AI Image Recognition Dataset Creation

Client

A global urban mobility technology company

Objective

The objective of the study was to support the development of an AI image recognition model. The client aimed to build an AI system capable of accurately understanding and interpreting a base image (background scene), and to evaluate whether the AI could detect and identify changes made to that base image — such as the addition of objects or people — and correctly recognize the inserted elements.

Methodology

The study was designed to create image datasets for training and evaluating an AI image recognition model. Two types of image sets were developed.

1. Human-Inclusive Image Set

A single natural base image (original scene) was used as the starting point. One human and two objects were captured separately and later combined digitally to create the final composite image. Each image set consisted of five images:

• Original base image (natural scene)

• Individual image of the human

• Individual image of Object B

• Individual image of Object C

• Final composite image containing the base image, one human, and two objects

2. Object-Inclusive Image Set

A single natural base image was used as the starting point. Three objects were captured separately and later combined digitally to create the final composite image. Each image set consisted of five images:

• Original base image (natural scene)

• Individual image of Object A

• Individual image of Object B

• Individual image of Object C

• Final composite image containing the base image and all three objects

Image Capture Process

To ensure consistency and image quality, all photographs were captured using a tripod to maintain a stable camera position throughout the image collection process. A variety of real-world backgrounds were selected to create a diverse and natural image dataset, including office buildings, outdoor parking areas, public parks, gyms, and employees' homes. For each location, a natural base image was captured first, followed by individual photographs of the human and objects under the same camera setup, later combined to create the final composite image while maintaining consistent perspective, lighting, and alignment.

Scale & Client Feedback

Approximately 2,000 image sets were created and submitted for AI model training and validation. Following the client's quality control review, around 1,000 image sets met the required quality standards and were accepted for use. The remaining sets were rejected due to issues such as lighting variation, image saturation, and inconsistent exposure — reinforcing the importance of tight quality control at the point of capture.

Key Learnings

• Diverse real-world backgrounds improved the robustness and variety of the AI training dataset.

• Using a tripod ensured consistent framing and camera positioning, reducing variation across image sets.

• Capturing the base image, human, and objects separately enabled accurate, realistic composite images.

• Maintaining consistent lighting, exposure, and image quality control was critical, as even minor variations significantly impacted dataset acceptance.

• The structured five-image sequence provided a reliable dataset for training and evaluating the AI model's ability to detect and understand scene changes.

Building an image dataset for AI model training? Market Xcel can design and execute structured, quality-controlled image capture at scale — get in touch to discuss your dataset requirements.