v0.6.4: Chamfer Distance Numerical Stabilization & NaN Elimination

Release Notes - Conquer3D v0.6.4

Features

• Chamfer & Hausdorff Distance Numerical Stabilization & NaN Elimination:
  • Replaced raw thrust::device_vector system cudaMalloc allocations with PyTorch CUDACachingAllocator device tensors, preventing CUDA memory allocator aborts under GPU memory pressure.
  • Implemented dedicated single-reference CUDA kernel (one_sided_chamfer_single_point_kernel) computing direct $O(N)$ squared distances without KD-Tree construction overhead for $M=1$.
  • Added robust empty-set safeguards for $N=0$ and $M=0$, preventing __builtin_clz(0) undefined behavior and cudaErrorInvalidConfiguration kernel launch crashes.
  • Applied critically small $\epsilon = 10^-26$ numerical clamp (torch.clamp(distances, min=1e-26)) before torch.sqrt() to eliminate NaN artifacts from negative sub-normal floating-point drift.
  • Guarded priority queue insertion in kdtree::query_kdtree_loop against non-finite float coordinates.

Refactors & Build

• PyBind11 Dispatcher: Handled $N=0$ and $M=0$ gracefully in C++ PyBind wrapper returning valid pre-allocated empty/infinite tensors.
• Version Bump: Bumped version to 0.6.4 in pyproject.toml.

Examples

• Standalone verification suite in scratch/verify_chamfer_nan_fixes.py validating 100% numerical parity against PyTorch cdist across empty, single-point, identical, and large-scale random point clouds.