superpoint在jetson NX平台的推理耗时问题

[ustccxy]

您好，我们在jetson NX平台上进行测试。使用develop分支构建镜像。使用realsense D435i双目相机。 开启tensonrt加速，进行superpoint前向推理时，单张图片的耗时在30ms以上，影响到了实时性。而在jetson orin上，单张推理的耗时在15ms左右，想请问你们在进行测试时的耗时情况，以及如何解决推理耗时问题。

[allegorywrite]

Hi, I have the same problem with Xavier NX. Our result is written below

Without tensorRT：

SuperPoint: 62.0ms
NetVLAD: 15.5ms
SuperGlue: 83.1ms

With tensorRT_fp16：

SuperPoint: 33.9ms
NetVLAD: 5.9ms
SuperGlue: 80.1ms

With tensorRT_int8：

SuperPoint: 28.0ms

If we believe the data above, when using a stereo vision camera, two superpoints process (left and right) are required per timestep, so the processing speed cannot exceed 1000ms / 2*30 ms = 16.7 Hz. Is this correct? In the paper, the authors say they tested with Nvidia Xavier NX, but what were the imu and image frequencies required to get a stable odometry? It seems that @ustccxy 's question has been unanswered for a long time, can anyone answer this question?

[allegorywrite]

@xuhao1 @Peize-Liu