A stereo depth camera that returns clean maps all morning and frames of speckle by mid-afternoon looks like a failing sensor, and it almost never is. The signature case runs fine for roughly 25-30 capture sessions of about 15 s each, then degrades, and a USB replug clears it in seconds. That is a saturated USB 3 link, not a worn camera. The host silently drops packets, incomplete frames pile up, and the stereo engine is left matching half-empty images. What follows is the bandwidth arithmetic, a five-minute viewer test that convicts or acquits the link in one session, the settings that cut the stream to a comfortable rate, and the cabling and code changes that keep it fixed on a production line.
Why does stereo depth turn noisy only after long runs?
The failure pattern is a link-budget signature, not a sensor fault. Clean depth for hours, then frames of speckle after roughly 25-30 capture sessions of about 15 s each, cured instantly by a replug: the USB 3 host is discarding packets faster than your pipeline notices.
The noise itself is diagnostic. It arrives as whole bad frames — full-image speckle, holes, random disparity — alternating with good ones, not as soft edge fuzz. Unplugging the camera restores a clean stream immediately, and after the first episode the fault returns sooner each time. Both details point at the transport. Silicon that wears out does not recover on replug, and a pipeline that has already choked refills fast under the same load. A flat on-device temperature trace across a failing session is the third clue, and it lets you rule the sensor's thermal state out before you start.
How much USB 3 bandwidth does 1080p RGB plus 720p depth at 30 fps use?
Uncompressed, the pair of streams moves about 242 MB/s — call it 1.94 Gbit/s. That is half or more of the roughly 3.2-3.8 Gbit/s a USB 3 port delivers in practice, and every other device on that controller competes for what is left.
The arithmetic assumes uncompressed RGB and raw 16-bit depth, the usual case for stereo depth output; compressed RGB would shrink the first term only.
USB 3.0 advertises 5 Gbit/s, but its 8b/10b coding leaves 4 Gbit/s of payload, and framing and protocol overhead bring a sustained stream's practical ceiling to roughly 3.2-3.8 Gbit/s. Your two streams take 51-61% of that before headers, timestamps or a second device, and a front-panel case header or a hub shaves the ceiling further. The margin is thin enough that one burst — an OS copy job, another camera on the same controller — tips the link into dropped packets, and each dropped packet becomes an incomplete frame. The same budget arithmetic, generalized to any interface, is in the camera bandwidth budget guide.
What fails first: the link, the cable, the power or the heat?
Rank the causes in this order — link saturation, cable, power, heat. Saturation shows up as an incomplete-frame counter that climbs, cable faults survive a settings change, power faults track the projector switching on, and heat shows a rising temperature trace.
- Link saturation (most likely). Depth speckles in whole frames and the viewer logs incomplete frames while your application runs. Test it in five minutes without your code, as in the next section.
- Cable and connector. A kinked, stretched or repeatedly flexed cable, or a run through a front-panel case header, drops packets even at modest rates. Test: swap in a short certified cable straight into a rear motherboard port and re-run the counter.
- Power budget. A bus-powered hub or a weak port browns the camera out just as the projector switches on, which reads as noise or disconnects at stream start. Test: move to a powered hub or a rear port. The budget's extreme failure — no image at all — is covered in depth camera black image and power checks, and the projector-start signature in depth camera disconnects when the IR stream starts.
- Heat (least likely here). Read the on-device temperatures during a failing session. A flat trace — a projector holding near 30 °C and an ASIC under about 40 °C — clears the sensor; a climbing trace means mounting and enclosure work, not a cable swap.
How do you confirm the cause in five minutes, without your code?
Close your application, open the vendor's viewer, stream the exact same resolutions and frame rate, and watch the log. If "incomplete video frame" warnings accumulate and the depth map speckles with your code out of the loop, the host link is dropping packets and your software is exonerated.
- Close your application completely, so it holds no buffers on the device.
- Open the vendor's depth viewer and set the production stream set: 1920 × 1080 at 30 fps RGB plus 1280 × 720 at 30 fps depth.
- Stream for the full failing session: 10-15 minutes continuous covers the 25-30 runs of 15 s that showed the fault.
- Watch two things: the incomplete-frame warnings ("incomplete video frame" in most viewers) and the IR image, where a working structured-light projector shows a dense white dot pattern.
- Halve exactly one variable — depth to 848 × 480 at 30 fps, or the whole set to 15 fps — and confirm the warnings stop.
- If warnings persist at the reduced rate, go back to cause 2: swap cable and port before touching settings again.
Commit to this reading: at these settings a healthy link logs zero incomplete frames in a 15-minute run. A few per minute is a saturated link, and the fix is bandwidth, not code. Leave the viewer at the application's exact settings; a viewer that quietly negotiates a lower mode will hide the fault you are chasing.
Why does replugging fix it, and why does it come back?
Unplugging forces the host to re-enumerate the device. Stalled endpoints are cleared, buffers are flushed and the pipeline restarts from a clean state, which is why the cure is instant — and why the fault returns, because the saturated link that caused it is unchanged.
The camera's stereo silicon never degraded, which is why recovery is complete rather than partial. Recurrence comes sooner after the first episode because the shortfall is still there, and the host stack re-enters a state it has already failed in: half-flushed queues retrying into a saturated wire. The production-safe equivalent of the replug is an in-software device reset, where your SDK exposes one. Trigger it from the error handler when the incomplete-frame count spikes, and treat it as a recovery path, not the fix. The shorter-cycle version of the same stale-session behavior, and its reset, is in frames that never arrive within 5 seconds.
Which settings cut the bandwidth enough to stop the drops?
Cut the depth stream first: 848 × 480 at 30 fps needs about 24 MB/s instead of 55 MB/s, and 1280 × 720 at 15 fps needs about 28 MB/s. Either one turns a saturated link into one with more than half its capacity spare.
| Stream set | Raw rate | Share of 3.2-3.8 Gbit/s | Verdict |
|---|---|---|---|
| 1080p RGB + 720p depth, 30 fps | 242 MB/s (1.94 Gbit/s) | 51-61% before overhead | Marginal — drops under load |
| 1080p RGB + 720p depth, 15 fps | 121 MB/s (0.97 Gbit/s) | 26-30% | Fits, halves update rate |
| 720p RGB + 848 × 480 depth, 30 fps | 107 MB/s (0.86 Gbit/s) | 23-27% | Comfortable |
| 720p RGB only, 30 fps | 83 MB/s (0.66 Gbit/s) | 17-21% | Trivial |
Decide by what the inspection needs. Keep 1080p RGB only when you align depth to color for measurement; otherwise 720p RGB plus 848 × 480 depth is the set to order around. The 848 × 480 mode coarsens the depth image about 1.5× per axis: a feature spanning 30 depth pixels at 720p spans about 20 at 480p, so check your smallest feature against that before committing. Aligning depth to color happens on the host and cuts no wire traffic — the fix has to be at the source. Disable any auxiliary stream you do not consume; every enabled stream shares the one cable.
Do you still need code changes, or is new hardware enough?
Do both. Requeue every incomplete frame instead of filtering it, and move the camera to a rear motherboard port on a short certified cable. The code fix stops poisoned frames from reaching your filters; the hardware fix stops the drops from happening.
In code, treat an incomplete frame as poison. After each grab, read the frame's completeness flag as your SDK reports it; when it is incomplete, requeue the buffer and grab a fresh one. Never run a partial frame through alignment or the spatial and hole filters — that pipeline is exactly what turns a few dropped packets into the full-frame speckle you saw. Count every discard and log it per session as a health metric; the incomplete frame discard-and-requeue pattern shows the buffer handling.
On the hardware side: rear motherboard port, no hub, no passive extension, no front-panel case header, and a certified cable. Treat 3 m as the practical ceiling for passive USB 3 copper, and go active beyond it. Where the camera rides a moving gantry and the cable flexes every cycle, specify a drag-chain-rated USB3.0 camera cable; the 3 m high-flex LS-CBL-0219 is built for that duty:
When is the noise not bandwidth?
If the viewer shows zero incomplete frames while the depth still speckles, the link is clean and the noise is optical. Check that the IR image shows the projector's dot pattern, that the part holds still for the full exposure, and that the exposure fits the frame period.
Work that checklist in order. A dark IR image means the projector is off or failed, and active stereo has no texture to match; a sparse pattern explains noisy edges on dark or shiny parts. Exposure is a hard ceiling: at 30 fps the frame period is 33.3 ms, so an exposure set past it caps the effective rate and adds blur — the depth camera frame rate and exposure cap article works that limit, and the 2D stream exposure and gain defaults give starting values. Camera or part motion inside the exposure corrupts disparity regardless of link health; the mounting stillness guide covers the fixture. When counters stay at zero, the dot pattern is dense and speckle persists on edges, the residual noise is stereo matching, which has its own stereo disparity noise troubleshooting guide.
What do you change so it never comes back?
Prevention is a short list. Keep the camera on a rear motherboard port, keep the cable certified, short and unmoving, log the incomplete-frame count every session, and re-specify the sensor only when the viewer test proves the platform itself is the problem.
The decision per case: counter climbs with your code closed — cut depth to 848 × 480 at 30 fps or the whole set to 15 fps, and drop RGB to 720p when alignment allows. Counter still climbs at the reduced rate — new certified cable into a rear port, hub removed. Disconnects when the projector starts — powered hub or a different port. Counter silent but the map speckles — optics and exposure, not the link.
Commissioning checklist:
- Stream the production settings in the viewer before the first shift and record the incomplete-frame count as the baseline; expect zero.
- Route the cable with a service loop, away from power lines, clamped so it cannot flex at the connector.
- Put the incomplete-frame counter on your dashboard with an alert above one per minute sustained.
- Inspect the cable jacket and connectors every maintenance interval; a greyed, kinked jacket is a scheduled replacement, not a wait-for-failure part.
- After any PC or hub change, re-run the five-minute viewer test — the budget changed even though nothing on the bench did.
If the task has outgrown the camera rather than the link — range, resolution, field of view — a guide to selecting an active stereo depth camera walks the platform trade-offs, and the re-specification starts from these stereo and structured-light depth sensors:
- Type
- Stereo / structured light
- Type
- Stereo / structured light
For parts and help sizing the stream set around your part, browse stereo and structured-light 3D sensors or the camera finder, match a certified run from camera data and power cables, or start at the vision configurator. The 3D topic page collects the neighboring fixes, and the team answers spec questions through the contact page.
Questions
Can the camera run on USB 2.0 instead?
USB 2.0 offers about 35 MB/s of practical bulk throughput. Depth-only at 1280 × 720, 16-bit, 15 fps needs about 27.6 MB/s, so depth alone fits with a little margin; adding 1080p RGB does not. Most stereo depth cameras fall back to a reduced mode on a USB 2.0 port, so plan on lower frame rate and resolution, not the full stream set.
Is a hub or an extension cable ever acceptable?
Yes, with conditions. Use a powered hub so each port has its own power budget. Plug the camera into a hub port whose controller is not shared with another camera, and re-run the incomplete-frame count after installing it. Passive extension cables are the usual hidden fault: replace one with a single certified cable of the full length, or with an active cable.
Does this fault wear the camera out?
No. Dropped packets cost the silicon nothing; the sensor, projector and processor are not degraded by an overloaded link. The wear items are mechanical: cable jackets flexed at the connector and USB connectors cycled by repeated replugging. Fix the link budget and the replug ritual disappears on its own, which also extends connector life.
What incomplete-frame rate should trigger an alert?
At these settings a healthy link logs zero incomplete frames in a 15-minute session. Alert above one per minute sustained, or any spike past ten in an hour. Treat the first alert as the cue to re-run the five-minute viewer test before the depth maps go visibly bad.
Does switching the projector off reduce the load?
The projector costs almost no USB bandwidth, so switching it off does not fix a saturated link. On untextured parts it makes depth noisier, because the stereo engine loses the texture it matches on. It does trim the camera's power draw slightly, which only matters on a marginal port. Leave it on; recover bandwidth by lowering depth resolution or frame rate instead.


