Ant Group Launches Multimodal Remote Sensing Model “SkySense”

According to the Chinese media outlet Sina Tech, Ant Group has launched a 2 billion parameter multimodal remote sensing base model called SkySense. This is the latest research achievement of Ant’s Bailing large model in the multimodal field, and its paper has been accepted by the top computer vision conference CVPR 2024. SkySense can be used for terrain, crop observation, interpretation, and effectively assist agricultural production and management.

SkySense has been evaluated on a total of 17 international authoritative public datasets, with testing tasks including land use monitoring, high-resolution target identification, and object change detection, among 7 common remote sensing perception tasks. It has been compared with 18 other global mainstream models, including Prithvi jointly developed by IBM and NASA. The data shows that SkySense ranked first in all 17 evaluations. For example, in the international high-definition remote sensing object detection list FAIR1M 2.0, SkySense’s average precision (mAP) leads the second place by over 3%.

In the just-released CVPR 2024 paper acceptance results, the research achievements of SkySense have also been included. It is understood that Ant Group is planning to open up the SkySense model parameters for industry collaboration to promote the development of intelligent remote sensing technology and applications.

SkySense was jointly developed by Ant’s AI innovation research department NextEvo and Wuhan University. NextEvo is Ant’s core AI technology research and development team, leading the research and development of the BaiLing large model. Its research directions include computer vision, natural language processing, multimodal, AI governance and compliance, digital humans, AI engineering, and other core technologies. Last year, the department upgraded its multimodal team, with Yang Ming leading the comprehensive layout of multimodal technology.

SEE ALSO: Jiang Lu, Head of Google’s VideoPoet Project, Joined TikTok