Successfully Reproduced WildGS-SLAM with 8GB GPU

[ureeey]

Successfully Reproduced WildGS-SLAM

I have successfully run WildGS-SLAM in my environment.
Here is my environment and the configuration I set, hope it helps others.

Environment

Item	Detail
Machine	Dell G15 5530 Laptop
GPU	NVIDIA RTX4060 Laptop 8GB
OS	Ubuntu 22.04 + 6.0.0-1020-oem kernel + Xfce X11 gdm3
NVIDIA DRIVER	535-server

Configuration

Item	Detail
gui	False
cam H_out	240
cam W_out	400
depth	metric3d_vit_small
feature_extractor	dinov2_vits14

GPU memory usage

Peak usage reaches 5–6 GB in the final stage, and is around 3 GB during tracking.

Result

python run.py ./configs/Dynamic/Wild_SLAM_Mocap/crowd_demo.yaml work well.
Thanks to the authors for the great work!

[Jianhao-zheng]

Thank you for sharing!

[ureeey]

I'd like to contribute to the project by adapting it for new hardware, with a plan to deploy it on a specific Jetson model. Do you have any advice?

[Jianhao-zheng]

It will depend on how much GPU vRAM the Jetson has. I see you successfully make it run in lower GPU memory by using reduced camera resolution. At least, it should be able to run. However, reducing resolution will always lead to worse performance. You can do some tests with different resolutions.

Meanwhile, I have a new work that can run much faster than WildGS-SLAM, while having similar/better tracking results. But that method only does camera tracking. There is no online mapping. I will release it this month.

[ureeey]

Thanks for your detailed reply and suggestions! I will do some tests.

I am really looking forward to your new work! Could you please confirm two points:

1. Does the new work use the same uncertainty method as WildGS-SLAM?
​
2. Is it still developed and tested on an RTX 4090 desktop GPU?

Thank you again for sharing your progress!

[Jianhao-zheng]

1. No, the bottleneck of the running time is jointly optimizing the uncertainty MLP and the map. So, I pretrained a motion mask detector to directly output a motion mask from a pair of images.
2. Yes.

[ureeey]

Thanks! Looking forward to your new work.

…

---- Replied Message ---- | From | Jianhao ***@***.***> | | Date | 03/17/2026 06:16 | | To | ***@***.***> | | Cc | jinchunyang ***@***.***>***@***.***> | | Subject | Re: [GradientSpaces/WildGS-SLAM] Successfully Reproduced WildGS-SLAM with 8GB GPU (Issue #61) | Jianhao-zheng left a comment (GradientSpaces/WildGS-SLAM#61) No, the bottleneck of the running time is jointly optimizing the uncertainty MLP and the map. So, I pretrained a motion mask detector to directly output a motion mask from a pair of images. Yes. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[maheshreddyz]

Hey @Jianhao-zheng,
How can a motion mask detector be pretrained? Whether the object is moving or not will only be determined by the current scene settings, right? Is it based on the motion between the 2 frames? And how was it optimised, I mean, what was the ground truth?

[maheshreddyz]

Thanks for your detailed reply and suggestions! I will do some tests.

I am really looking forward to your new work! Could you please confirm two points:

1. Does the new work use the same uncertainty method as WildGS-SLAM?​2. Is it still developed and tested on an RTX 4090 desktop GPU?

Thank you again for sharing your progress!

Hey, I have a similar idea of running this setup on Jetson. I was wondering, wouldn't it be difficult to run the dino on Jetson? might it be too heavy?

[ureeey]

Hey, @maheshzebu , I think the VRAM should be okay, but FPS might drop a bit because of Jetson’s CUDA and TensorCore limitations.Worth trying anyway! Which Jetson model are you going to use?

[maheshreddyz]

Hey, @maheshzebu , I think the VRAM should be okay, but FPS might drop a bit because of Jetson’s CUDA and TensorCore limitations.Worth trying anyway! Which Jetson model are you going to use?

I am planning to use orin nx 16 gb, how much is it taking to run on the 8GB gpu? what is the computation time?