LO-World: Unified BEV Latent Modeling for Joint 4D Occupancy and LiDAR Forecasting
The authors report a 10.9% improvement in occupancy forecasting IoU and a 30.1% reduction in LiDAR Chamfer Distance over prior state-of-the-art baselines.
LO-World forecasts semantic occupancy and LiDAR point clouds in a shared bird’s-eye-view latent space, allowing semantic and geometric information to reinforce each other. Its single-stream design removes the need for separate forecasting pipelines. A hybrid range-voxel autoencoder works with a flow-matching generator that caches historical context to reduce sampling latency. The paper reports throughput above 25 frames per second on a single RTX 4090.
HF Daily Papers' note
LO-World forecasts semantic occupancy and LiDAR point clouds in a shared bird’s-eye-view latent space, allowing semantic and geometric information to reinforce each other. Its single-stream design removes the need for separate forecasting pipelines. A hybrid range-voxel autoencoder works with a flow-matching generator that caches historical context to reduce sampling latency. The paper reports throughput above 25 frames per second on a single RTX 4090.
HF Daily Papers' note
score 4