Disentangling random and cyclic effects in time-lapse sequences

Time-lapse image sequences offer visually compelling insights into dynamic processes that are too slow to observe in real time. However, playing a long time-lapse sequence back as video often results distracting flicker due random effects, such weather, well cyclic the day-night cycle. We introduce problem of disentangling way allows separate, after-the-fact control overall trends, and effects images, describe technique based on data-driven generative models achieves this goal. This enables us "re-render" ways would not be possible with input images alone. For example, we can stabilize focus plant growth over many months, under selectable, consistent weather. Our approach is Generative Adversarial Networks (GAN) conditioned time coordinate sequence. architecture training procedure designed so networks learn model variations, using GAN's latent space, disentangle trends variations by feeding conditioning label Fourier features specific frequencies. show our robust defects data, enabling amend some practical difficulties capturing sequences, temporary occlusions, uneven frame spacing, missing frames.