Scientists in India used deep learning techniques and Microsoft’s COCO dataset to train a convolutional neural network (CNN) to search for a certain person based on details like height and clothing color as well as gender.
A simple request such as “men wearing green shirts who are 180cm tall” can narrow down the search for an entire array of different videos to just the frames that contain the men meeting the criteria.
The researchers state that the algorithm is capable of correctly identifying the target 28 times out of 41, while working with a very challenging dataset.
The CNN system could also prove very useful in cases where a timeline is needed for a certain individual, which could be based on surveillance footage. Just think about the case of a missing person, for example, where footage that might stretch out into hundreds of hours of multiple surveillance feeds could be narrowed down just to the clips that feature the relevant person.
The researchers hope for further development of the system and are still attempting to improve it. You can read the entire paper here, if you want to know more.