Neural Networks for the Intel® NCS 2 and the Intel® Movidius™ Neural Compute Stick with the OpenVINO™ toolkit

This directory contains multiple subdirectories. Each subdirectory contains software, data, and instructions that pertain to using a specific neural network (based on any framework) with a Neural Compute device such as the Intel® NCS 2. Along with the trained network itself, examples are provided via Makefile that show how the OpenVINO Model Optimizer can be used to compile the network to Intermediate Representation (IR) and also how to create a program that uses that IR model for inferencing. The sections below are categorized by network type and include a brief explaination of each network.

This directory should be a preferred location for neural networks rather than the caffe or tensorflow directories. The caffe and tensorflow directories are legacy directories. New networks should created in this directory.

Image Classification Networks for Neural Compute devices

Image Classification Network	Description
age_gender_net	Network that classifies a face image into age ranges.
alexnet	Network that classifies images based on the 1000 categories described in Large Scale Visual Recognition Challenge 2012 (ILSVRC2012).
googlenet_v1	BAIR/BLVC GoogleNet is a network based on googlenet_v1, the winner of ILSVRC 2014, that classifies images based on the 1000 categories described in Large Scale Visual Recognition Challenge 2012 (ILSVRC2012).
inception_v1	Network that classifies images based on the 1000 categories described in Large Scale Visual Recognition Challenge 2012 (ILSVRC2012).
inception_v2	Network that classifies images based on the 1000 categories described in Large Scale Visual Recognition Challenge 2012 (ILSVRC2012).
inception_v3	Network that classifies images based on the 1000 categories described in Large Scale Visual Recognition Challenge 2012 (ILSVRC2012).
inception_v4	Network that classifies images based on the 1000 categories described in Large Scale Visual Recognition Challenge 2012 (ILSVRC2012).
resnet_50	Deep Residual network with 50 layers that classifies images based on the 1000 categories described in Large Scale Visual Recognition Challenge 2012 (ILSVRC2012).
mnist	Network that classifies handwritten digits. This network is based on tensorflow mnist_deep. This project also includes an option to train the network yourself.
mobilenets	The mobilenets (as described in the MobileNets Paper) are small, low-latency, low-power Convolutional Neural Networks for Mobile Vision Applications. They are parameterized for a variety of different uses. Multiple trained networks with different parmameter values are compiled in this directory.
squeezenet_v1.0	Accuracy similar to AlexNet with many fewer parameters and small model size as described int the squeezenet paper. Network that classifies images based on the 1000 categories described in Large Scale Visual Recognition Challenge 2012 (ILSVRC2012).
ssd_inception_v2_traffic_light	Single Shot Detector with inception v2 that was trained on 2 different labels on traffic light (green and red).

Object Detection Networks for Neural Compute devices

Object Detection Network	Description
face_detection_retail_0004	This is a nework that is trained to find faces in general. More information specific to this network is available.
ssd_inception_v2_gesture	Single Shot Detector with inception v2 that was trained on 7 different 7 different hand gestures.
ssd_inception_v2_food	Single Shot Detector with inception v2 that was trained on 10 different foods.
ssd_mobilenet_v1_caffe	MobileNet Single Shot Detector takes an image, detects the 20 PASCAL object classes as specified in the (Visual Object Classes Challenges), their bounding boxes, and classifications.
tiny_yolo_v1	This Tiny You Only Look Once model is based on tiny-yolo v1 DarkNet model . Given an image, detects the 20 PASCAL object classes as specified in the (Visual Object Classes Challenges), their bounding boxes, and classifications. Requires some post processing of results to narrow down relevant boxes.
tiny_yolo_v2	This Tiny You Only Look Once model is based on tiny-yolo v2 DarkNet model . Given an image, detects the 20 PASCAL object classes as specified in the (Visual Object Classes Challenges), their bounding boxes, and classifications. Requires some post processing of results to narrow down relevant boxes.
tiny_yolo_v3	Tiny Yolo (You Only Look Once) v3 network. For more information see https://github.com/mystic123/tensorflow-yolo-v3.git

Misc Networks for Neural Compute devices

Network	Description
facenet	FaceNet is a nework that is trained to find and quantify landmarks on faces in general. By comparing the face landmark quantification values (network inference output) on two images, it is possible to determine how likely the two faces are of the same person. This is based on work by David Sandberg.
segmantic segmentation adas 0001	Semantic segmentation adas 0001 is a nework that is trained to do semantic segmentation on 20 different classes. More information specific to this networkis avaialble.