Reproduce “Implicit Occupancy Flow Fields for Perception and Prediction in Self-Driving” on OpenResearch

About this paper

Ben Agro, Quinlan Sykora, Sergio Casas, Raquel Urtasun

IMPLICITO introduces an implicit neural network for unified perception and prediction in self-driving, modeling continuous spatio-temporal occupancy and motion flow. This approach allows on-demand querying, outperforming existing object-based and explicit grid methods in accuracy, calibration, and computational efficiency for motion planners on urban and highway datasets.

Key insights

• Implicit representations enable a more efficient and expressive approach to spatio-temporal perception and prediction by allowing continuous querying and avoiding redundant computations of dense grids.
• An attention mechanism with learned offsets can effectively expand the receptive field of the model, allowing it to retrieve relevant information from distant spatial and temporal locations crucial for accurate future prediction.
• Unified, end-to-end training of perception and prediction, directly from raw sensor data, leads to superior performance and better calibrated predictions compared to decoupled or object-based pipelines.

Summary by alphaXiv.