Dense Semantic Forecasting with Multi-Level Feature Warping

Anticipation of per-pixel semantics in a future unobserved frame is also known as dense semantic forecasting. State-of-the-art methods are based on single-level regression subsampled abstract representation recognition model. However, cannot account for skip connections from the backbone to upsampling path. We propose address this shortcoming by warping shallow features observed images with upsampled feature flow. Our goal not straightforward, since coarse flow introduces noise into forecasted features. therefore base our work single-frame models that more resistant connections. To achieve this, we training procedure enables operate reasonably well or without Validation experiments reveal interesting insights influence particular accuracy. forecasting method delivers 70.2% mIoU 0.18 s and 58.5% 0.54 future. These show 0.6 points improved accuracy respect baseline promising directions work.