Towards generalizable and interpretable three-dimensional tracking with inverse neural rendering

Today, the most successful methods for image-understanding tasks rely on feed-forward neural networks. Although this approach offers empirical accuracy, efficiency and task adaptation through fine-tuning, it also comes with fundamental disadvantages. Existing networks often struggle to generalize across different datasets, even on the same task. By design, these networks ultimately reason about high-dimensional scene features, which are challenging to analyse. This is true especially when attempting to predict three-dimensional (3D) information based on two-dimensional images. We propose to recast vision problems with RGB inputs as an inverse rendering problem by optimizing through a differentiable rendering pipeline over the latent space of pretrained 3D object representations and retrieving latents that best represent object instances in a given input image. Specifically, we solve the task of 3D multi-object tracking by optimizing an image loss over generative latent spaces that inherently disentangle shape and appearance properties. Not only do we investigate an alternative take on tracking but our method also enables us to examine the generated objects, reason about failure situations and resolve ambiguous cases. We validate the generalization and scaling capabilities of our method by learning the generative prior exclusively from synthetic data and assessing camera-based 3D tracking on two large-scale autonomous robot datasets. Both datasets are completely unseen to our method and do not require fine-tuning.