Training camera-based 3D semantic occupancy models typically requires 3D labels. A recent line of work avoids this by rendering predictions into 2D images and supervising with pseudo labels. The arXiv preprint 'Where Do the Gaussians Go? Foreground Attraction for Camera-Based Weakly Supervised 3D Occupancy' examines what happens when this approach is applied to sparse Gaussian representations.
According to the abstract, such models suffer from a problem: the Gaussians are not placed correctly, and the paper identifies a phenomenon it calls foreground attraction. The proposed method appears to counteract this by adjusting how the model attends to foreground regions, though the abstract does not provide implementation details or quantitative results.
Because the source is a single preprint with a truncated abstract, this summary is limited to the problem statement and the proposed direction. No comparison with other methods or experimental outcomes is available in the provided text.