Hangzhou, China/ April 26, 2019 Dahua Technology, a world-leading video-centric smart IoT solution and service provider, recently took the 1st place in the KITTI Semantic Segmentation Evaluation and broke the world record based on its image semantic segmentation technology of deep learning algorithm, surpassing other top-notch AI companies and leading academic research institutions, which marks the leading position of Dahua Technology in the field of semantic segmentation.
Dahua Technology has established large-scale computing centers and data centers for algorithm training, focusing on the research and commercialization of multiple algorithms domains, and has formed its core competitiveness. In 2017, Dahua Technology achieved the first place in the field of Scene Flow, Optical Flow as well as Text Recognition Detection respectively. In 2018, Dahua Technology won the first place in the 2D Vehicle Object Detection Evaluation, MOT Tracking, and Pedestrian Re-identification. In the early 2019, Dahua Technology won the first place in the international competition of Case Segmentation Evaluation. This time in the field of Semantic Segmentation Algorithm, Dahua Technology once again made new breakthroughs.
Established by Karlsruhe Institute of Technology and Toyota Technological Institute at Chicago, KITTI dataset is one of the largest international computer vision algorithm benchmark datasets. Datasets are used to evaluate the performance of Stereo, Optical Flow, Visual Odometry, Object Detection and Tracking, Road, Semantics and other computer vision technologies in the vehicular environment. KITTI contains real images collected from a variety of scenes such as urban areas, rural areas, highways, etc. Each image contains up to 15 vehicles and 30 pedestrians, with varying degrees of occlusion and truncation.
Semantic segmentation refers to the process of classifying each pixel of a computer's input image to the object category. It is not only the basic task of computer vision, but also plays a vital role in the application of autonomous driving, robot scene understanding and virtual reality.
In the KITTI semantic segmentation task, 19 types of objects such as automobiles, pedestrians, roads, motorcycles, bicycles, traffic signs, buildings and vegetation in various scenes need to be accurately segmented. At the same time, the training set provides only 200 pieces of data, which is small sample learning.
In this evaluation, Dahua Technology incorporated the advantages of image classification, single target segmentation, full pixel semantic segmentation and other advanced algorithms to improve the accuracy of semantic segmentation and constructed a global attention mechanism based on multi-position and channel features. Besides, migration learning and incremental learning methods were also adopted to effectively improve the segmentation accuracy of the algorithm.