NeuroGeom: Revised Research Proposal
Revised proposal for submission to CVPR 2027 / ICCV 2027 / NeurIPS 2027.
See neurogeom_revised_proposal.md for the full document.
Key Changes from v1
- Reframed novelty around GAT refinement + topology learning (not primitive codebook learning)
- Added DNMP and DBW as primary comparisons; corrected Neural-Pull characterization
- Fixed view-direction conditioning in decoder
- Specified differentiable codebook assignment via Gumbel-Softmax
- Softmax-weighted density composition to avoid bleeding
- Extended Gumbel temperature annealing schedule
- Three-phase training to coordinate instance params and GNN learning rates
- Zero-mean deformation constraint
- CARLA promoted to primary evaluation for structural metrics
- SSCBench methods as semantic baselines instead of MonoScene alone
- Cross-dataset generalization in main results
- Honest rendering quality and speed assessment
- Bounded GNN corrections to prevent overpowering direct optimization
- Added 12 ablation studies