Semi-Supervised Hyperspectral Object Detection Challenge
Semi-supervised learning has developed into a highly researched problem as it minimizes the labeling costs while still achieving performance levels comparable to a fully labeled dataset. However, most semi-supervised learning algorithms are based on pre-trained models on ImageNet and are thus challenging to port to other image domains, especially those with more than three bands. In this competition, we present the application of a newly acquired dataset collected from a university rooftop with a hyperspectral camera to perform object detection. The camera produces images at a spatial resolution of 189-212 x 1600 pixels with 371 spectral bands, which we downsample to 51 bands. Each image in the dataset consists of visible and near-infrared measurements captured over three days on a fixed viewpoint. We use data captured on morning day as the training set. The other two days collectively constitute the validation and test sets on the competition server.
For the task of semi-supervised learning, we use 10% of the data as the labeled data across three categories, vehicles, bus, and bike, and strive to improve hyperspectral representation learning by leveraging popular deep learning techniques.
For each image, we provide the spectral calibrated radiance image and a corresponding mask to the area of interest. A starter code (based on mmdetection) with evaluation criteria will be provided with the CodaLab server as the baseline.
The timeline of this competition is as follows:
Train set, and Validation set released
Validation server online
Test set released and Test server online
Test output submission deadline
Fact sheet submission deadline
Acknowledgements: This work is supported by the Dynamic Data Driven Applications Systems Program, Air Force Office of Scientific Research under Grants FA 9550-19-1-0021 and FA 9550-15-1-0444.