Post-training Model Quantization Using GANs for Synthetic Data Generation

Quantization is a widely adopted technique for deep neural networks to reduce the memory and computational resources required. However, when quantized, most models would need a suitable calibration process to keep their performance intact, which requires data from the target domain, such as a fraction of the dataset used in model training and model validation (i.e. calibration dataset). In this study, we investigate the use of synthetic data as a substitute for the calibration with real data for the quantization method. We propose a data generation method based on Generative Adversarial Networks that are trained prior to the model quantization step. We compare the performance of models quantized using data generated by StyleGAN2-ADA and our pre-trained DiStyleGAN, with quantization using real data and an alternative data generation method based on fractal images. Overall, the results of our experiments demonstrate the potential of leveraging synthetic data for calibration during the quantization process. In our experiments, the percentage of accuracy degradation of the selected models was less than 0.6%, with our best performance achieved on MobileNetV2 (0.05%). The code is available at: https://github.com/ThanosM97/gsoc2022-openvino

LandCover.ai: Dataset for Automatic Mapping of Buildings, Woodlands and Water from Aerial Imagery

Monitoring of land cover and land use is crucial in natural resources management. Automatic visual mapping can carry enormous economic value for agriculture, forestry, or public administration. Satellite or aerial images combined with computer vision and deep learning enable the precise assessment and can significantly speed up the process of change detection. Aerial imagery usually provides images with much higher pixel resolution than satellite data allowing more detailed mapping. However, there is still a lack of datasets that were made for the segmentation of buildings with other highly publicly important environmental instances like woods or water. Here we introduce LandCover.ai (Land Cover from Aerial Imagery) dataset that propose semantic segmentation. We collected images of 216.27 sq. km lands across Poland, a country in Central Europe, 39.51 sq. km with resolution 50 cm per pixel and 176.76 sq. km with resolution 25 cm per pixel and manually fine annotated three following classes of objects: buildings, woodlands, and water. Additionally, we report simple benchmark results, achieving 86.2% of mean intersection over union on the test set. It proves that the automatic mapping of land cover is possible and can be applied in various domains. The dataset is publicly available at http://landcover.ai