LXC on Proxmox, Core i3 N305, Edge TPU - but no object detection

[degudejung]

Hi,
I run Scrypted NVR with a subscription for 5 cameras on the following setup

• Intel Core i3 N305
• Coral Edge TPU (M.2 PCIe)
• the "official" LXC container installed via the script on Proxmox 8.2, 8 GB, unprivileged
• Intel UHD iGPU and Coral Edge TPU passed through / allowed for the LXC (see below)
• 4 cameras connected successfully via ONVIF, 1 via RTSP

Integration of the cameras worked flawlessly. Recordings in NVR are OK. When I open the plugins in the management console, both tensor-flow-lite and open-vino and manually paste snapshots from the cameras, it does detect cars, potted plants, person etc. However, whatever I chose for object detection, there are "no events in the past 24 hours".

Here's how devices look from inside the LXC:

ls -l /dev/apex*  
crw-rw-rw- 1 nobody nogroup 120, 0 Apr 28 18:49 /dev/apex_0  
---------- 1 root   root         0 Apr 29 15:37 /dev/apex_1
ls -l /dev/dri/ 
drwxr-xr-x 2 nobody nogroup       80 Apr 28 18:49 by-path
crw-rw-rw- 1 nobody nogroup 226,   0 Apr 28 18:49 card0
crw-rw-rw- 1 nobody nogroup 226, 128 Apr 28 18:49 renderD128

...so it looks to me as if the Coral TPU is correctly installed and drivers loaded on the host and that the iGPU is correctly made accessible to the LXC.

Open-vino shows in the console:

Available devices:
CPU :
        SUPPORTED_PROPERTIES:
                AVAILABLE_DEVICES: 
                RANGE_FOR_ASYNC_INFER_REQUESTS: 1, 1, 1
                RANGE_FOR_STREAMS: 1, 4
                EXECUTION_DEVICES: CPU
                FULL_DEVICE_NAME: Intel(R) Core(TM) i3-N305
                OPTIMIZATION_CAPABILITIES: FP32, FP16, INT8, BIN, EXPORT_IMPORT
                DEVICE_TYPE: Type.INTEGRATED
                DEVICE_ARCHITECTURE: intel64
                NUM_STREAMS: 1
                AFFINITY: Affinity.CORE
                INFERENCE_NUM_THREADS: 0
                PERF_COUNT: False
                INFERENCE_PRECISION_HINT: <Type: 'float32'>
                PERFORMANCE_HINT: PerformanceMode.LATENCY
                EXECUTION_MODE_HINT: ExecutionMode.PERFORMANCE
                PERFORMANCE_HINT_NUM_REQUESTS: 0
                ENABLE_CPU_PINNING: True
                SCHEDULING_CORE_TYPE: SchedulingCoreType.ANY_CORE
                ENABLE_HYPER_THREADING: True
                DEVICE_ID: 
                CPU_DENORMALS_OPTIMIZATION: False
                LOG_LEVEL: Level.NO
                CPU_SPARSE_WEIGHTS_DECOMPRESSION_RATE: 1.0
                DYNAMIC_QUANTIZATION_GROUP_SIZE: 0
                KV_CACHE_PRECISION: <Type: 'float16'>

available devices: ['CPU']
model/mode/precision: scrypted_yolov8n_320/AUTO/FP32
/root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/scrypted_yolov8n_320/FP32/best.xml /root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/scrypted_yolov8n_320/FP32/best.bin /root/.scrypted/volume/plugins/@scrypted/openvino/files/scrypted_labels.txt
EXECUTION_DEVICES ['CPU']
/root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/scrypted_yolov9c_flt/FP32/best.xml /root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/scrypted_yolov9c_flt/FP32/best.bin
/root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/vgg_english_g2/FP32/best.xml /root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/vgg_english_g2/FP32/best.bin
/root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/inception_resnet_v1/FP32/best.xml /root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/inception_resnet_v1/FP32/best.bin
/root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/craft/FP32/best.xml /root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/craft/FP32/best.bin
/root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/vgg_english_g2/FP32/best.xml /root/.scrypted/volume/plugins/@scrypted/openvino/files/v5/vgg_english_g2/FP32/best.bin

...and under Settings-> Available Devices there is also only"CPU". Should there not also be the iGPU?

In Tensorflow-lite console it shows:

coral edge tpu library loaded successfully
edge tpus [{'type': 'pci', 'path': '/dev/apex_0'}]
model: efficientdet_lite0_320_ptq
added tpu {'type': 'pci', 'path': '/dev/apex_0'}
/root/.scrypted/volume/plugins/@scrypted/tensorflow-lite/files/v13/efficientdet_lite0_320_ptq_edgetpu.tflite /root/.scrypted/volume/plugins/@scrypted/tensorflow-lite/files/v13/coco_labels.txt

..and under Settings it says "Detected Edge TPU [{"type": "pci", "path": "/dev/apex_0"}]". Model is set to "default"

I am not a Linux pro but so far all this looks pretty OK to me.

BUT :-(

Whatever I pick on the cameras "Object Detection"->"Advanced" ribbon, be it default, tensorflow-lite or open-vino, I NEVER get any events in the camera stream inside the NVR. I can easily browse through the last hours and see objects appear in great detail but there is no single detection. I even brought the threshold down to 0.3 and tried all three "Decoder" options default, WebAssembly and FFmpeg. I have the same (lack of) effect on the ONVIF and RTSP camera(s).

Please support regarding
1 Should I opt for the Coral TPU or the Intel iGPU (Scrypted NVR is the only reason I bought the Coral TPU)?
2 What am I missing to get things going with object detection? How to start troubleshooting?

Thank you in advance

[koush]

motion sensor issue probably https://docs.scrypted.app/scrypted-nvr/troubleshooting.html#object-detection

[degudejung]

Thank you @koush. That was it indeed 🤦‍♂️
I had not read the pipeline that way and thought I could do object detection directly without motion detection as the trigger. Now I just have to fine-tune the MD sensitivity to reduce all the false alarms and it should be good.

Still, I don't understand why the Intel iGPU - the one that you recommend so much - will not show up as hardware in any plugin, especially open-vino. Is there anything I can do to troubleshoot this?