保留正文与补采附件。源页面当前未重新核验；具体限制见交付清单。

Revisit consent button We value your privacy We use cookies to enhance your browsing experience, serve personalised ads or content, and analyse our traffic. By clicking "Accept All", you consent to our use of cookies. Customise Reject All Accept All Customise Consent Preferences We use cookies to help you navigate efficiently and perform certain functions. You will find detailed information about all cookies under each consent category below. The cookies that are categorised as "Necessary" are stored on your browser as they are essential for enabling the basic functionalities of the site. ... Show more NecessaryAlways Active Necessary cookies are required to enable the basic features of this site, such as providing secure log-in or adjusting your consent preferences. These cookies do not store any personally identifiable data.

Functional Functional cookies help perform certain functionalities like sharing the content of the website on social media platforms, collecting feedback, and other third-party features.

Analytics Analytical cookies are used to understand how visitors interact with the website. These cookies help provide information on metrics such as the number of visitors, bounce rate, traffic source, etc.

Performance Performance cookies are used to understand and analyse the key performance indexes of the website which helps in delivering a better user experience for the visitors.

Advertisement Advertisement cookies are used to provide visitors with customised advertisements based on the pages you visited previously and to analyse the effectiveness of the ad campaigns.

Uncategorised Other uncategorised cookies are those that are being analysed and have not been classified into a category as yet. No cookies to display. Reject All Save My Preferences Accept All Toggle SearchToggle Mobile Navigation Publish AI, ML & data-science insights to a global community of data professionals. Sign inSubmit an Article LatestEditor's PicksNewsletterShipAI

Write For TDS LinkedInXToggle Search Artificial Intelligence

Easy Object Detection with Yolo-NAS

Learn how to do object detection with Python using yolo-NAS Ivo Bernardo August 3, 20246 min read Photo by googledeepmind @ Unplash.com Photo by googledeepmind @ Unplash.com YOLO (You only look once) revolutionized the computer vision arena. The first version of YOLO was released in 2016 by Joseph Redmon et. al and it smashed benchmarks both in terms of speed and accuracy. When it comes to object detection, YOLO has been a favorite of Data Scientists and Machine Learning engineers and the go-to model when it comes to segmenting entities in images. Since it was launched, YOLO had many new iterations that improved several setbacks of previous versions, namely:

The last version of YOLO is YOLO v9 (https://arxiv.org/abs/2402.13616). one important thing to be aware is that every computer vision and object detection model is evaluated on two parameters: Accuracy (defined by metrics related to computer vision segmentation) __ and Speed (defined by latency in the inference). One example of how CV algorithms are evaluated is shown below: Learn this step by step with the interactive AI Engineer, AI and Data Scientist and Machine Learning roadmaps. Comparison of different YOLO models vs. EfficientDet - Image by https://blog.roboflow.com/yolov5-is-here/ Comparison of different YOLO models vs. EfficientDet - Image by https://blog.roboflow.com/yolov5-is-here/ In this example, I'm going to show you how to run your own YOLO models on images and videos, so that you can rform object detection and segmentation. The entire code I'm executing is linked here, on this Google Colab Notebook. Part of the code and blog post was inspired by this great notebook by Harpreet Sahota. Let's start!

Loading the Model

We'll start by training our own computer vision model... Of course not, that would be a waste of time :-) We'll start by loading the pre-trained version of a large YOLO model. In this case, we'll use an implementation of YOLO with NAS (Neural Architecture Search). Neural architecture search is a common implementation of neural network optimization techniques to improve the automatic choice of parameters in the deep learning model. The model we'll use was trained on the COCO dataset (common objects in context). These weights already contain a very good baseline to detect objects on images. First, let's add a couple of libraries we will need to our environment, namely:

diff

!pip install torchinfo
!pip install super_gradients

Next, we'll load the functions and libraries we will need: javascript

from super_gradients.training import models
from torchinfo import summary
import torch

Note: If you are using google colab and have a prompt to restart the kernel after installing super_gradients , do it. It's a known issue from using super_learners in that environment Then, we start by loading our COCO pretrained model into memory using super_gradients : ini

# Using Yolo NAS Large
yolo_nas_l = models.get("yolo_nas_l", pretrained_weights="coco")

Cool! We have our pretrained model stored in yolo_nas_l . Picking up these pretrained versions of YOLO is super easy with models . torchinfo gives us a very cool view of our model architecture: python

summary(model=yolo_nas_l,
        input_size=(16, 3, 640, 640),
        col_names=["input_size", "output_size", "num_params", "trainable"],
        col_width=20,
        row_settings=["var_names"]
)

Out of curiosity, if you would like to check the YOLO architecture (from the original paper) check out the image below: Image Source: https://www.researchgate.net/publication/329038564_Complexity_and_accuracy_analysis_of_common_artificial_neural_networks_on_pedestrian_detection Image Source: https://www.researchgate.net/publication/329038564_Complexity_and_accuracy_analysis_of_common_artificial_neural_networks_on_pedestrian_detection Here's a preview of the architecture of our YOLO NAS model: Preview of Yolo-NAS model - image by author. Preview of Yolo-NAS model - image by author. With the YOLO weights loaded, it's time to play around with image segmentation!

Segmenting Images

Let's start with a simple image featuring a desk as the centerpiece, surrounded by various objects. In this image, we can see:

Image 1 - Image by dawidliberadzki at Unsplash Image 1 - Image by dawidliberadzki at Unsplash In our code, we only request bounding boxes where our model has a confidence level exceeding 55%. Only labels for these items will show up in the segmented image. makefile

url = "https://images.unsplash.com/photo-1476365518243-f738bf58443d?q=80&w=2787&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D"
yolo_nas_l.predict(url, conf=0.55).show()

Labeled Image with YOLO - Image by Author Labeled Image with YOLO - Image by Author Nice! Our model was able to confidently detect the chairs, laptop, desk and plant (although it missed the lamps, headphones and that hi-tech jug). Let's see what happens if we lower our confidence threshold: makefile

url = "https://images.unsplash.com/photo-1476365518243-f738bf58443d?q=80&w=2787&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D"
yolo_nas_l.predict(url, conf=0.40).show()

Labeled Image with YOLO 40% Confidence - Image by Author Labeled Image with YOLO (40% Confidence) - Image by Author Ah, interesting! Lowering our threshold has led to our first instance of misclassification. In this case, the headphones were incorrectly identified as a computer mouse. Reducing the threshold will inevitably result in more errors in object detection and segmentation. Let's try an image with a more crowded scenario: Image by madebyvadim @unsplash.com Image by madebyvadim @unsplash.com What objects will our YOLO model segment? As this image has more elements, I'll lower the confidence of the inference: makefile

url = "https://images.unsplash.com/photo-1496664444929-8c75efb9546f?q=80&w=2940&auto=format&fit=crop&ixlib=rb-4.0.3&ixid=M3wxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8fA%3D%3D"
yolo_nas_l.predict(url, conf=0.3).show()

Labeled Image with YOLO 30% Confidence - Image by Author Labeled Image with YOLO (30% Confidence) - Image by Author As we can see, YOLO fails to segment some of the objects - particularly smaller / thiner objects. Nevertheless, performance is quite good, particularly for such a crowded image. To improve some the performance, transfer learning with images related to your use case is extremely important. Now, with**super_gradients**, we're abstracting most of code that actually prepares and predicts the boundaries on our image. If you would like to have a deeper insight on the foundations of these models, check out the article Training a Computer Vision Algorithm from my Pytorch Introduction Series . In the next section, we'll learn how to use YOLO-NAS, but in the context of segmenting videos.

For this section, I'll be using a traffic short video on Youtube: Traffic Video - https://www.youtube.com/watch?v=CftLBPI1Ga4 Traffic Video - https://www.youtube.com/watch?v=CftLBPI1Ga4 We'll start by downloading this video into our Google Colab environment. We can also display the video by using IPython 's handy YoutubeVideo function: java

from IPython.display import YouTubeVideo

video = YouTubeVideo('CftLBPI1Ga4')
display(video)

To download our video, I'll use youtube-dl: bash

video_url = f'https://www.youtube.com/watch?v=CftLBPI1Ga4'

!pip install -U "git+https://github.com/ytdl-org/youtube-dl.git"
!python -m youtube_dl -f 'bestvideo[ext=mp4]+bestaudio[ext=m4a]/mp4' "$video_url"

We have almost everything ready. We just need to set the output path for the segmented video and we're ready to pass our video through the YOLO model! ini

input_video_path = f"/content/Cars, Busy Streets, City Traffic - No Copyright Royalty Free Stock Videos-CftLBPI1Ga4.mp4"
output = "traffic.mp4"

All set! Let's run our segmentation model: lua

device = 'cuda' if torch.cuda.is_available() else "cpu"
yolo_nas_l.to(device).predict(input_video_path).save(output)

Note: This code may take a bit to run, depending on the configurations of your local system and the availability of GPUs on Google Colab .. and that's it! You can check out the segmented video on my youtube channel "The Data Journey". Here's a cool preview: Segmented Traffic Video - Image by Author Segmented Traffic Video - Image by Author

YOLO models are extremely powerful! You can use them to segment images or videos quite fast with just a couple lines of code. Using this code that I've shown here, it's super easy to play around with your own media and experiment with this great computer vision model. As you may noticed, particularly in complex images or videos, YOLO base models still have some flaws that need to be corrected. As I've mentioned this can be improved quite well with model fine tuning. Check out this post for more info. And, if you are curious on how the building blocks of Deep Learning work, feel free to check my pytorch introduction series, that should give you a good overview on the basics of deep learning:

Thank you for reading this post! And again, thank you to Harpreet for the awesome notebook. If you want to read / see more content related to AI and DS, subscribe my youtube Channel "The Data Journey ": https://www.youtube.com/@TheDataJourney42 https://www.youtube.com/@TheDataJourney42

Register Now to Continue Reading

Unlock Towards Data Science with a Free Account

All articles on Towards Data Science are available to registered users. Create a free account today to access this article and thousands more. Register for freeSign in to your account

Written By

Ivo Bernardo

See all from Ivo Bernardo Artificial IntelligenceComputer VisionData ScienceDeep LearningYolo Share This Article

Towards Data Science is a community publication. Submit your insights to reach our global audience and earn through the TDS Author Payment Program. Write for TDS

Related Articles

Hadoop vs Spark: Overview and Comparison

Programming A summary and comparison of Spark and Hadoop Rian Dolphin March 1, 20227 min read

LightGBM: The Fastest Option of Gradient Boosting

Machine Learning Learn how to implement a fast and effective Gradient Boosting model using Python Gustavo Santos January 12, 20257 min read

How to work with object detection datasets in COCO format

Machine Learning A comprehensive guide to defining, loading, exploring, and evaluating object detection datasets in COCO format using FiftyOne Eric Hofesmann February 19, 202110 min read

BakeryScan and Cyto-AiSCAN

Machine Learning and the unexpected impact of an AI that identifies bread. Matthew Turner September 7, 20217 min read

Don't Do Laundry Today, It Will Be Cheaper Tomorrow

Machine Learning Analysing electricity price changes in London through causal inference Alvaro Corrales Cano October 21, 202417 min read

3 Subtle Mistakes That Can Undermine Your SQL Query Performance

Artificial Intelligence And how to mitigate them Soner Yıldırım May 11, 20234 min read

Deploy Any Python Project to Kubernetes

Programming It's time to dive into Kubernetes and bring your Python project to Cloud with this fully-fledged project template! Martin Heinz April 15, 20206 min read

Get Uncertainty Estimates in Regression Neural Networks for Free

Artificial Intelligence Given the right loss function, a standard neural network can output uncertainty as well Robert Kübler April 22, 202210 min read

How to Perform Bulk Insert/Update/Upsert Actions with SQLAlchemy ORM

Data Science Learn efficient ways to perform bulk actions with the latest version of SQLAlchemy Lynn G. Kwong May 14, 202411 min read

Your home for data science and AI. The world's leading publication for data science, data analytics, data engineering, machine learning, and artificial intelligence professionals. Subscribe to Our NewsletterWrite For TDS•About•Advertise•Privacy Policy•Terms of Use © Insight Media Group, LLC 2026•The New Stack•roadmap.sh•Creator Network

Recommended Articles

Close