Feature flow: In-network feature flow estimation for video object detection

• A type of shallow modules are proposed to directly predict the feature flow for alignment in a single network. Self-supervision learning is introduced further improve quality predicted flow. new state-of-the-art performance shown by comparing with other methods, while fast inference speed maintained. Optical flow, which expresses pixel displacement, widely used many computer vision tasks provide pixel-level motion information. However, remarkable progress convolutional neural network, recent approaches solve problems on feature-level. Since displacement vector not consistent common approach forward optical network and fine-tune this task dataset. With method, they expect fine-tuned produce tensors encoding feature-level In paper, we rethink about de facto paradigm analyze its drawbacks video object detection task. To mitigate these issues, propose novel (IFF-Net) an I n-network F eature low estimation module (IFF module) detection. Without resorting pre-training any additional dataset, our IFF able indicates displacement. Our consists module, shares features branches. This compact design enables IFF-Net accurately detect objects, maintaining speed. Furthermore, transformation residual loss (TRL) based self-supervision , improves IFF-Net. outperforms existing methods achieves ImageNet VID.