Consumer tech leader builds 160K-image zero-AI CV dataset

Overview

A leading global consumer technology company was building the ground-truth dataset behind its next-generation computer-vision models ” object removal, background-fill, and synthetic-media detection.

Those models are only as honest as the data underneath them. Train an object-removal model on edited or AI-generated images and it learns artifacts, not reality.

The company needed 160,000 authentic, highly diverse RAW images ” plus 5,000 counterfactual pairs showing the same scene with and without a physical object ” captured under controlled, repeatable conditions, with zero pixel-level manipulation, generative AI, or synthetic data anywhere in the set.

It also needed the capture to span real sensors and real places: professional full-frame, APS-C, and medium-format cameras alongside approved smartphones, across natural, artificial, and low-light conditions and thousands of geographic contexts.

Firstsource ran the collection end to end ” sourcing, controlled capture, paired counterfactual shoots, quality control, and structured metadata. This was Intelligence that Operates: model-grade ground truth produced on the cadence the program needed.

Challenges

The dataset carried three hard constraints that generic image collection cannot meet:

• Counterfactual pairs leave no room for error: Each pair had to show the identical scene with and without a physical object ” same exposure, same focal parameters, an absolutely static background while the object was removed. A single pixel of drift between frames makes the pair useless for training object-removal models.
• Authenticity with zero synthetic shortcuts: The mandate was 100% authentic capture ” no generative AI, no synthetic data, no pixel-level editing. At 160,000 images the pull toward augmentation is constant, and the dataset's value as ground truth depended on never giving in to it.
• Diversity across sensors, light, and place: The set had to hold up across a wide mix of professional cameras and approved smartphones, three lighting conditions, and thousands of geographic contexts ” so models trained on it stay sensor-agnostic and robust to real-world lens and lighting variation.

How We Made It Happen

Firstsource designed a controlled-capture program built for repeatability and authenticity, then ran it end to end across sensors and locations.

• Controlled, repeatable capture: A disciplined capture standard held exposure, focal parameters, and background fixed across each counterfactual pair, so the only difference between two frames was the object itself.
• Multi-sensor, multi-context sourcing: Capture spanned a deliberate mix of professional cameras and approved smartphones, across varied lighting and thousands of locations ” diversity built in at the source, not added afterward.
• In-process quality control and structured metadata: Image quality was verified before anything was uploaded, and every image shipped with EXIF telemetry and structured metadata associations ” so the dataset arrived clean, paired, and ready to train.

Conclusion

A computer-vision model is only as trustworthy as the images beneath it, and authentic counterfactual ground truth at this scale is what lets a model learn reality instead of artifacts. Firstsource delivered that dataset ” and the capture discipline behind it ” as Intelligence that Operates.