Some questions about my custom datasets

[bjutlss]

According to README, I made my own dataset using ObjectDatasetsTools and modified the attributes in obj.data (such as camera parameters, diameters, etc.) and got the following results via train.py:

-----------------------------------
  tensor to cuda : 0.000346
    forward pass : 0.012606
get_region_boxes : 0.003043
 prediction time : 0.015995
            eval : 0.112702
-----------------------------------
2023-02-15 11:57:06 Results of deli
2023-02-15 11:57:06    Acc using 5 px 2D Projection = 91.13%
2023-02-15 11:57:06    Acc using 10% threshold - 0.007801272195673763 vx 3D Transformation = 71.29%
2023-02-15 11:57:06    Acc using 5 cm 5 degree metric = 95.09%
2023-02-15 11:57:06    Mean 2D pixel error is 2.552845, Mean vertex error is 0.006240, mean corner error is 4.039984
2023-02-15 11:57:06    Translation error: 0.006158 m, angle error: 2.358766 degree, pixel error:  2.552845 pix

But when I predict an image from the same camera in real time, the results are always bad.I'm not sure where is the problem, you can provide some help?