Dahua Technology’s ARI 2D object detection method recently took the 1st place in the KITTI Vision Benchmark Suite’s Object Detection Evaluation 2012 on July 26th, 2018 – with an accuracy of 91.48% based on a moderate difficulty level.
The KITTI vision benchmark suite is a systematic benchmarking platform designed to evaluate computer vision performance. Funded by the Karlsruhe Institute of Technology and the Toyota Technological Institute (KITTI) in Chicago, it is probably the world’s first and largest benchmarking suite for vision based autonomous driving. KITTI includes real life images collected from a variety of scenery, from urban streets in the mid-size city of Karlsruhe to rurals roads and highways. Each image contains sophisticated scenarios involving at most 15 vehicles and 30 pedestrians with varying levels of overlapping. The KITTI vision Benchmark suite comprises of real-world benchmarks for stereo, optical flow, visual odometry, object detection and tracking.
This competition has provided Dahua with an excellent opportunity to further its independent research and development of deep learning algorithms. Based upon the advantages of network structures such as Resnet, Dahua Technology has successfully improved the structures of its deep learning detection algorithm. Utilising reinforcement learning and other training techniques, as well as multi-model fusion technology, Dahua Technology has made a significant improvement in the detection rate of small and/or overlapped targets. 2D object detection algorithms are able to realise target detection in videos. They then capture the targets after classifying them. 2D object detection is widely applied in the company’s newly launched intelligent products, especially the cameras, NVRs and servers based upon AI deep-learning
