecognizedObjectObservation边界框时偏移错误

我使用Vision框架进行物体检测，获取到[VNRecognizedObjectObservation]后，会对归一化矩形做转换再展示：

let transform = CGAffineTransform(scaleX: 1, y: -1).translatedBy(x: 0, y: -CGFloat(height))
VNImageRectForNormalizedRect(normalizedRect, width, height) // 因使用SwiftUI展示，应用该变换
    .applying(transform)

宽度和高度取自SwiftUI的GeometryReader：

Image(...)
    .resizable()
    .scaledToFit()
    .overlay {
        GeometryReader { geometry in // 后续通过ZStack和ForEach遍历[VNRecognizedObjectObservation]，id取\.uuid，代码如下：
            let calculatedRect = calculateRect(boundingBox, geometry)
            Rectangle()
                .frame(width: calculatedRect.width, height: calculatedRect.height)
                .offset(x: calculatedRect.origin.x, y: calculatedRect.origin.y)
        }
    }

但目前存在问题：即使是正方形图片，多数边界框位置都不正确（仅少数准确）。该问题与模型无关——使用同一个MLModel处理同一张图片，在Xcode模型预览中边界框定位非常精准。

应用内展示效果：

Xcode模型预览效果：

更新：最小可复现代码 #

在macOS SwiftUI项目的ContentView.swift中使用以下代码，并将YOLOv3Tiny.mlmodel添加至项目包，即可复现相同问题：

import SwiftUI
import Vision
import CoreML

class Detection: ObservableObject {
    let imgURL = URL(string: "https://i.imgur.com/EqsxxTc.jpg")! // Xcode预览生成的对应图片：https://i.imgur.com/6IPNQ8b.png
    @Published var objects: [VNRecognizedObjectObservation] = []

    func getModel() -> VNCoreMLModel? {
        if let modelURL = Bundle.main.url(forResource: "YOLOv3Tiny", withExtension: "mlmodelc") {
            if let mlModel = try? MLModel(contentsOf: modelURL, configuration: MLModelConfiguration()) {
                return try? VNCoreMLModel(for: mlModel)
            }
        }
        return nil
    }

    func detect() async {
        guard let model = getModel(), let tiff = NSImage(contentsOf: imgURL)?.tiffRepresentation else {
            fatalError("Either YOLOv3Tiny.mlmodel is not in project bundle, or image failed to load.")
            // YOLOv3Tiny模型获取地址：https://ml-assets.apple.com/coreml/models/Image/ObjectDetection/YOLOv3Tiny/YOLOv3Tiny.mlmodel
        }
        let request = VNCoreMLRequest(model: model) { (request, error) in
            DispatchQueue.main.async {
                self.objects = (request.results as? [VNRecognizedObjectObservation]) ?? []
            }
        }
        try? VNImageRequestHandler(data: tiff).perform([request])
    }

    func deNormalize(_ rect: CGRect, _ geometry: GeometryProxy) -> CGRect {
        let transform = CGAffineTransform(scaleX: 1, y: -1).translatedBy(x: 0, y: -CGFloat(geometry.size.height))
        return VNImageRectForNormalizedRect(rect, Int(geometry.size.width), Int(geometry.size.height)).applying(transform)
    }
}

struct ContentView: View {
    @StateObject var detection = Detection()

    var body: some View {
        AsyncImage(url: detection.imgURL) { img in
            img.resizable().scaledToFit().overlay {
                GeometryReader { geometry in
                    ZStack {
                        ForEach(detection.objects, id: \.uuid) { object in
                            let rect = detection.deNormalize(object.boundingBox, geometry)
                            Rectangle()
                                .stroke(lineWidth: 2)
                                .foregroundColor(.red)
                                .frame(width: rect.width, height: rect.height)
                                .offset(x: rect.origin.x, y: rect.origin.y)
                        }
                    }
                }
            }
        } placeholder: {
            ProgressView()
        }
        .onAppear {
            Task { await self.detection.detect() }
        }
    }
}

编辑补充： 经进一步测试确认，Vision返回的位置数据完全正确，deNormalize()函数计算出的矩形位置和尺寸也无问题，因此问题根源必然在SwiftUI的展示逻辑中。

内容的提问来源于stack exchange，提问作者Swiftly