| misc: | |
| exp_dir: ocfnet_geo | |
| mode: train | |
| gpu_mode: True | |
| verbose: True | |
| verbose_freq: 500 | |
| snapshot_freq: 1 | |
| pretrain: '' | |
| model: | |
| model: OCFNet | |
| in_feats_dim: 1 | |
| out_feats_dim: 32 | |
| conv1_kernel_size: 5 | |
| normalize_feature: True | |
| dist_type: L2 | |
| overlap_attention_module: | |
| attention_feats_dim: 256 | |
| num_head: 4 | |
| nets: ['self','cross','self','cross','self','cross'] # GeoTransformer / CoFiNet: 3 interleaved rounds | |
| # 'self' layers are masked multi-head self-attention (torch's fused attention) with a 3D | |
| # rotary position embedding on the voxel coordinates; 'cross' layers get no position | |
| # embedding, the two clouds are not registered yet | |
| # RoPE variant: positions enter the attention as a rotary embedding of the voxel | |
| # coordinates (relative, applied to query and key) instead of the paper's geometric | |
| # encoding added to the features. Set both to True to keep the two together. | |
| rope: True | |
| rope_theta: 100.0 | |
| geometric_pos_enc: False # CoFiNet ape: False -- no positional encoding at the coarse level | |
| pos_enc_k: 5 # neighbours of the geometry-aware positional encoding (Eq. 1) | |
| coarse_to_fine: | |
| sinkhorn_iters: 100 # CoFiNet uses 100; at 20 the plan is still 1e-1 from its marginals | |
| similarity: inner # CoFiNet Eq.(1): einsum(F_X, F_Y)/sqrt(d), unnormalised | |
| sinkhorn_epsilon: 0.05 | |
| coarse_threshold: 0.05 # tau_t of Eq. (3) | |
| # A patch is the Voronoi cell of its super-point -- disjoint, so a point belongs to | |
| # exactly one patch -- capped at `patch_size` members. | |
| use_overlap_head: True # head trained everywhere; only the marginals differ | |
| coarse_marginals: uniform # CoFiNet's plain OT; 'overlap' is Eq. (5) | |
| fine_marginals: uniform | |
| patch_size: 64 | |
| max_coarse_matches: 256 # patches refined per training step | |
| # the overlap head: the paper normalises before the sigmoid, which fixes its mean | |
| # at 0.5 while the targets average `overlap_prior` (scripts/diagnose_ocfnet.py) | |
| overlap_head_norm: True | |
| overlap_prior: 0.08 | |
| # let the two patches of a match talk to each other before matching, so the | |
| # transport marginal means "matchable against this patch" rather than | |
| # "matchable somewhere" -- 74% of a patch's points have their partner elsewhere | |
| patch_attention: True | |
| readout_mutual: True # keep only mutually-agreed matches | |
| readout_deduplicate: True # and one per source point: 0.356 -> 0.526 inlier ratio | |
| min_coarse_matches: 128 # at inference, fall back to the top-k pairs if the | |
| # threshold selects fewer than this | |
| eval_pairs: 300 # benchmark pairs registered after every epoch | |
| matching_radius: 0.1 # CoFiNet's pos_margin; the radius the inlier ratio uses | |
| loss: | |
| # L = L_C + L_F + L_CO + L_FO (section 2.4) | |
| w_coarse_loss: 1.0 | |
| w_coarse_circle_loss: 1.0 # GeoTransformer overlap-aware circle loss on the super-point features | |
| w_fine_loss: 1.0 | |
| w_inlier_loss: 1.0 # expected inlier ratio, maximised directly | |
| w_coarse_inlier_loss: 1.0 # the same, for the coarse transport | |
| w_coarse_overlap_loss: 1.0 | |
| w_fine_overlap_loss: 1.0 | |
| # ranking supervision the paper's losses lack: a circle loss on the descriptors | |
| # the transport cost is built from, and a row-wise classification inside each | |
| # patch, which trains the argmax that Eq. (6) reads out | |
| w_descriptor_loss: 1.0 | |
| w_patch_infonce_loss: 1.0 | |
| w_coarse_infonce_loss: 1.0 | |
| pos_margin: 0.1 | |
| neg_margin: 1.4 | |
| log_scale: 24 | |
| pos_radius: 0.0375 | |
| safe_radius: 0.1 | |
| overlap_radius: 0.0375 | |
| matchability_radius: 0.05 | |
| w_circle_loss: 1.0 | |
| w_overlap_loss: 1.0 | |
| w_saliency_loss: 0.0 | |
| max_points: 256 | |
| optimiser: | |
| # OCFNet optimises cross-entropy over transport plans, not Predator's circle loss, so the | |
| # SGD schedule of the Predator configs does not carry over. This follows CoFiNet, whose | |
| # loss is of the same family: Adam at 1e-4. | |
| optimizer: ADAM | |
| max_epoch: 150 # CoFiNet trains for 150 | |
| lr: 0.0003 | |
| momentum: 0.8 | |
| weight_decay: 0.000001 | |
| bn_momentum: 0.05 | |
| scheduler: ExpLR | |
| scheduler_gamma: 0.95 | |
| scheduler_freq: 1 | |
| iter_size: 1 | |
| dataset: | |
| dataset: indoor | |
| benchmark: 3DMatch | |
| root: data/indoor | |
| voxel_size: 0.025 | |
| # The attention is masked per pair and the transport plans are solved per pair, so the | |
| # pairs of a batch never mix (see scripts/test_batch_equivalence.py). | |
| # Memory grows with the voxel count, about 20 GiB per million voxels on this model | |
| # (scripts/bench_batch_size.py); 32 peaks around 18 GiB of the 63 GiB card, which leaves | |
| # room for a batch that happens to draw only large scenes. Throughput is already flat | |
| # here: 0.047 s/pair at 32 against 0.044 at 48. | |
| # Aligned with CoFiNet (configs/tdmatch/tdmatch.yaml), whose published numbers this | |
| # implementation is compared against: their effective batch is 4 (batch 1, iter_size 4), | |
| # so an epoch is 5160 optimiser steps against 645 at batch 32 -- eight times more learning | |
| # per epoch. Batch 8 recovers most of that while keeping the GPU busy. | |
| batch_size: 12 | |
| num_workers: 6 | |
| augment_noise: 0.005 | |
| train_info: configs/indoor/train_info.pkl | |
| val_info: configs/indoor/val_info.pkl | |
| demo: | |
| src_pcd: assets/cloud_bin_21.pth | |
| tgt_pcd: assets/cloud_bin_34.pth | |
| n_points: 1000 | |