Xidian News(By Yang Yuting, Wang Dan) On March 31, the organizing committee of 2022 IEEE GRASS Data Fusion Contest sent a congratulation mail to School of Artificial Intelligence of Xidian University for winning the championship in the official contest of the International Geoscience and Remote Sensing Symposium (IGARSS), a top conference in remote sensing. The champion team “Lu Xiaoqiang & Cao Guojin” will be invited to report their results at the 2022 IEEE International Conference on Geoscience and Remote Sensing Symposium (IGARSS 2022) to be held at the Kuala Lumpur Convention Center in Malaysia in July. This contest is funded by the National Natural Science Foundation of China and the Chinese Association for Artificial Intelligence-Huawei MindSpore Academic Awards.
In addition to the champion team, the other nine participating teams of the university all entered the top 30 of the ranking. Among them, three teams entered the top ten and six teams the top twenty. The participating teams are mainly composed of the first-year graduate students. The school adheres to the philosophy of “promoting learning and education by contests” to train junior graduate students, so as to enable them to learn the latest professional knowledge faster and better from the contest, and improve their ability in scientific research and professional practice.
For the multi-modal semi-supervised semantic segmentation dataset MF-DFC22 proposed in this 2022 IEEE GARSS Data Fusion Contest, the champion team presented an Adaptive Pixel-Rebalancing Self-Training (APRST) algorithm based on active learning. APRST uses SegFormer-B0 and SeMask-FPN-Swin-B as the basic segmenters respectively. The training process is divided into two stages: 1. using a small amount of labeled data for fully supervised training, and after the convergence of the model, initializing it to a teacher model and a student model respectively; 2. using the teacher model to make predictions on unlabeled data.
The Semi-Supervised Learning Challenge of the 2022 IEEE GARSS Data Fusion Contest is organized by the Image Analysis and Data Fusion Technical Committee (IADF TC) of the IEEE GARSS, University of Bretagne-Sud, ONERA and ESA Φ-lab. It aims to promote the innovation of automatic land cover classification and to label training data only from the portion consisting of VHR RGB images.
MiniFrance-DFC22 (MF-DFC22) is the dataset used in the contest, which is an extension and modification of the MiniFrance dataset for semi-supervised semantic segmentation. The multimodal MF-DFC22 data contains aerial images, elevation models and land use (land cover) maps of 19 large cities in France and their surrounding areas, which embrace urban and rural scenes -- residential, industrial and commercial areas, fields, forests, coasts and low mountains.