Recently, CVPR 2023, one of the top three global conferences in the field of computer vision, was held in Vancouver, Canada. The conference announced the research achievements of multiple projects and the results of the participating teams. The DeepBlue team achieved excellent results of "one championship, three runner-ups, and two third places" in the six projects they participated in. Since 2019, the DeepBlue team has won the championship in the CVPR challenge for five consecutive years.

CVPR is an authoritative academic conference in the field of computer vision, organized by the Institute of Electrical and Electronics Engineers (IEEE). It is known as the "Oscars" of the computer vision field and attracts top companies, research institutions, and universities from the industry worldwide to participate. It is a stage for global AI research teams to demonstrate their scientific and technological innovation and test their fundamental research achievements.

2023 VizWiz Grand Challenge

The 2023 VizWiz Grand Challenge Workshop is part of the 2023 CVPR workshop. The purpose of this workshop is to enable researchers to understand the technological needs of visually impaired individuals and authorize researchers to improve algorithms to meet these needs. All the data comes from photos taken by visually impaired individuals and the questions they raise. The photos have problems such as blurriness and out-of-focus.

To address these issues, the DeepBlue team chose state-of-the-art multimodal models, improved the task settings for model training, and made fine adjustments to the parameters. As a result, they achieved excellent results of "one championship, one runner-up, and one third place" in the competition.

CVPR 2023 AVA

The Accessibility Vision and Autonomy (AVA) Challenge is part of the CVPR 2023 Workshop on Accessible Vision and Autonomy. It aims to bring together researchers, students, and advocates from the fields of accessibility, computer science, and self-service systems. The goal of this competition is to use generated data to solve problems in the sharing and principled development of data-driven visual accessibility systems.

The challenge covers two tracks: instance segmentation and keypoint detection. They share the same dataset, which is synthetically generated by a rendering engine. The data mainly presents challenges such as significant domain differences between natural and generated data, small and weak targets, and long-tailed distributions.

In the instance segmentation track, the DeepBlue team selected appropriate model algorithms in a very limited time and used effective model training methods.

In the keypoint detection track, the DeepBlue team combined the instance segmentation model from track one to effectively detect pedestrian targets and further perform keypoint detection on pedestrians. This saves the additional time required to train a pedestrian detector. In the end, they achieved excellent results of second place and third place in the two tracks of the competition, respectively.

Plant Traits 2023

Plant Traits 2023 is a branch of the CVPR 2023 FGCV Fine-Grained Visual Categorization Challenge. The challenge requires participants to use deep learning regression models to predict plant attributes (e.g., length, germination rate, and more than 30 other attributes) from plant photos.

The DeepBlue team chose to use climate data as auxiliary features and the original images as the main features for regression prediction. They adjusted the data distribution, increased the regression loss, and achieved the second-place result through model ensembling.

So far, the DeepBlueAI team has won more than 40 championships in various international AI competitions. Among them, they have accumulated more than 20 championships in the top three computer vision events: CVPR, ECCV, and ICCV. The outstanding performance of the DeepBlue team at CVPR 2023 not only fully demonstrates their solid strength in AI theoretical research breakthroughs and technological innovation to their global peers but also expands the application space of computer vision technology in various industries throughout society. It has laid a solid theoretical foundation and technical reserves for the future widespread application of AI technology in fields such as autonomous driving, urban governance, robotics, and industrial intelligence.