AInnovation Ranks First in PASCAL VOC, the World’s Authoritative Object Detection Challenge

Recently, PASCAL VOC, the world’s authoritative public evaluation datasets for object detection, released its latest rankings. “AInnoDetection”, the algorithm developed by AInnovation surpassed the performance of many well-known AI enterprises and university laboratories in comp4, including Google, Microsoft, Carnegie Mellon, Tsinghua University, Alibaba, Sogou and Tencent. Across the twenty evaluation criteria, AInnovDetection ranked first in ten criteria, reflecting the sophistication of AInnovation’s AI algorithm.

Object detection is one of three basic tasks in computer vision, and it’s also a challenging problem in the field of computer vision. Pascal VOC Challenge is one of the three most important computer vision challenges in the world. The datasets have high quality annotations, complex scenarios and diverse targets; and therefore, it is a prime choice for fast testing algorithms. Pascal VOC does not have big datasets, but it has rich scenes, therefore it tests for the design and innovation ability of AI.

Pascal VOC is the main venue for AI companies to compete, and competition is fierce. It attracts nearly 100 professional teams from across the world, including Google, Microsoft, Carnegie Mellon, Tsinghua University, Alibaba, Ping An Technology, Sogou and Yi+.

Pascal VOC datasets have 20 classes, including humans, animals, vehicles and indoor objects. AInnovation’s computer vision team participated in the competion4 subtask. Across all twenty classes of the competition, AInnoDetection ranked first in ten classes, and ranked first overall by total score.

The AInnoDetection Algorithm

The AInnoDetection model proposed by the team is based on the classic two-stage detection pipeline, with data augmentation including pasting small objects and mix-up methods to enhance the performance when detecting small and obscured objects, respectively. Also, the data augmentation makes the network more robust.

The famous two-stage detection network was implemented, the backbone of ResNext152 for multi-level feature extraction, neck of 6-level FPN structure and head of 3x cascade RCNN, each of them had two subnets, one for classification with focal loss and another for bounding box regression with smooth L1 loss. Meanwhile, no matter training or inference, multi-scale images are taken, forcing the network to focus and learn how to recognize objects of different scales.

Read More: Decentralized AI Alliance Reaches 50 Members Worldwide

Finally, the algorithm adopted a multiple model fusion strategy with NMS and obtained the best results.

Commercial Value of the Object Detection Algorithm

Since its founding, AInnovation has strived to develop commercial products using high-quality AI algorithms for the retail, manufacturing, finance and other industries. It focuses on the advancement and maturity of AI algorithms, and provides various commercial AI products and solutions. As a result, the company has developed at a rapid and steady pace. At present, AInnovation has built an industry-leading machine learning platform and AI industry vision platform named ManuVision. Furthermore, it has produced several Top-level Conference Papers which improve the accuracy of the algorithms in the business scenarios, improve the speed of model training, and integrate computing resources efficiently.

AInnovation’s visual algorithm has been applied in the field of commodity identification, intelligent container, industrial vision and smart community. The commercial effect is at the leading level in the industry. The object detection algorithm adopted in PASCAL VOC competition has already been applied in scenarios such as product detection, industrial defect detection, and channel display monitoring. AInnovation has won other championships at several of the world’s top artificial intelligence algorithm competitions; for example, most recently, it placed first at the famous WIDER FACE.

Read More: HIPAA-Compliant Geospatial Health Analytics – Now Available for Tableau