YoloV8Translator加载模型执行后结果错误有偏移，通过YoloV8TranslatorFactory加载模型则没有问题，这个是什么原因，怎么解决？

[zhu-weixin]

我想请问是我下面的通过YoloV8Translator加载模型的处理有问题吗，还是后续对图片的处理错误，导致出现这个问题？

下面是我分别通过两种方式加载模型，getPredictor1是通过YoloV8TranslatorFactory加载模型，getPredictor2是通过YoloV8Translator加载模型的代码

@Slf4j
public class Main4 {

    public static void main(String[] args) throws IOException, ModelException, TranslateException {
        Path imgPath = Paths.get("D:/Project/aggreg/image/src/main/resources/img/bus.jpg");

        try {
            Predictor<Image, DetectedObjects> predictor1 = getPredictor1();
            Predictor<Image, DetectedObjects> predictor2 = getPredictor2();
            Path outputPath = Paths.get("D:/Project/aggreg/image/src/main/resources/img/");
            Files.createDirectories(outputPath);
            //第一种构建方式
            Image img1 = ImageFactory.getInstance().fromFile(imgPath);
            DetectedObjects detection1 = predictor1.predict(img1);
            if (detection1.getNumberOfObjects() > 0) {
                img1.drawBoundingBoxes(detection1);
                Path output = outputPath.resolve("yolo_detected_1.png");
                try (OutputStream os = Files.newOutputStream(output)) {
                    img1.save(os, "png");
                }
                log.info(">>>>>>>>>>>>>>>>>>>>>>>>>>>");
                log.info("Detected object1: {}", detection1);
                log.info("Detected object saved in: {}", output);
                log.info(">>>>>>>>>>>>>>>>>>>>>>>>>>>");
            }
            //第二种构建方式
            Image img2 = ImageFactory.getInstance().fromFile(imgPath);
            DetectedObjects detection2 = predictor2.predict(img2);
            if (detection2.getNumberOfObjects() > 0) {
                img2.drawBoundingBoxes(detection2);
                Path output = outputPath.resolve("yolo_detected_2.png");
                try (OutputStream os = Files.newOutputStream(output)) {
                    img2.save(os, "png");
                }
                log.info(">>>>>>>>>>>>>>>>>>>>>>>>>>>");
                log.info("Detected object2: {}", detection2);
                log.info("Detected object2 saved in: {}", output);
                log.info(">>>>>>>>>>>>>>>>>>>>>>>>>>>");
            }
        } catch (Exception e) {
            e.printStackTrace();
        }
    }

    /**
     * 通过YoloV8TranslatorFactory加载模型
     *
     * @return
     * @throws ModelNotFoundException
     * @throws MalformedModelException
     * @throws IOException
     */
    public static Predictor<Image, DetectedObjects> getPredictor1() throws ModelNotFoundException, MalformedModelException, IOException {
        Criteria<Image, DetectedObjects> criteria =
                Criteria.builder()
                        .setTypes(Image.class, DetectedObjects.class)
                        .optDevice(Device.gpu())
                        .optModelUrls(Main4.class.getResource("/yolo").getPath())
                        .optModelName("yolo11n.onnx")
                        .optEngine("OnnxRuntime")
                        .optArgument("width", 640)
                        .optArgument("height", 640)
                        .optArgument("resize", true)
                        .optArgument("toTensor", true)
                        .optArgument("applyRatio", true)
                        .optArgument("threshold", 0.6f)
                        .optTranslatorFactory(new YoloV8TranslatorFactory())
                        .optProgress(new ProgressBar())
                        .build();
        ZooModel<Image, DetectedObjects> model = criteria.loadModel();
        Predictor<Image, DetectedObjects> predictor = model.newPredictor();
        return predictor;
    }

    /**
     * 通过YoloV8Translator加载模型
     *
     * @return
     * @throws ModelNotFoundException
     * @throws MalformedModelException
     * @throws IOException
     */
    public static Predictor<Image, DetectedObjects> getPredictor2() throws ModelNotFoundException, MalformedModelException, IOException {
        Translator<Image, DetectedObjects> translator = YoloV8Translator.builder()
                .addTransform(new Resize(640, 640))
                .addTransform(new ToTensor())
                .optApplyRatio(true)
                .optThreshold(0.6f)
                .optSynsetArtifactName("synset.txt")
                .build();
        Criteria<Image, DetectedObjects> criteria =
                Criteria.builder()
                        .setTypes(Image.class, DetectedObjects.class)
                        .optDevice(Device.gpu())
                        .optModelUrls(Main4.class.getResource("/yolo").getPath())
                        .optModelName("yolo11n.onnx")
                        .optEngine("OnnxRuntime")
                        .optArgument("maxBox", 1000)
                        .optTranslator(translator)
                        .optProgress(new ProgressBar())
                        .build();
        ZooModel<Image, DetectedObjects> model = criteria.loadModel();
        Predictor<Image, DetectedObjects> predictor = model.newPredictor();
        return predictor;
    }
}

下面是我的控制台打印的结果（当然我知道有内存溢出的问题）：

18:28:56.194 [main] INFO xyz.zwxin.image.service.yolo.Main4 - >>>>>>>>>>>>>>>>>>>>>>>>>>>
18:28:56.194 [main] INFO xyz.zwxin.image.service.yolo.Main4 - Detected object1: [
	{"className": "bus", "probability": 0.90587, "boundingBox": {"rect":[0.01778717041015625,0.2105330228805542,0.9935930252075195,0.685427975654602]}},
	{"className": "person", "probability": 0.89490, "boundingBox": {"rect":[0.06084575653076172,0.36660306453704833,0.29915738105773926,0.8391817331314086]}},
	{"className": "person", "probability": 0.89393, "boundingBox": {"rect":[0.8278806686401368,0.3668684959411621,0.9998627662658692,0.814603328704834]}},
	{"className": "person", "probability": 0.88770, "boundingBox": {"rect":[0.2755547523498535,0.37552340030670167,0.42212188243865967,0.7971038579940797]}},
	{"className": "person", "probability": 0.72946, "boundingBox": {"rect":[1.4168620109558105E-4,0.5128156661987304,0.09215204119682313,0.8079570293426513]}}
]

18:28:56.226 [main] INFO xyz.zwxin.image.service.yolo.Main4 - Detected object saved in: D:\Project\aggreg\image\src\main\resources\img\yolo_detected_1.png
18:28:56.226 [main] INFO xyz.zwxin.image.service.yolo.Main4 - >>>>>>>>>>>>>>>>>>>>>>>>>>>
18:29:01.987 [main] INFO xyz.zwxin.image.service.yolo.Main4 - >>>>>>>>>>>>>>>>>>>>>>>>>>>
18:29:01.987 [main] INFO xyz.zwxin.image.service.yolo.Main4 - Detected object2: [
	{"className": "bus", "probability": 0.90590, "boundingBox": {"rect":[0.050815309797014506,0.6015426090785435,2.8388173239571706,1.9583563123430525]}},
	{"className": "person", "probability": 0.89488, "boundingBox": {"rect":[0.17384161267961776,1.047433308192662,0.8547286987304688,2.3976612091064453]}},
	{"className": "person", "probability": 0.89395, "boundingBox": {"rect":[2.3653640747070312,1.0481888907296317,2.856751033238002,2.3274388994489397]}},
	{"className": "person", "probability": 0.88761, "boundingBox": {"rect":[0.7872960226876395,1.0729188919067383,1.2060600008283342,2.277428899492536]}},
	{"className": "person", "probability": 0.72934, "boundingBox": {"rect":[4.0481771741594585E-4,1.465179443359375,0.2632885490145002,2.308438846043178]}}
]

18:29:01.997 [main] INFO xyz.zwxin.image.service.yolo.Main4 - Detected object2 saved in: D:\Project\aggreg\image\src\main\resources\img\yolo_detected_2.png
18:29:02.008 [main] INFO xyz.zwxin.image.service.yolo.Main4 - >>>>>>>>>>>>>>>>>>>>>>>>>>>

yolo_detected_1.png的图片是这样的：

yolo_detected_2.png的图片是这样的：

[frankfliu]

@zhu-weixin
Thanks for reporting this issue. There is a bug in BaseImageTranslator.builder() class. This was introduced when refactor Yolov8TranslatorFacotry. We will raise a PR a short after.

For the time being, you can use the following workaround:

        Map<String, Object> args = new ConcurrentHashMap<>();
        args.put("width", 640);
        args.put("height", 640);
        args.put("resize", true);
        args.put("toTensor", true);
        args.put("applyRatio", true);
        args.put("threshold", 0.6f);
        args.put("maxBox", 1000);

        Translator<Image, DetectedObjects> translator = YoloV8Translator.builder(args).build();