Hi everyone,
We are seeing a reproducible failure with an OAK-D Pro PoE when it is powered by a PoE injector and connected directly to the host. The same camera, host, code and configuration works reliably through a PoE switch. We'd like to know whether this is a known issue and what you recommend.
Setup
Device: OAK-D Pro PoE (board R4M3E4), DeviceId 19443010A15D842F00
Bootloader: NETWORK 0.0.29 (also the version embedded in DepthAI)
Firmware (from crash dump): 40f5e0b83ee8148d5871e08a74521f05ee0760cd
Host: Jetson AGX Orin, L4T R39.2 (Ubuntu 24.04), Python 3.12
DepthAI: 3.10.0 (first seen on 3.9.0)
Pipeline: Camera node on CAM_A (IMX378), 1920x1080 @ 60 fps, ISP resize on device, either VideoEncoder H.265 or raw NV12 to the host. IR projectors not enabled.
Power/network:
Failing: PoE injector <model / 802.3af/at / gigabit> → camera, injector data port → Orin RJ45 directly (1 Gbps link).
Working: unmanaged PoE switch (100 Mbps link) with camera and Orin on the same switch.
What happens (injector)
Power on the injector, open the device, stream: works fine (e.g., H.265 at ~43 Mbps, LeonOS CPU ~60%).
Close the device to change the output resolution (normal close() / context-manager exit). The camera reboots via the watchdog, as expected.
Wait until the device is announced again in X_LINK_BOOTLOADER state (we poll getAllAvailableDevices() and require two consecutive sightings), then open it again.
The device opens, but no frame is ever produced (TX:0 in device stats). After ~15–30 s the firmware crashes and a crash dump is collected (errorId 9001).
After one or two more cycles of this, the camera drops the Ethernet link completely: no carrier on the host NIC, no ping, no discovery. Only a power cycle of the injector brings it back. The first session after the power cycle works again.
This happened on every reconnect we tried through the injector: different resolutions, encode on and off, with and without heavy link load.
Same procedure on the PoE switch: 5 out of 5 reconnects succeeded. Each time the device reappeared ~13–15 s after close and streamed normally (encode and raw).
Relevant lines from the device crash dump (failed session)
$$
[E] hai_cm_sensor_select_mode():164: No suitable sensor mode. Selecting default one - 0 for start 1920x1080 at 0x0 fps min 60.000000 max 60.000000
============================ Configuring IMX378 with mode 0
[5.766] [Camera(0)] [debug] Original intrinsics for 3840x2160: ...
Patched intrinsics for 4056x3040: ...
MIPI DPHY 4 error, stat 0x30
[5.893] [VideoEncoder(1)] [warning] VideoEncoder settings coherency check changed 1 additional parameter(s) besides level
... TX:0 (no data transmitted afterwards)
$$
The same 1920x1080@60 mode was selected fine in the previous session, just before the close. To us this looks like the IMX378/MIPI path does not come back cleanly after the watchdog reboot when the camera is powered by the injector. Also visible in the boot log: Failed to probe IR driver LM3644 on bus 1 (LM3644 detected on bus 2); we don't know if that is normal.
Two full crash dumps (.tar from CrashDump.toTar) are attached.
Possibly related host-side crash (DepthAI 3.9.0)
While the device was in this state with setMaxReconnectionAttempts(3, cb), the host process segfaulted in a DepthAI thread after RECONNECTING → RECONNECT_FAILED:
$$
dai:😃eviceGate::getState()
dai:😃eviceBase::closeImpl()
dai:😃eviceBase::close()
dai:😛ipelineImpl::stop()
$$
We have since set reconnection attempts to 0 and handle reconnection ourselves.
Already checked
Bootloader updated (0.0.29). Bootloader version is not the issue.
Injector and cables tested; the camera works on first power-up every time.
Not a routing issue: the failure is loss of the physical link, not discovery.
Similar older reports: depthai-core #415 (OAK-D-Pro-PoE on injector dying after ~6 connections, fixed in 2.17.1) and forum thread 4854 (OAK-D W PoE freezing on one PoE source but fine on a TP-Link switch). Neither mentions the MIPI/sensor-mode symptom.
Questions
Is there a known issue with the sensor/MIPI not re-initializing after a watchdog (soft) reboot depending on PoE power source?
Are there power requirements for the reboot sequence beyond 802.3af (e.g., inrush, hold-up time) that an injector may not meet?
Is there a way to force a full sensor power reset on reconnect without cutting PoE power?
Is the DeviceGate::getState segfault in closeImpl for RVC2 PoE known/fixed?
Thanks!