There is no new message. I need to make clear about the steps to run the repo

[trungUTELai]

I have an issue while running the repo. As I have ran the roslaunch d2vins quadcam.launch (for Omni-VINS only case) nothing happened except the appearance of topic names (occurred when I checked rostopic list). However, when I tried to echo or hz it, there was no message published even with topic /oak_ffc_4p/assemble ).
I wonder if it needs any thing to run the code, for example starting the oak_ffc_4p_ros_driver first?

[trungUTELai]

Currently, I have to run the oak_ffc_4p_ros_driver first, then the code in this repo start running.
Another problem related to the latency, why the latency seem so high:
I ran the command: roslaunch d2vins quadcam.launch depth_gen:=true enable_pgo:=true show:=true

[Peize-Liu]

Your device only has 15W, please change to MAXN

[trungUTELai]

@Peize-Liu , Hello, thank you for the advice and appreciate for the fast replication.
I have changed the system and the performance enhanced(from ~500ms to ~280ms time cost). However, It still so slow for real time application.

In fact, when I tried to run without enable depth_gen. It's great for real time (time cost ~13ms)

I am using jetson orin nx, and cameras from oak-ffc-4p only.
I am running the code with the available models in the branch.
Can you give me some guidance for how to improve the system performance.

[trungUTELai]

Regarding to the official paper
Now, I will try to convert the HITNET model to tensorrt + fp16 to achieve the 15Hz frequency as the paper told.

Moreover, one of the issue from D2SLAM have tried to convert NetVlad to tensorrt+fp16 and achived the 6ms inference time. Have you tried it before? If yes, was the trade-off is acceptable ?

[Peize-Liu]

The existing code will automatically convert the HITNET Onnx to the tensorRT Engine.
Can you show more information about your platform? Like nvidia-smi.
For NetVlad, I also implemented it before, but the performance improvement is not that evident.

[trungUTELai]

Yes, currently, my jetson don't have nvidia-smi. Is it necessary to install nvidia driver on jetson orin ?

Currently, I am running the the dockerFile

Here is some information about my platform showed by jtop:

[trungUTELai]

Thank you for your support, your code run as good as expected, I have some mistake while setup the configuration so the performance is not good as expected.

[Peize-Liu]

Good to know.