Cameras and other photography systems include the capability to detect objects of interest. Object detectors trained using sample data have a classification loss, e.g., due to insufficient training owing to a finite number of negative samples used during training. An increase in complexity of the object detector increases running time which necessitates a tradeoff between recall, precision, and speed. A common approach is to minimize the average loss across the entire training database. This disclosure proposes a new framework that takes the final image quality into account while training an object detector, by using a modified loss calculation function for the object detection framework used in photography. The framework enables better decisions regarding the various tradeoffs involved in loss calculation. The loss of classification during training is calculated by comparison of an image captured with and without successful detection of the object. The object detector, trained to take into account the impact of detecting or missing an object on a captured image, can improve the quality of captured images.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Yang, Ruiduo and Levoy, Marc, "Training a machine-learning based object detector for use in photography", Technical Disclosure Commons, (February 28, 2019)