AI models

Through the NNComponent, DepthAI SDK abstracts:

1. AI model sourcing using blobconverter from Open Model Zoo (OMZ) and DepthAI Model Zoo (DMZ).

2. AI result decoding - currently SDK supports on-device decoding for YOLO and MobileNet based results using YoloDetectionNetwork and MobileNetDetectionNetwork nodes.

3. Decoding of the config.json which allows an easy deployment of custom AI models trained using our notebooks and converted using https://tools.luxonis.com.

4. Formatting of the AI model input frame - SDK uses BGR color order and Planar / CHW (Channel, Height, Width) layout conventions. If model accepts color images, it should accept 3 channels (B,G,R), and if it accepts grayscale images, it should accept 1 channel.

Netron app allows you to check model’s input layout

5. Integration with 3rd party tools/services (Roboflow).

SDK supported models

With NNComponent you can easily try out a variety of different pre-trained models by simply changing the model name:

  from depthai_sdk import OakCamera

  with OakCamera() as oak:
      color = oak.create_camera('color')
-     nn = oak.create_nn('mobilenet-ssd', color)
+     nn = oak.create_nn('vehicle-detection-0202', color)
      oak.visualize([nn], fps=True)
      oak.start(blocking=True)

Both of the models above are supported by this SDK, so they will be downloaded and deployed to the OAK device along with the pipeline.

The following table lists all the models supported by the SDK. The model name is the same as the name used in the NNComponent constructor.

Name	Model Source	FPS*
age-gender-recognition-retail-0013	OMZ	33
emotions-recognition-retail-0003	OMZ	33
face-detection-adas-0001	OMZ	18
face-detection-retail-0004	OMZ	33
facemesh_192x192	DMZ	32
facial_landmarks_68_160x160	32	DMZ
human-pose-estimation-0001	OMZ	8
mobilenet-ssd	OMZ	31
mobilenetv2_imagenet_embedder_224x224	DMZ	/
pedestrian-detection-adas-0002	OMZ	19
person-detection-0200	OMZ	14
person-detection-retail-0013	OMZ	15
person-reidentification-retail-0288	OMZ	33
person-vehicle-bike-detection-crossroad-1016	OMZ	12
sbd_mask_classification_224x224	DMZ	64+
vehicle-detection-0202	OMZ	14
vehicle-detection-adas-0002	OMZ	14
vehicle-license-plate-detection-barrier-0106	OMZ	29
yolo-v3-tf	OMZ	3.5
yolo-v3-tiny-tf	OMZ	33
yolov4_coco_608x608	DMZ	1.1
yolov4_tiny_coco_416x416	DMZ	32
yolov5n_coco_416x416	DMZ	32
yolov6n_coco_640x640	DMZ	26
yolov6nr3_coco_640x352	DMZ	32
yolov7tiny_coco_640x352	DMZ	23
yolov7tiny_coco_416x416	DMZ	29
yolov8n_coco_640x352	DMZ	22

* - FPS was measured using only color camera (1080P) and 1 NN using callbacks (without visualization)

Got questions?

Head over to Discussion Forum for technical support or any other questions you might have.