Which model to use and where to download dataset

[RGVGreatCoder]

Hello,

I want to test the person-vehicle-bike-detection-crossroad on both: my Raspberry Pi 4 (4Gb) with an Intel Neural Compute Stick 2 (NCS2) and on my Ubuntu 20.x in my WSL (I successfully setup the OpenVINO toolkit on both devices).  However, I see several types of this Pre-Trained Models which are person-vehicle-bike-detection-crossroad-2000, ...2001, 2002, 2003, 2004, 0078, 1016 and 1020. I like the person-vehicle-bike-detection-crossroad-1016 (https://docs.openvino.ai/latest/omz_models_model_person_vehicle_bike_detection_crossroad_1016.html) only because I can better understand the Specification values than the other ones but, I am not sure this is the best one to use. I am also struggling on finding the correct dataset and images for -any of Intel's Pre-trained models. I am following this page (https://docs.openvino.ai/latest/workbench_docs_Workbench_DG_Generate_Datasets.html#doxid-workbench-docs-workbench-d-g-generate-datasets) but it is not clear to me where to download the respective dataset. However, I downloaded the "2017 Val images and 2017 Train/Val annotations" files from this OpenVINO page (https://docs.openvino.ai/latest/workbench_docs_Workbench_DG_Dataset_Types.html#coco) but I can't tell if these files are the correct ones for the person-vehicle-bike-detection-crossroad-1016 pre-trained model.

So I have the following two Questions:

1. Which pre-trained model should I use in my Pi4 with an NCS2 (and why)?
2. Where can I find the right dataset (annotations and images) for every dataset (particularly, the recommended intel's pre-trained model for my Pi4)?

Thank you and will be waiting for your reply,

[Wan_Intel]

Hi RGVGreatCoder,

Thanks for reaching out to us.

To answer your first question, the Inference Engine MYRIAD plugin has been developed for inference of neural networks on Intel® Neural Compute Stick 2 (Intel® NCS2).

First, to configure your Intel® NCS2, please execute the following command:

sudo usermod -a -G users "$(whoami)"

source /opt/intel/openvino_2021/bin/setupvars.sh

sh /opt/intel/openvino_2021/install_dependencies/install_NCS_udev_rules.sh

For more information, please refer to Add USB Rules for an Intel® Neural Compute Stick 2 device.

Next, OpenVINO™ Toolkit provides a set of Intel’s Pre-Trained Models and Public Pre-Trained Models that you can use for learning and demo purposes or for developing deep learning software.

Please refer to Intel’s Pre-Trained Models Device Support and Public Pre-Trained Models Device Support and choose any model that is supported by MYRIAD plugin to use with Intel® NCS2.

For example, person-vehicle-bike-detection-crossroad-1016 is supported by MYRIAD plugin and it can be used with Intel® NCS2. However, resnet18-xnor-binary-onnx-0001 is not supported by MYRIAD plugin, hence it cannot be used with Intel® NCS2.

For your second question, you don’t have to get the dataset (annotations and images) to run inference in your Raspberry Pi with Intel® NCS2 or WSL2.

If you would like to obtain sample media files that are available for the demo application, you can use the images and videos from https://github.com/intel-iot-devkit/sample-videos.

Please get back to us shall you face any issues.

Regards,

Wan

[RGVGreatCoder]

Thank you for your reply.

Can you please provide me a link with a Python sample code where I can infer the person-vehicle-bike-detection-crossroad-1016 model to detect an object in a picture and in a video?

[Wan_Intel]

Hi RGVGreatCoder,

You can use Single Human Pose Estimation Demo, a Python demo from Open Model Zoo that showcases top-down pipeline for human pose estimation on video or image.

You can run the following command to do inference on a CPU:

python3 single_human_pose_estimation_demo.py -d CPU --model_od <path_to_model>/mobilenet-ssd.xml --model_hpe <path_to_model>/ person-vehicle-bike-detection-crossroad-1016.xml --input <path_to_video_or_image>

Supported models for this demo application are shown as follows:

• mobilenet-ssd
• pedestrian-and-vehicle-detector-adas-0001
• pedestrian-detection-adas-0002
• person-detection-retail-0013
• person-vehicle-bike-detection-crossroad-0078
• person-vehicle-bike-detection-crossroad-1016
• ssd300
• ssd512
• ssd_mobilenet_v1_coco
• ssd_mobilenet_v1_fpn_coco
• ssd_mobilenet_v2_coco
• ssdlite_mobilenet_v2
• single-human-pose-estimation-0001

Regards,

Wan

[Wan_Intel]

Hi RGVGreatCoder,

This thread will no longer be monitored since we have provided a suggestion. 

If you need any additional information from Intel, please submit a new question.

Regards,

Wan