tiny yolo v3

Introduction

The tiny yolo v3 network can be used for object recognition and classification. This model was trained with the Coco data set and can detect up to 80 classes. See https://pjreddie.com/darknet/yolo/ for more information on this network.

This sample utilizes the OpenVINO Inference Engine from the OpenVINO Deep Learning Development Toolkit and was tested with the 2020.1 release.

The provided Makefile does the following:

1. Clones a Tensorflow yolo repo that will help to convert the darknet weights to tensorflow.
2. Download the labels and weights from the Tiny Yolo v3 site.
3. Converts the weights and generates a Tensorflow frozen pb file.
4. Compiles the Tensorflow frozen pb file to an IR (intermediate representation) using the Model Optimizer.
5. Runs the provided tiny_yolo_v3.py program that does a single inference on a provided image as an example on how to use the network using the Inference Engine Python API.

Model Information

Inputs

• name: ‘inputs’, shape: [1x3x416x416], Expected color order is BGR. Original network expects RGB, but for this sample, the IR is compiled with the –reverse_input_channels option to convert the IR to expect the BGR color order.

  Outputs

• name: ‘detector/yolo-v3-tiny/Conv_12/BiasAdd/YoloRegion’, shape: [1, 255, 26, 26].
• name: ‘detector/yolo-v3-tiny/Conv_9/BiasAdd/YoloRegion’, shape: [1, 255, 13, 13].

Tiny yolo v3 divides the image into 13x13 and 26x26 grid cells. Each grid cell has 3 anchor boxes and each anchor box has an object score, 20 class scores, and 4 bounding box coordinates. The tiny_yolo_v3.py code reads the number of classes through the –labels argument. If your model has a different amount of classes from the default model, please make sure your labels file has the correct amount of classes.

Note: The ‘26’ and ‘13’ values in each of the outputs represent the number of grid cells for each output. Note: The ‘255’ in each output represents each of the three anchor box class scores, object scores and bounding boxes. 80 class scores + 4 bounding box values + 1 object score = 85 values. 85 values * 3 anchor boxes= 255 values.

All scores will need to be filtered to achieve desired results. Please see Algorithm Thresholds section.

Running this Example

make run

Note: You can also specify the camera using the command: make run INPUT=cam Note: You may need to adjust some values to better detect objects

Algorithm Thresholds

There are a few thresholds in the code you may want to tweek if you aren’t getting results that you expect:

• DETECTION_THRESHOLD: This is the minimum probability allowed for boxes returned from tiny yolo. This should be between 0.0 and 1.0. A lower value will allow more boxes to be displayed.
• IOU_THRESHOLD: Determines which boxes from Tiny Yolo should be separate objects vs identifying the same object. This is based on the intersection-over-union calculation. The closer this is to 1.0 the more similar the boxes need to be in order to be considered around the same object.

Makefile

Provided Makefile describes various targets that help with the above mentioned tasks.

make run or make run_py

Runs a sample application with the network.

make help

Shows makefile possible targets and brief descriptions.

make all

Makes the follow items: deps, data, compile_model.

make compile_model

Uses the network description and the trained weights files to generate an IR (intermediate representation) format file. This file is later loaded on the Neural Compute Stick where the inferences on the network can be executed.

make install-reqs

Checks required packages that aren’t installed as part of the OpenVINO installation.

make uninstall-reqs

Uninstalls requirements that were installed by the sample program.

make clean

Removes all the temporary and target files that are created by the Makefile.