AI with Human-like Capability: Tracking Any Object Amodally

Webmaster’s Home (ChinaZ.com) December 25 news: Tracking Any Object Amodally is a project developed by Carnegie Mellon University and Toyota Research Institute to enable artificial intelligence to behave like humans Track an object in its entirety and understand its complete structure even when partially obscured or not fully visible.

In this project, we trained a computer to "understand" and track the complete shape and position of a partially occluded object. This is critical for autonomous vehicles as it can more safely and reliably identify and track partially obscured pedestrians or other vehicles in complex environments

Project address: https://tao-amodal.github.io/ Project address: https://tao-amodal.github.io/

Code link: https://github.com/WesleyHsieh0806/TAO-Amodal

In order to improve object tracking technology, they specially designed a data set called TAO-Amodal. This data set contains numerous video sequences containing various occluded or partially visible objects, and provides detailed annotation information to help artificial intelligence better understand and track those objects that we can only partially see

TAO-Amodal dataset contains 880 different categories covering thousands of video sequences. The dataset includes amodal and modal bounding boxes for completely invisible, partially out-of-box, and occluded objects. The main purpose of this dataset is to evaluate the capabilities of current trackers in occlusion reasoning by tracking amodal perception of any object

In addition, the project also developed a lightweight module called the "Amodal Expander Plug-in" to enhance the functionality of the object tracker. This plugin converts a standard Modal tracker into an Amodal tracker, making it more efficient and accurate when tracking partially obscured or not fully visible objects

According to test results on the TAO-Amodal data set, this technology achieved improvements of 3.3% and 1.6% in detecting and tracking occluded objects. Especially in terms of tracking people, the performance is improved by 2 times compared with existing modal tracking technology. The success of this project will greatly improve the intelligence of computer vision systems, making them more human-like when dealing with occluded objects, thereby playing a greater role in fields such as autonomous driving and video surveillance.

The above is the detailed content of AI with Human-like Capability: Tracking Any Object Amodally. For more information, please follow other related articles on the PHP Chinese website!