onnxruntime 部署部分为什么是将角度作为类回归，这样我测试了没有原本模型检测准确度

[LHJ-YM]

def preprocess(self, img, new_shape):
    img = self.letterbox(img, new_shape, auto=False)[0]
    img = img.transpose((2, 0, 1))[::-1]  # HWC to CHW, BGR to RGB
    img = np.ascontiguousarray(img).astype('float32')
    img /= 255  # 0 - 255 to 0.0 - 1.0
    if len(img.shape) == 3:
        img = img[None]  # expand for batch dim
    return img

def postprecess(self, prediction, src_img, new_shape):

    nc = prediction.shape[2] - 5 - 180  # number of classes

    xc = prediction[..., 4] > CONF_THRES
    outputs = prediction[:][xc]

    generate_boxes, bboxes, scores = [], [], []

    for out in outputs:

        cx, cy, longside, shortside, obj_score = out[:5]
        class_scores = out[5: 5+nc]
        class_idx = np.argmax(class_scores)

        max_class_score = class_scores[class_idx] * obj_score
        if max_class_score < CONF_THRES:
            continue

        theta_scores = out[5+nc:]
        theta_idx = np.argmax(theta_scores)
        theta_pred = (theta_idx - 90) / 180 * PI

        bboxes.append([[cx, cy], [longside, shortside], max_class_score])
        scores.append(max_class_score)
        generate_boxes.append([
            cx, cy, longside, shortside, 
            theta_pred, max_class_score, class_idx
        ])

    indices = cv2.dnn.NMSBoxesRotated(
        bboxes, scores, CONF_THRES, NMS_THRES
    )
    det = np.array(generate_boxes)[indices.flatten()]

    pred_poly = self.rbox2poly(det[:, :5])

    pred_poly = self.scale_polys(new_shape, pred_poly, src_img.shape)
    det = np.concatenate((pred_poly, det[:, -2:]), axis=1) # (n, [poly conf cls])

    for *poly, conf, cls in reversed(det):
        c = int(cls)
        label = f'{CLASSES[c]} {conf:.2f}'
        self.poly_label(src_img, poly, label, COLORS[c])
    
    cv2.imshow('Result', src_img)
    cv2.waitKey(0)
    cv2.destroyAllWindows()