Automated Object Labeling System for Scalable and Efficient Real-Time Recognition

Object recognition in real-time enables transformative applications across various industries. Real-time solutions are crucial in areas such as quality control, security, and inspection. Applications like sorting waste or recycling streams with numerous product types each existing in various states, require recognizing thousands or even tens of thousands of object classes. These systems rely on large datasets of manually labeled images for each class, which is expensive and time-consuming.

To address these challenges, we developed a method for automatically identifying and labeling objects for training recognition systems, along with a machine to implement it. This low-cost device uses a camera to capture images in both UV and visible light spectrums, allowing marked objects to be recognized. The objects are moved along a conveyor, imaged from multiple angles, and shuffled for repeated imaging. The machine then generates training labels, including bounding boxes or object masks, by combining data from both light spectrums. Initial testing shows promising results, and we are working to scale it for more object classes. The impact of labeling errors on recognition accuracy is also demonstrated by introducing controlled variations in the label boxes.

In conclusion, our automated labeling system speeds up the process, reduces human error, scales efficiently to large datasets, and opens new possibilities for object recognition applications.