RDIR: Capturing Temporally-Invariant Representations of Multiple Objects in Videos

Piotr Zielinski, Tomasz Kajdanowicz

Abstract: Learning temporally coherent representations of multiple objects in videos is crucial for understanding their complex dynamics and interactions over time. In this paper, we present a deep generative neural network, which can learn such representations by leveraging pretraining. Our model builds upon a scale-invariant structured autoencoder, extending it with a convolutional recurrent module to refine the learned representations through time and enable information sharing among multiple cells in multi-scale grids. This novel approach provides a framework for learning perobject representations from a pretrained object detection model, offering the ability to infer predefined types of objects, without the need for supervision. Through a series of experiments on benchmark datasets and real-life video footage, we demonstrate the spatial and temporal coherence of the learned representations, showcasing their applicability in downstream tasks such as object tracking. We analyze the method's robustness by conducting an ablation study, and we compare it to other methods, highlighting the importance of the quality of objects' representations.