Compatibility checklist when LiDAR point order ≠ TJ4D

[qiujialiang]

I am using a 4-D mmWave radar dataset whose point layout is
[range, azimuth, elevation, doppler, power, x, y, z] (8-D),
while the original TJ4D format is [X, Y, Z, V_r, Range, Power, Alpha, Beta] (4-D).
Training runs without crash, but test mAP drops to ~30 %, indicating that feature order is silently ignored.
I would like a canonical list of places that must be updated so that both training and inference consume the correct 8-D layout.

[13125548701]

我使用的是一个 4-D 毫米波雷达数据集，其点布局为 [范围、方位角、仰角、多普勒、功率、x、y、z] （8-D），而原始 TJ4D 格式为 [X、Y、Z、V_r、范围、功率、Alpha、Beta] （4-D）。训练运行时没有崩溃，但测试 mAP 下降到 ~30%，这表明特征顺序被静默忽略。我想要一个必须更新的地方的规范列表，以便训练和推理都使用正确的 8-D布局。

我也遇到这个问题啦，您解决了吗