LetOccVote: Learning Weakly Supervised 3D Occupancy through Consensus

arXiv:2609.04846v1 proposes LetOccVote, a weakly supervised 3D occupancy prediction method based on consensus mechanisms. This method optimizes geometric and semantic supervision through cross-frame voting: it corrects pseudo-depth values and eliminates contradictory estimates through deep voting, and aggregates pseudo-semantic observations in shared 3D space through semantic voting to identify reliable evidence. LetOccVote is trained solely with 2D pseudo-labels, without the need for 3D occupancy annotations. On the Occ3D-nuScenes dataset, this method achieved a IoU of 53.27 and mIoU of 20.39, reaching the highest performance among similar methods.