[Support]: N150/OpenVINO/mobilenet_v2 upgrade questions.

[vespaman]

So, I am on 0.15, with 5 cameras and N150/OpenVINO and mobilenet v2. This has been working sufficiently for a year or two. Now, I'm moving all my different servers to proxmox, and while doing it, it would be nice to improve the current setup.

Mainly, I have false detection every night on one camera pointing to trees, backlit by streetlight, and missing detections on others. I only want people detection (nothing fancier).

If I have understood correctly, going for a YOLO setup might be better, and in my proxmox cluster I have a few N150's and one AMD 7H255 at hand. I can also insert a Hailo 8L into one of the N150.
Can you suggest which route (detector/on which hardware).

With the current setup (300x300) openvino mobilenet v2, I have about 20-25% CPU and inference of about 13ms. I guess I could in worst case split cameras between multiple containers, but ideally I'd like them all on one.

[inkeep-ai-assistant[bot]]

Based on your setup with 5 cameras and the hardware options available, here are the detector and hardware recommendations:

Recommended Setup

Best Option: AMD 7H255 with OpenVINO and YOLOv9

The AMD 7H255 (780M integrated GPU) with OpenVINO detector running a YOLOv9 model would be your best choice.(1) The 780M shows an inference time of approximately 14ms for YOLOv9 at 320x320 resolution, which is comparable to your current 13ms with mobilenet_v2.(1) This setup supports YOLOv9 models which offer improved accuracy over the older mobilenet_v2 architecture.(2)

For this configuration:

detectors:
  ov:
    type: openvino
    device: GPU

model:
  model_type: yolo-generic
  width: 320
  height: 320
  input_tensor: nchw
  input_dtype: float
  path: /config/model_cache/yolo.onnx
  labelmap_path: /labelmap/coco-80.txt

(3)

Alternative: Hailo 8L in N150

If you add a Hailo 8L to one of your N150 systems, you could use YOLOv6n models with inference times around 11ms.(1) The Hailo detector automatically selects the default YOLOv6n model for Hailo-8L hardware.(3)

Configuration example:

detectors:
  hailo:
    type: hailo8l
    device: PCIe

model:
  width: 320
  height: 320
  input_tensor: nhwc
  input_pixel_format: rgb
  input_dtype: int
  model_type: yolo-generic
  labelmap_path: /labelmap/coco-80.txt

(3)

Why Upgrade from mobilenet_v2

YOLOv9 and YOLOv6n models are newer architectures that offer "slightly higher accuracy and improved detection of small objects" compared to mobilenet-based models.(2) This should help with your missing detections issue.

Addressing False Positives

For the tree/backlight false positives, upgrading to a YOLO-based model may help, but you should also consider using Frigate+ models if you want further improvements. Frigate+ models are "specifically designed for the way Frigate NVR analyzes video footage" and "offer higher accuracy with less resources."(2) You can submit both true and false positives to train a model optimized for your specific conditions.(2)

Single Container Feasibility

With the AMD 7H255's integrated GPU, you should be able to run all 5 cameras in one container. The documentation notes that "when using many cameras one detector may not be enough to keep up" and suggests defining multiple detectors if needed.(3) However, given your current 20-25% CPU usage and the similar inference times, a single detector on the 7H255 should handle your workload.

📚 Sources:

1. Recommended hardware
2. Models
3. Supported Hardware

---

Have another question? Just tag @inkeep.

[NickM-27]

The AI is not quite correct here. You can't use the GPU with openvino. You could start by trying yolov9 on the N150 and see how it performs

[vespaman]

OK, thank you for correcting him, :) I'll try yolov9.