Language-Based Object Detection Using Negative Descriptions and Images

The patent segments the label space into positive labels (objects present in the image) and negative labels (objects absent from the image). This segmentation allows the system to independently process and utilize negative labels, resolving the contradiction by organizing the expanded label space into distinct functional categories that can be effectively managed despite the overall increase in label quantity.

The patent inverts the traditional object detection approach by explicitly generating and processing negative labels alongside positive labels. Instead of only detecting what is present, the system generates descriptions of what is absent and uses these negative labels for training, thereby improving negative label utilization while maintaining the expanded label space capability.

The patent applies preliminary action by pre-generating negative images and their corresponding negative labels before the actual object detection training process. This allows the computationally intensive generative modeling to be performed in advance, creating a ready-to-use augmented dataset that can then be efficiently utilized during training without repeatedly invoking complex generative models.

The patent uses copying by generating negative images that are modifications or transformations of existing positive images rather than creating entirely new images from scratch. By copying the structure and content of original images and selectively modifying them to represent absent objects, the system reduces computational complexity while maintaining data augmentation effectiveness.