01Problem
The client had already accumulated enough source material to support the next stage of its computer vision product, but the image library had not been created under a single data collection standard.
The target product appeared under different lighting conditions, against different surfaces, and at different angles. Some images had clear boundaries, while others contained reflections off plastic or glass device surfaces, weak contrast, partial occlusions, cropping, or nearby objects with similar shapes and colors. Before the data could be used for model training, the product had to be separated from the surrounding image through an accurate segmentation mask.
The main difficulty was not drawing one acceptable outline. It was ensuring that the same interpretation was applied thousands of times. Without detailed rules, two annotators could produce visually reasonable but inconsistent masks. One might include a narrow product component that another treated as background. Shadows and reflective edges — especially on clear or glossy plastic medical device surfaces — could also be handled differently depending on individual judgment.
The client's internal computer vision team could review representative examples and define the intended output, but it could not manually inspect the full dataset. The project therefore required a delivery partner capable of managing both annotation volume and quality assurance.
