License Plate Detection - NVIDIA TAO MBv2 SSD - Edge Impulse

Alif Semiconductor / License Plate Detection Public

Target: Alif HP (Cortex-M55 400MHz + U55-256MACC) Clone this project

NVIDIA TAO RetinaNet settings

Training settings

Training processor

Please provide a valid training processor option

Backbone

Number of training cycles

Number of training cycles should be a number

Minimum learning rate

Minimum learning rate should be a number

Maximum learning rate

Maximum learning rate should be a number

Augmentation settings

Random crop min scale

Random crop min scale should be a number

Random crop max scale

Random crop max scale should be a number

Random crop min aspect ratio

Random crop min aspect ratio should be a number

Random crop max aspect ratio

Random crop max aspect ratio should be a number

Zoom out min scale

Zoom out min scale should be a number

Zoom out max scale

Zoom out max scale should be a number

Brightness

Brightness should be a number

Contrast

Contrast should be a number

Saturation

Saturation should be a number

Hue

Hue should be a number

Random flip

Random flip should be a number

Advanced training settings

Validation set size

Please provide a valid number for the train/validate split (between 0 and 1)

%

Split train/validation set on metadata key

Profile int8 model

Soft start

Soft start should be a number

Annealing

Annealing should be a number

Quantization aware training

IoU threshold

IoU threshold should be a number

Confidence threshold

Confidence threshold should be a number

Batch size

Batch size should be a number

Regularizer type

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Freeze blocks

Freeze blocks should not be empty

Neural network architecture

Save

Input layer (76,800 features)

Edge Impulse Inc. / NVIDIA TAO SSD

Output layer (1 classes)

Model

Model version:

Last training performance (validation set)

Precision score

49.1%

Metrics (validation set)

Metric	Value
mAP	0.14
mAP@[IoU=50]	0.37
mAP@[IoU=75]	0.05
mAP@[area=small]	-1.00
mAP@[area=medium]	0.00
mAP@[area=large]	0.23
Recall@[max_detections=1]	0.26
Recall@[max_detections=10]	0.26
Recall@[max_detections=100]	0.26
Recall@[area=small]	-1.00
Recall@[area=medium]	0.01
Recall@[area=large]	0.43

On-device performance

Engine:

Inferencing time

0 ms.

Peak RAM usage

0.0K

Flash usage

0.0K

Did you know? You can customize your model through the Expert view (click on to switch),
or can even bring your own model (in PyTorch, Keras or scikit-learn).

Model	Author
MobileNetV2 SSD FPN-Lite (320x320 only) Officially supported A pre-trained object detection model designed to locate up to 10 objects within an image, outputting a bounding box for each object detected. The model is around 3.7MB in size. It supports an RGB input at 320x320px. For other resolutions, use FOMO or an NVIDIA TAO model.	Edge Impulse
FOMO (Faster Objects, More Objects) MobileNetV2 0.1 Officially supported An object detection model based on MobileNetV2 (alpha 0.1) designed to coarsely segment an image into a grid of background vs objects of interest. These models are designed to be <100KB in size and support a grayscale or RGB input at any resolution.	Edge Impulse
FOMO (Faster Objects, More Objects) MobileNetV2 0.35 Officially supported An object detection model based on MobileNetV2 (alpha 0.35) designed to coarsely segment an image into a grid of background vs objects of interest. These models are designed to be <100KB in size and support a grayscale or RGB input at any resolution.	Edge Impulse
yolon-quant Enterprise yolo5vn with quantization	Alif Semiconductor
YOLOv5 for Renesas DRP-AI Community Transfer learning model using YOLOv5 v5 branch with yolov5s.pt weights. This block is only compatible with Renesas DRP-AI.	Renesas
YOLOv5 Community Transfer learning model based on Ultralytics YOLOv5 using yolov5n.pt weights, supports RGB input at any resolution (square images only).	Community blocks
YOLOX for TI TDA4VM Community TI's EDGEAI YOLOX. https://github.com/TexasInstruments/edgeai-yolox. Outputs ONNX v7 model format both with and without final detect layers using PyTorch 1.7.1. See the implementation https://github.com/edgeimpulse/example-custom-ml-block-ti-yolox/tree/onnx-v7	Texas Instruments
NVIDIA TAO RetinaNet Professional Enterprise Object detection model with superior performance on smaller objects. Configurable backbones optimized for targets from MCU to GPU. Supports rectangular input. Image width and height must be multiples of 32. Training requires GPU.	Edge Impulse Inc.
NVIDIA TAO YOLOV3 Professional Enterprise Object detection model that is fast and accurate. Configurable backbones optimized for targets from MCU to GPU. Supports rectangular input. Image width and height must be multiples of 32. Training requires GPU.	Edge Impulse Inc.
NVIDIA TAO YOLOV4 Professional Enterprise Object detection model that is fast and accurate. Configurable backbones optimized for targets from MCU to GPU. Supports rectangular input. Image width and height must be multiples of 32. Training requires GPU.	Edge Impulse Inc.
NVIDIA TAO SSD Professional Enterprise Object detection model for general purpose use. Configurable backbones optimized for targets from MCU to GPU. Supports rectangular input. Image width and height must be multiples of 32. Training requires GPU.	Edge Impulse Inc.

The backbone is the foundational element of a neural network model, responsible for extracting meaningful features from input data, enabling subsequent parts to perform anomaly scoring or classification.

Description	Author
EfficientNet V2B0 Uses around 755K RAM based on your input size, and between 240-1675K ROM depending on the number of layers, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 0.35 Uses around 636K RAM based on your input size, and between 69-134K ROM depending on the number of layers, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 0.50 Uses around 643K RAM based on your input size, and between 78-168K ROM depending on the number of layers, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 0.75 Uses around 1263K RAM based on your input size, and between 109-266K ROM depending on the number of layers, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 1.0 Uses around 1273K RAM based on your input size, and between 126-372K ROM depending on the number of layers, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 0.1 Uses around 622K RAM based on your input size, and 58K ROM, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse

The scoring function operates on the features extracted by the backbone, and returns anomaly scores for each segment of the image.

Description	Author	Recommended
MobileNetV2 SSD FPN-Lite (320x320 only) A pre-trained object detection model designed to locate up to 10 objects within an image, outputting a bounding box for each object detected. The model is around 3.7MB in size. It supports an RGB input at 320x320px. For other resolutions, use FOMO or an NVIDIA TAO model.	Edge Impulse
FOMO (Faster Objects, More Objects) MobileNetV2 0.1 An object detection model based on MobileNetV2 (alpha 0.1) designed to coarsely segment an image into a grid of background vs objects of interest. These models are designed to be <100KB in size and support a grayscale or RGB input at any resolution.	Edge Impulse
FOMO (Faster Objects, More Objects) MobileNetV2 0.35 An object detection model based on MobileNetV2 (alpha 0.35) designed to coarsely segment an image into a grid of background vs objects of interest. These models are designed to be <100KB in size and support a grayscale or RGB input at any resolution.	Edge Impulse
yolon-quant yolo5vn with quantization	Edge Impulse
YOLOv5 for Renesas DRP-AI Transfer learning model using YOLOv5 v5 branch with yolov5s.pt weights. This block is only compatible with Renesas DRP-AI.	Edge Impulse
YOLOv5 Transfer learning model based on Ultralytics YOLOv5 using yolov5n.pt weights, supports RGB input at any resolution (square images only).	Edge Impulse
YOLOX for TI TDA4VM TI's EDGEAI YOLOX. https://github.com/TexasInstruments/edgeai-yolox. Outputs ONNX v7 model format both with and without final detect layers using PyTorch 1.7.1. See the implementation https://github.com/edgeimpulse/example-custom-ml-block-ti-yolox/tree/onnx-v7	Edge Impulse
NVIDIA TAO RetinaNet Object detection model with superior performance on smaller objects. Configurable backbones optimized for targets from MCU to GPU. Supports rectangular input. Image width and height must be multiples of 32. Training requires GPU.	Edge Impulse
NVIDIA TAO YOLOV3 Object detection model that is fast and accurate. Configurable backbones optimized for targets from MCU to GPU. Supports rectangular input. Image width and height must be multiples of 32. Training requires GPU.	Edge Impulse
NVIDIA TAO YOLOV4 Object detection model that is fast and accurate. Configurable backbones optimized for targets from MCU to GPU. Supports rectangular input. Image width and height must be multiples of 32. Training requires GPU.	Edge Impulse
NVIDIA TAO SSD Object detection model for general purpose use. Configurable backbones optimized for targets from MCU to GPU. Supports rectangular input. Image width and height must be multiples of 32. Training requires GPU.	Edge Impulse

These users will be notified over email when this job finishes.

	Name
	Joshua Buck Edge Impulse staff