Your depth topics show up in rostopic list, but rostopic hz prints "no new messages", rostopic echo returns nothing, and rviz stays blank. The setup behind it is common: the camera driver runs in one Docker container, your tools run in another, and both point at the same ROS master. The cause is almost never ROS itself. A USB depth camera lets exactly one Linux process own the device, and a container is just another process on the same kernel. You get the test sequence that separates device ownership from container networking, three topologies that fix it, and the bandwidth arithmetic that tells you whether images can cross containers once the graph is healthy.
Symptom: the master answers, but no image data moves
The signature looks like this. Container one starts the driver with depth-to-color alignment enabled. Container two sets ROS_MASTER_URI to the host IP on port 11311 and finds every topic, including one named like /camera/aligned_depth_to_color/image_raw. Yet echo on that topic prints nothing, hz prints "no new messages" forever, and rviz draws no frames. docker ps shows both containers Up, so nothing has crashed.
Two separate connections are in play, and only one of them is being tested. The list command talks to the master. The image data flows on a direct connection from the publishing node to each subscriber. A topic can be listed, subscribed and still deliver zero bytes, because the publisher has no frame to send or because the subscriber cannot reach the publisher's endpoint. The three causes below cover both paths in the order they turn up.
Cause 1: one process owns the camera, containers included
Depth camera drivers open the device exclusively: one process on the machine holds the camera, and every other open fails. That policy lives in the vendor driver, below Docker. Containers share the host kernel, so --privileged and a mount of /dev or /dev/bus/usb change permissions, not ownership. A container that can reach the device node still has to win an open nobody else holds.
That is why list succeeds while hz starves. A second driver instance, a viewer left open on the host, or a stale container holds the camera; your driver's open fails; it publishes no frames. The topic names still register at startup, echo waits politely, and hz prints its familiar line.
Test by elimination. Stop every container that touches the camera, start exactly one driver, and run hz inside that same container. If the rate appears, the camera side was the problem and you can move subscribers out with confidence. To find the holder before you restart, list the processes that keep the video device nodes open; the standard tools that map open files to processes name the driver, the viewer or the stale container directly.
The same rule bites without Docker: one owner per USB device also blocks a color stream when another process already holds the camera, which the one-owner-per-USB-device fix for OpenCV color streams walks through.
Cause 2: the publisher's endpoint is unreachable from the second container
When the camera is publishing, the suspect moves to the data connection. The master answers requests on port 11311 from any container that can route to the host, but the publishing node advertises its own address for the direct subscriber connection. Under the default bridge network that address lives inside the container's network namespace, and your second container cannot reach it. List works; echo and hz do not.
A test publisher settles it in one minute. Run a known publisher inside the driver container, with the roscpp_tutorials babbler node started by rosrun as the usual choice, and echo it from the second container. Messages arriving means the graph and routing are healthy, which moves suspicion back to the camera. Nothing arriving means this cause: run both sides with host networking (--net=host), or keep publisher and subscriber in one container.
Ethernet-linked 3D cameras hit the same container boundary in their own way; the Ethernet 3D camera in Docker fix covers discovery and streaming across it.
Cause 3: a stale driver process still holds the device
The third pattern appears after a day of restarts. A container exits without releasing the device, or a driver from an earlier session still runs on the host. The new driver fails its open with little console noise and publishes nothing. Many depth cameras show a status LED while a driver owns them, so a dark LED next to a nominally running driver is a reliable tell.
Check docker ps for containers that are Up but should not be, stop them, and confirm nothing holds the device nodes before the next start. Then run hz inside the driver container before you add subscribers, so every topology change starts from a camera you have just seen streaming.
The test sequence, in order
- Publish a test topic from the driver container and echo it from the second container; the babbler tutorial node is enough. Arrives: continue. Silent: fix the network first, with host networking on both sides or one container for both roles.
- Guarantee a single owner. Stop every driver on the host and in every container, confirm nothing holds the device nodes, then start exactly one driver.
- Run rostopic hz on the aligned depth topic inside the driver container. A steady rate in the tens of hz proves camera and driver are healthy.
- Move the subscriber to its own container, keeping the driver as the sole owner. If the rate survives the move, the topology is right.
Worked example: a 1280 x 720, 30 fps depth stream crossing from the driver container to the host
Inputs, and where they come from: depth images are mono16, 16 bits per pixel, the encoding behind 16-bit depth and 8-bit infrared streams on one depth camera. Take a working format of 1280 × 720 px at 30 fps; both are settings you choose in the driver, so substitute your own and the arithmetic keeps its shape.
The rule in words: bytes per second equals width times height times bytes per pixel times frame rate, and a link budget in bits per second is that product times eight.
The arithmetic: 1280 × 720 = 921,600 px per frame. At 2 B per pixel that is 1,843,200 B per frame, and at 30 fps it is 1,843,200 × 30 = 55,296,000 B/s, about 55.3 MB/s. Times eight: 55,296,000 × 8 = 442,368,000 bit/s, about 442 Mbit/s.
Result: the depth topic alone asks for 442 Mbit/s, which is 44% of a 1000 Mbit/s Gigabit link, so it crosses between containers on one host with over half the line rate free. Add an 8-bit infrared stream in the same format and it contributes 921,600 × 1 × 30 × 8 = 221 Mbit/s, for 664 Mbit/s combined, still about two thirds of the line. Add RGB at 8 bits per channel and the third stream alone is 921,600 × 3 × 30 × 8 = 664 Mbit/s; the three together reach about 1,327 Mbit/s, past the line rate.
Decision: depth-only, or depth plus IR, crosses the container boundary comfortably at this format. Three full-rate streams do not fit one Gigabit link: drop the frame rate to 15 fps and the combined load falls to about 664 Mbit/s, or split the streams across separate network interfaces. Keep combined streams under roughly 90% of the line rate and this arithmetic stays your go/no-go.
Fix: three topologies that keep one owner and move images over ROS
Everything in one container is the bench default. Driver, subscribers and rviz run side by side, ownership is trivially single, and data never leaves the container; your DISPLAY mount already handles the GUI. When you want to hand over the camera node rather than all of /dev, passing a USB camera into a container with a CDI spec is the tidy route.
Driver on the host, tools in containers keeps the owning process outside any namespace. With host networking on both sides every endpoint advertises a reachable address, and the images cross loopback, where 442 Mbit/s is easy work.
One driver container, consumers anywhere reachable is the production shape. One container owns the camera for its whole life and everything else subscribes over ROS; on a single host this is again host networking, and across hosts it needs routed endpoints between them.
If hz already reports a healthy rate but rviz draws nothing, stop reading camera logs. Rendering is the suspect, and the black point cloud in Docker fix covers that branch.
Prevention: checks before you restart the stack
- One driver instance per device, counted across the host and every container. Two starts that both "succeed" mean one publishes nothing.
- Mount the specific device instead of all of /dev; the blanket mount is how a second container quietly reaches a camera it must never open.
- Set ROS_MASTER_URI identically in every terminal and container, and confirm the address you typed is the one the master answers on.
- Prove hz on the aligned depth topic inside the driver container before adding subscribers, so each change starts from a known-good stream.
- Two cameras planned? Each needs its own owning process and its own stable device identity; the multiple depth cameras on one host fix covers enumeration order and USB bandwidth.
When the usual answer is wrong
- The test publisher fails even between two terminals on the plain host: the ROS network is the problem, not the camera. Work the path, master address, subnet and firewall, the way the depth camera latency and network path fix does for Ethernet cameras.
- hz reports a steady rate but frames arrive black: ownership and routing are solved; look at rendering and drivers next.
- The camera works on the host and disappears inside a VM after a reboot: the same one-owner rule at a different boundary; who owns the device on Windows and a Linux VM maps that case.
Taking the topology to a production cell
A laptop proves the topology; a cell needs a controller that runs the driver container and its subscribers around the clock, with enough LAN ports to feed the cameras. Our vision controllers category filters by LAN port count, PoE and memory.
Both of these cover the topology above:
- LAN ports
- 3-4 LAN ports
- USB 3 ports
- 2 USB 3 ports
- LAN ports
- 5+ LAN ports
- USB 3 ports
- 3-4 USB 3 ports
The MV-VC2140-128G40-1T covers a single-camera bench with 4 built-in light channels and 8 GB of memory. The MV-VC5525-256G70-NN adds PoE camera ports, 5+ LAN ports and 32 GB of memory for multi-camera cells. To have the whole chain chosen around your camera, start at the system configurator; more fixes in this series sit on the 3D camera topic page, and the 3D sensors finder splits laser profilers from stereo cameras when the camera itself is the open question.
If the rate still refuses to appear after the sequence above, send your docker ps output and the driver log through the contact page, and we will read the topology with you.
Questions
Why does rostopic list work while rostopic echo prints nothing?
rostopic list talks only to the master, which answers on port 11311 even across a container boundary. rostopic echo opens a second, direct connection to the publishing node. When that node advertises an address your container cannot route to, the list succeeds and the data never arrives. Check that the publisher's advertised address resolves from inside the subscriber container, or run both sides with host networking.
Does --privileged or a full /dev mount give my container its own camera?
No. Privileged mode and device mounts change permissions, so the container's processes may reach the device node, but it is still one physical camera on one kernel. The vendor driver's one-owning-process rule counts every process on the host, inside or outside containers. Two containers that each mount /dev/bus/usb share one camera between them, not two cameras.
Why does the topic appear in rostopic list if the camera never published?
Registration and data are separate steps. A node announces its topics to the master when it starts, often before the first frame exists. If the device open failed after startup, the names stay on the master while no bytes ever flow, which is why list looks healthy while hz starves. Read an empty echo as a failed open or an unreachable publisher, never as a missing topic.
What changes when I add a second depth camera?
Every camera needs its own owning process, so run one driver instance per device with a distinct, stable device identity each. Enumeration order is not stable across reboots, and two streams double the load: two 442 Mbit/s depth streams will not share one 100 Mbit/s link. The guide on multiple depth cameras on one host walks the enumeration and USB bandwidth checks.
Does the same fix apply to ROS 2?
The device side does: one owning process per camera still holds, containers included. The graph side differs because ROS 2 has no master; ros2 topic hz reports only what DDS discovery has matched, and a QoS mismatch can leave a topic silent while the camera publishes fine. Check device ownership first, then QoS, then container networking.

