VoxFormer

Why we included this project

VoxFormer tackles a hard 3D perception problem: rebuilding a complete, semantically labeled 3D volume from ordinary camera images alone, including occluded and invisible regions. Its two-stage design proposes sparse voxel queries from monocular depth first, then densifies them with self-attention, which keeps training GPU memory under 16GB while staying competitive on the SemanticKITTI scene completion benchmark. That makes it a solid reference for researchers and engineers working on camera-only occupancy prediction where lidar is unavailable. The repo includes full training and evaluation code, pretrained weights, and a deformable-attention variant, so reproducing the reported numbers or adapting the architecture to your own data is straightforward. One catch: the non-commercial NVIDIA license means this is for research, not for shipping directly into a commercial product.