How long does deployment take, and why is the first run slow?

Deployment takes about 10 minutes. The first run builds an inference engine for this GPU on the device, about 8 minutes; it is kept locally and reused on restart.

Does it work offline, and is video sent to the cloud?

After deployment it runs offline; the first deploy needs internet access to pull the image. Depth is estimated on the Jetson GPU and the stream stays on the LAN.

How accurate is the depth? Can it measure distance?

It is a single-camera model estimate, suited to judging relative distance, prototypes and demos. For reliable ranging or safety-related decisions, add a measuring sensor such as lidar, stereo or ToF.

How do I get depth results into my application?

The default exit is an RTSP stream for viewing. Inside the container runs a ROS 2 pipeline, where the depth map and point cloud live; subscribe to those nodes for raw data, or add your own data interface inside the container.

The stream stutters or does not show. What should I check?

Open rtsp://device-ip:8554/depth in VLC first to confirm the stream exists. If you see only the raw camera picture, the stream source is not the inference output; if the log cannot read frames, set the camera index to auto or 1 and redeploy.

Can I change the model or the output?

Frame rate, bitrate, panel size and point count are deploy-time parameters. Another Depth Anything model means replacing the model file and rebuilding the inference engine on the device.

What should we check on licensing before commercial use?

The deployment scripts ship with this design, and the model weights keep the Depth Anything publisher's terms. Before shipping a closed product, confirm the licence on the weights you use or swap in a compatibly-licensed model.

Depth Anything V3 on Jetson

One-click SSH deployment of Depth Anything V3 on NVIDIA Jetson with a prebuilt Docker image.

Beginner10minVision & Video AISource Code

Reusable Site Templates

Teams that need distance estimates from an ordinary camera, and want the depth output running before they build the application.

Robot and mobile-platform prototypes — estimate how far objects are from a single camera to trial obstacle avoidance or grasping.
Trade-show and customer demos — one stream shows the original picture, a colour depth map and a 3D point cloud side by side.
Repeatable test setups — the same image deploys to several Jetsons with one click, so every unit matches.

Not a fit: millimetre-grade ranging or anything needing safety certification. Depth here comes from a single-camera model estimate, not from lidar or stereo.

How it works

Depth Anything V3 device composition: a USB camera connects over USB to a reComputer J40 Series host, which sends an RTSP stream to a player or preview page

Two devices: a USB camera and a Jetson host. Depth estimation runs on the host GPU, and the result goes out as a video stream on the LAN, viewable in any RTSP player or the platform preview.

One USB camera plugged into the Jetson, captured at 640x480 by default.

Camera found automatically: the camera index is auto-detected; set 0 or 1 by hand if detection fails.
One camera per host: the default flow takes a single stream.

The depth model, inference engine and streaming service all run on this machine, packaged as one Docker container.

Item	Detail
Host	reComputer J40 Series (Jetson Orin NX)
Model	Depth Anything V3 Metric Large
Prerequisites	Docker and the NVIDIA container runtime installed, SSH reachable
First deploy	Builds the inference engine on the device, about 8 minutes; restarts reuse it

How well it works

The table lists what this design provides on the reComputer J40 Series, and its limits.

What it does	Limit
Estimates depth from a USB camera in real time, rendering a colour depth map and a 3D point cloud	Single-camera estimation: relative distance is usable, but it does not replace lidar ranging
Publishes the result as an RTSP stream that ordinary players can open	Streams at 6 fps by default, for viewing and demos rather than as a control-loop data feed
Starts with the device, and restarts reuse the inference engine	The first deploy builds the inference engine on the device, about 8 minutes

What you get out of it

Once deployed, this is where the result comes out.

Exit	What comes out	Notes
RTSP 8554, path /depth	One frame combining the original picture, a colour depth map and a 3D point cloud	6 fps and 1500 kbps by default, adjustable at deploy time

The stream stays on the LAN. To feed depth data into your own program, read it from the ROS 2 nodes inside the container; the design does not expose a data API by default.

How to deploy

Check the setup and its product list, then follow the deployment guide.

Deployment Options

Jetson One-Click Depth Demo

Deploy Depth Anything V3 to Jetson via SSH without manual command input.

2 steps · ~10min

Open the deployment guide →

Edge Device

recomputer_j40

FAQ

One reComputer J40 Series host and a USB camera. The host needs Docker and the NVIDIA container runtime, and must accept SSH. The container is capped at 4GB of memory.

Didn't find your answer?

We Are Glad to Be Your Hardware Partner !