Running on an embedded platform

[dapao-0808]

May I ask, what is the average processing time for an image on an embedded platform? I spent about 700 milliseconds on Nvidia Orin nx

[mingwz]

tensorrt ？

[meyiao]

@dapao-0808 I have tested it on rk3562 which is much weaker than Orin Nx, it takes ~60ms per frame.

[zw-92]

@dapao-0808 I have tested it on rk3562 which is much weaker than Orin Nx, it takes ~60ms per frame.

Hello, I'm also trying to deploy it on the rk3562. The feature extraction of a single image takes about 60ms, and there are some other post-processing procedures which are very time-consuming. How much time does it take for you in total? Also, I previously evaluated your C++ code on MegaDepth1500 and found that the accuracy was a bit lower than that of the xfeat_match in the original paper. Now I'm trying to deploy the xfeat_match_star version on rk3562. How to evaluate the accuracy of the deployed version is also a problem. I hope we can communicate.

[meyiao]

@dapao-0808 I have tested it on rk3562 which is much weaker than Orin Nx, it takes ~60ms per frame.

Hello, I'm also trying to deploy it on the rk3562. The feature extraction of a single image takes about 60ms, and there are some other post-processing procedures which are very time-consuming. How much time does it take for you in total? Also, I previously evaluated your C++ code on MegaDepth1500 and found that the accuracy was a bit lower than that of the xfeat_match in the original paper. Now I'm trying to deploy the xfeat_match_star version on rk3562. How to evaluate the accuracy of the deployed version is also a problem. I hope we can communicate.

Hi, my own test shows that the accuracy loss is negligible, I wouldn't suggest pay too much attention to it. As for the post-process（point score softmax, nms, descriptor interpolation), it cost ~3ms on my PC, I guess it might cost ~20ms on the RK3562. The mutual-NN feature matching is time consuming also.