eWaSR | Luxonis

Our new model ZOO works with DepthAI V3. Find out more in our documentation.

0 Likes

Model Details

Model Description

eWaSR is an embedded-compute-ready maritime obstacle detection network. The model can segment 3 classes: sky, water, and obstacles. It can be used in applications where the safe navigation of autonomous surface vehicles is required.

Developed by: Matija Teršek et al.
Shared by:
Model type: Computer vision
License:
Resources for more information:

Training Details

Training Data

The model was trained on , created for training deep USV obstacle detection models. For more information about training data check the .

Testing Details

Metrics

Images from the are categorized into two categories: Overall and Danger Zone. The latter represents images where obstacles are less than 15m away which pose an immediate threat to the vehicle.

Metric	Overall	Danger Zone (< 15m)
F1 score	92.56	78.09

Results are taken from .

Technical Specifications

Input/Output Details

Input:
- Name: image
  - Info: NCHW BGR 0-255 image.
Output:
- Name: prediction
  - Info: Segmentation masks for every class.

Model Architecture

eWaSR architecture follows the encoder, feature mixer, and decoder architecture. The backbone features of intermediate encoder layers are resized, concatenated and processed by a lightweight scale-aware semantic extraction module (LSSE). The semantically-enriched features are injected into higher-layer backbone features by semantic-injection modules - SIM. The resulting features are then concatenated with the IMU mask and passed to the segmentation head.

Please consult the for more information on model architecture.

Throughput

Model variant: ewasr:512x384

• Input shape: [1, 3, 384, 512] • Output shapes: [[1, 3, 96, 128], [1, 256, 24, 32]]

Platform	Precision	Throughput (infs/sec)	Power Consumption (W)
RVC2	FP16	7.08	N/A
RVC4	FP32	312.28	3.84

* Benchmarked with , using 2 threads (and the DSP runtime in balanced mode for RVC4).

* Parameters and FLOPs are obtained from the package.

Quantization

RVC4 version of the model was quantized using a custom dataset. This was created by taking 40-image subset of .

Utilization

Models converted for RVC Platforms can be used for inference on OAK devices. DepthAI pipelines are used to define the information flow linking the device, inference model, and the output parser (as defined in model head(s)). Below, we present the most crucial utilization steps for the particular model. Please consult the docs for more information.

Install DAIv3 and depthai-nodes libraries:

pip install depthai
pip install depthai-nodes

Define model:

model_description = dai.NNModelDescription(
    "luxonis/ewasr:512x384"
)

nn = pipeline.create(ParsingNeuralNetwork).build(
    <CameraNode>, model_description
)

Inspect model head(s):

SegmentationParser that outputs message (segmentation mask for 19 classes).

Get parsed output(s):

while pipeline.isRuning():
    parser_output: dai.ImgFrame = parser_output_queue.get()

Example

You can quickly run the model using our script. It automatically downloads the model, creates a DepthAI pipeline, runs the inference, and displays the results using our DepthAI visualizer tool. To try it out, run:

python3 main.py \
    --model luxonis/ewasr:512x384 \
    -overlay

Maritime obstacle detection model.
License	Apache 2.0 Commercial use
Downloads	148
Tasks	Semantic Segmentation
Model Types	ONNX

Model Variants

Name	Version	Available For	Created At	Deploy
		RVC2, RVC4	10 months ago