Toolify Toolify logo
open
Home AI News Building a Custom Image Recognition Software: A Case Study

Building a Custom Image Recognition Software: A Case Study

Updated on Oct 18,2025

Table of Contents

In today's data-driven world, image recognition software is becoming increasingly important. This blog post details the compelling journey of CarSales.com.au in building their own image recognition system, Cyclops. It dives into the problems they faced, their innovative solution, and the significant benefits they achieved by implementing this technology. Discover how AI and machine learning transformed their processes, reduced costs, and improved accuracy in image classification. This story offers valuable insights for any business considering custom AI solutions. Image recognition, car sales, AI, Machine learning, software, technology, Cyclops.

Key Points

CarSales.com.au faced a challenge in manually categorizing thousands of car images daily.

They developed Cyclops, a custom image recognition software, to automate this process.

Building Cyclops involved overcoming challenges related to inconsistent human categorization and the need for a large training dataset.

Transfer learning was used as a technique to train Cyclops, to drastically reduce development time and costs.

Cyclops significantly improved image classification accuracy and increased productivity.

The Image Recognition Challenge at CarSales.com.au

Understanding the Problem

CarSales.com.au, a leading online marketplace for buying and selling cars in Australia, handles a massive amount of visual data daily. With approximately 230,000 cars listed at any given time and an influx of around 5,000 new car listings every day

, the task of categorizing car images was becoming increasingly complex and time-consuming. This task heavily relied on a team of photographers who visited dealerships and captured photos of the cars. These photographers, like Jane, would upload hundreds of photos daily, which then needed to be manually sorted and tagged. The sheer volume of images made this process not only tedious but also prone to human error.

The Bottleneck: The manual categorization process was a significant bottleneck. Each photographer uploaded their images to a media library, where they were presented with a screen requiring them to select the correct 'angle' or 'view' for each photo. This meant identifying whether the image showed the front, rear, driver's side, or interior of the vehicle.

The Costs: This manual process had substantial financial implications. Each photographer took about 30 minutes to upload, classify, and finalize a batch of 500 images. With 80 photographers doing this task, it translated to considerable man-hours and a staggering $250,000 per year in operational costs , not to mention the approximately 18 million mouse clicks! This expensive and manual method was clearly unsustainable, therefore creating a need to replace humans with a more efficient, automated system. An automated solution was required to handle a large scale image processsing operation.

The Need for Automation

Manual Image Categorization Inefficiencies: The pain point was quite clear: the existing manual system for classifying car images was inefficient and expensive. An efficient and reliable automated process was needed. The manual, labor-intensive system was simply not able to handle a very large scale product processing operation. The inconsistencies of humans processing the images made the data unreliable and therefore hard to use

Potential Benefits of AI: Automation promised several key benefits: Reduced operational costs, faster processing times, and improved accuracy. Automation with AI offered the possibility of significantly reducing operating costs, processing imagery far more quickly, as well as significantly improving accuracy. The goal was not to eliminate the photographers but to free them from the repetitive task of manual categorization, allowing them to focus on more valuable tasks.

Defining success: The goal was to drastically improve efficiency (by automating a significant portion of the categorization task) while maintaining or even improving the accuracy of image classification.

Overcoming Technical Hurdles in Cyclops' Development

Addressing Human Inconsistency

One of the initial challenges the team encountered was the inconsistent nature of human image classification. What one person considered a 'front view' might be classified differently by another. As humans, we can not always be relied upon to be consistant in our actions [t: 291]. However, because there is only one Cyclops, only one answer gets returned in a car image classification operation. To combat this, and in order to overcome human based inconstistencies, the AI was provided a very large dataset of images to learn from. [t: 356] The system was trained to adopt a uniform standard for categorization.

Enhancing Accuracy Through Feedback Mechanisms

To ensure Cyclops continuously improved, a feedback mechanism was integrated into the system. Whenever a human reviewer corrected Cyclops' classification, this correction was fed back into the AI's training data, which allowed Cyclops to increase the accuracy of its outputs. The AI would then use this information to improve its future performance

. As time went on, this feedback loop allowed it to become increasingly sophisticated, with the benefit that its reliance on human intervention steadily diminished.

Build vs Buy a Custom Image Recognition Software

👍 Pros

Complete customization to specific business needs and datasets.

Full control over the algorithms, training data, and infrastructure.

Opportunity to create a competitive advantage with unique features.

Potential for long-term cost savings (especially for large-scale applications).

👎 Cons

Significant upfront investment in development and infrastructure.

Requires specialized expertise in AI, machine learning, and data science.

Longer development timelines compared to off-the-shelf solutions.

Ongoing maintenance and updates require dedicated resources.

The Tangible Benefits of Implementing Cyclops

Improved Efficiency and Productivity

Since introducing Cyclops, the Image Recognition accuracy of the car images at CarSales.com.au increased from the 85% of when it was done by a human, to now reach 97%. This improvement in effeciency also has translated into improved employee efficiency. As such, the 30 minutes employees would spend tagging car images, multiplied by 80 employees working 225 days of the year, resulted in a total quarter of a million saved.

This reduced operational costs as employees spent less time processing the image. The technology has essentially allowed them to take a whole day off!

Higher Data Quality

Beyond speed and cost savings, Cyclops has significantly improved data quality. The consistently high accuracy of image classification enables CarSales.com.au to leverage its visual data more effectively. Accurate image categorization supports better search results, more targeted Advertising, and a more engaging user experience on their platform. This leads to more valuable and effective digital processes and operations, powered by more reliable data.

Frequently Asked Questions About AI Image Recognition

How can image recognition benefit my e-commerce business?
Image recognition can automate product categorization, improve search accuracy, enhance visual search capabilities, and help moderate user-generated content. These capabilities can lead to increased sales, improved customer satisfaction, and better brand protection. A big factor here is that the human labour, in terms of cost and the effort hours, gets dramatically reduced.
What are the key challenges in implementing image recognition?
The key challenges include acquiring high-quality training data, dealing with inconsistent data, managing computing resources for model training, and ensuring the AI model is accurate and reliable. AI projects can be complex and challenging, but by properly selecting AI for appropriate challenges and business problems, an AI solution can be a real game changer.
How does transfer learning help in building image recognition systems?
Transfer learning significantly reduces the time and resources required to train an image recognition model. By leveraging pre-trained models, businesses can quickly adapt existing AI capabilities to their specific needs with limited training data. This process enables developers to drastically cut down on training times, making AI easier than ever to apply. Transfer learning is a very common technique used for AI implementations.

Related Questions About Building Custom Image Recognition Software

What are some other AI and machine learning techniques used in the automotive industry?
AI and machine learning are transforming various aspects of the automotive industry. Predictive maintenance uses AI to forecast when vehicle components may need maintenance, reducing downtime and repair costs. Autonomous driving systems rely heavily on AI for object detection, path planning, and decision-making. AI is also used in optimizing supply chain management, personalizing the in-car experience, and detecting fraud in insurance claims. It is now easier than ever to apply such technology to automotive problems than it was in the past.
How is AI currently used for detecting vehicle damage in insurance claims?
When a customer reports vehicle damage, AI can analyze photos submitted as part of the claim. The AI algorithms can be trained to identify different types of damage (dents, scratches, broken parts) and to estimate repair costs. This automates and expedites the claims process, reducing human intervention and improving consistency. A reliable AI based automated damage assessment process ensures there is no inconsistancy in claims processing, and that assessment turnaround times remain constant.
What type of image recognition solution should I use, SaaS (Software as a Service) vs developing an AI model from scratch?
The choice between SaaS (Software as a Service) and developing an AI model from scratch really depends on the specific requirements, resources, and level of customization needed. If you’re looking for a straightforward, ready-to-use solution, SaaS is likely the faster and more cost-effective option. You don’t need a deep technical team or extensive expertise in AI – the SaaS provider takes care of the infrastructure, algorithms, and maintenance. However, SaaS solutions might not be ideal if you have unique, complex use cases that need very specific customizations. In this instance, you might want to consider developing an AI model from scratch that has all the niche functions you want to implement. SaaS systems usually have a fixed fee associated with them, but may be less cost effective compared to owning an AI model from scratch in the long run. The right choice will have to be made on a case by case basis.
Home Assistant AI Security Check: LLM Automation Guide
Editing AI-Generated Essays: A High School Toolkit

Most people like

Related Articles
Top AI Tools
More Tools
Recommend
Resourse
Browse by Alphabet
Top 1000 AI Tools Directory
Read more
About
Contact Us
toolify
Toolify, The Best AI Websites & AI Tools Directory
English 简体中文 繁體中文 한국어 日本語 Português Español Deutsch Français Tiếng Việt

Copyright ©2026 toolify