| # 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 |
| 1. Reframed novelty around GAT refinement + topology learning (not primitive codebook learning) |
| 2. Added DNMP and DBW as primary comparisons; corrected Neural-Pull characterization |
| 3. Fixed view-direction conditioning in decoder |
| 4. Specified differentiable codebook assignment via Gumbel-Softmax |
| 5. Softmax-weighted density composition to avoid bleeding |
| 6. Extended Gumbel temperature annealing schedule |
| 7. Three-phase training to coordinate instance params and GNN learning rates |
| 8. Zero-mean deformation constraint |
| 9. CARLA promoted to primary evaluation for structural metrics |
| 10. SSCBench methods as semantic baselines instead of MonoScene alone |
| 11. Cross-dataset generalization in main results |
| 12. Honest rendering quality and speed assessment |
| 13. Bounded GNN corrections to prevent overpowering direct optimization |
| 14. Added 12 ablation studies |
|
|
