YOLOv2 object detection algorithm using Darkflow

This story introduces how object detection can be done. It will be helpful if you plan to build an application which benefits from object detection. YOLOv2 algorithm is used behind the scene. However, I have used open source implementation, “Darkflow”, so you don’t need to worry about the detail.
You need 4 steps to perform object detection in short. There are some more details, so if you actually want to use my codes please visit my github repository for this story. There is the jupyter notebook for this story.

Installing Darkflow

Actually, you can visit Darkflow’s github repository and explore the installation guide. I am just going to leave the instruction as simple as possible here.
$ pip install Cython
$ git clone https://github.com/thtrieu/darkflow.git
$ cd darkflow
$ python3 setup.py build_ext --inplace
$ pip install .

Downloading Weights

There are two ways of downloading the pre-trained weights. First of all, you can download it from the official YOLO project webpage. Second, you can download it from here which Darkflow author’s own trained version. I decided to use the second method since I am using Darkflow implementation. You can give a shot for the first method if you want.
Red circle indicates the weight file you need to download

Building the Model

As you can see below, it is quiet simple to build the model. First, you need to define options object. Then, you need to instantiate TFNet class object with the options.
The options is a specification of the model and its environment. The screenshot below is the full list you can specify. I am going to use “model”, “load”, “threshold”, and “gpu” options for this story.
Darkflow Model Options
The model option asks you to specify what model you want to use. There are pre-defined model, and I used yolo.cfg here. This file contains a full of description of the model’s architecture.
The load option is for specifying which weight file you want to use. As I suggested to download yolo.weights from here, I specified it. If you have your own pre-trained weight files, this is where you let the model knows (like after you trained custom objects, your system will produce the specific weight file).
The gpu option is used when you have GPU on your system. It will boost the speed of predicting task.
The threshold option is the bottom line of confidence probability value for keeping detected objects. I set it to 0.1, and it seems pretty low. However, I chose this value to experiment how many object the model can detect, and I could filter later on.
After the model is fully loaded, you will have the output like above.

Detecting Objects and Draw Boxes on a Picture

The one line of code above is everything you need to do to detect objects. The only specified parameter is an image represented in numpy array style. If you print out the results, you will see there are a list of the object shown below. It looks very straight forward what each key means in the object.
I defined boxing function below so that I could reuse for testing on an image and video as well. I used opencv-python module for putting boxes and labels on an image.
With the boxing function, you can display the resulting image like below.
Or you can perform the same task on a video as well like follow.
* Please check supported video output format here depending on your operating system and set it on cv2.VideoWriter_fourcc function. In my case, I have tested on Windows 10, so I specified *’DIVX’.

Results and Evaluation

Here are some of the results I have experimented with.
It looks nice on various kinds of pictures. However, I see some drawbacks on the second and fourth pictures.
For the second picture’s case, It looks like the YOLO doesn’t expect a dog being in the middle of a picture. I assume that lots of pre-trained pictures had dogs on the ground.
YOLO is actually very good at detecting people. However, for the fourth picture’s case, people in a dynamic unexpected pose doesn’t look like being easily detectable. Also, when too many objects are overlapped, it looks like difficult to detect every single one of them.

Comments

Post a Comment

Popular posts from this blog

Maxpooling vs minpooling vs average pooling

Image
Pooling is performed in neural networks to reduce variance and computation complexity. Many times, beginners blindly use a pooling method without knowing the reason for using it. Here is a comparison of three basic pooling methods that are widely used. The three types of pooling operations are: Max pooling: The maximum pixel value of the batch is selected. Min pooling: The minimum pixel value of the batch is selected. Average pooling: The average value of all the pixels in the batch is selected. The batch here means a group of pixels of size equal to the filter size which is decided based on the size of the image. In the following example, a filter of 9x9 is chosen. The output of the pooling method varies with the varying value of the filter size. The operations are illustrated through the following figures. Average, Max and Min pooling of size 9x9 applied on an image We cannot say that a particular pooling method is better over others generally. The choice of pooling operation is made...
Read more

Generative AI - Prompting with purpose: The RACE framework for data analysis

Image
  Prompting with purpose: The RACE framework for data analysis We've already established that the quality of your prompts directly influences the quality of your AI-generated outputs. A well-crafted prompt acts as a guiding light, leading the AI toward insightful and actionable results. But how can we consistently achieve this level of clarity and precision in our prompts? Let’s look into the RACE framework—an invaluable tool to elevate your prompts, ensuring they are clear, insightful, and evidence-based. This framework offers a structured approach to prompt engineering, helping you break down complex tasks into manageable steps and communicate your intentions effectively to the AI. This reading will assist you with developing a powerful methodology with generative AI for your data analysis tasks, leading to more accurate, relevant, and impactful results. What is the RACE framework? RACE is an acronym that stands for Research, Answer, Cite, and Expand. It serves as a valuable fram...
Read more

Best Practices for Storing and Loading JSON Objects from a Large SQL Server Table Using .NET Core

 Best Practices for Storing and Loading JSON Objects from a Large SQL Server Table Using .NET Core Introduction JSON (JavaScript Object Notation) is widely used for storing structured data in SQL Server. However, when dealing with large tables, inefficient handling of JSON data can lead to performance bottlenecks, high storage costs, and slow queries. In this article, we will cover best practices for saving and loading JSON objects from large SQL Server tables using .NET Core. We will focus on storage strategies, indexing, efficient querying, and optimized data retrieval using Entity Framework Core (EF Core) and Dapper. --- 1. Choosing the Right Storage Strategy SQL Server supports JSON natively, but choosing the correct storage method depends on your use case. Option 1: Store JSON as NVARCHAR(MAX) (Best for Flexibility & Simplicity) ✔ Ideal for semi-structured or frequently changing data. ✔ Suitable when you need to store the entire JSON object but rarely query individual fiel...
Read more