Add illustration.png, pcam, requirements.txt, setup.py

pcam/configs/3dmatch/soft.yaml

@@ -0,0 +1,3 @@
+DATASET: '3dmatch'
+NUM_POINTS: 4096
+SPARSE_ATTENTION: False

pcam/configs/3dmatch/soft_filter.yaml

@@ -0,0 +1,5 @@
+DATASET: '3dmatch'
+NUM_POINTS: 4096
+SPARSE_ATTENTION: False
+# This threshold was optimised on the validation of 3dmatch
+THRESHOLD: .6

pcam/configs/3dmatch/soft_refinement.yaml

@@ -0,0 +1,6 @@
+DATASET: '3dmatch'
+NUM_POINTS: 4096
+SPARSE_ATTENTION: False
+# This threshold was optimised on the validation of 3dmatch
+THRESHOLD: .6
+DGR_OPTIM: True

pcam/configs/3dmatch/soft_safeguard.yaml

@@ -0,0 +1,10 @@
+DATASET: '3dmatch'
+NUM_POINTS: 4096
+SPARSE_ATTENTION: False
+# This threshold was optimised on the validation of 3dmatch
+THRESHOLD: .6
+DGR_OPTIM: True
+# Safeguard registration (as in DGR) when the confidence scores are below a threshold (WSUM_THRESHOLD)
+# This threshold was optimised to apply safeguard on the same number of scans as in DGR
+SAFEGUARD: True
+WSUM_THRESHOLD: 930

pcam/configs/3dmatch/sparse.yaml

@@ -0,0 +1,3 @@
+DATASET: '3dmatch'
+NUM_POINTS: 4096
+SPARSE_ATTENTION: True

pcam/configs/3dmatch/sparse_filter.yaml

@@ -0,0 +1,5 @@
+DATASET: '3dmatch'
+NUM_POINTS: 4096
+SPARSE_ATTENTION: True
+# This threshold was optimised on the validation of 3dmatch
+THRESHOLD: .5

pcam/configs/3dmatch/sparse_refinement.yaml

@@ -0,0 +1,6 @@
+DATASET: '3dmatch'
+NUM_POINTS: 4096
+SPARSE_ATTENTION: True
+# This threshold was optimised on the validation of 3dmatch
+THRESHOLD: .5
+DGR_OPTIM: True

pcam/configs/3dmatch/sparse_safeguard.yaml

@@ -0,0 +1,10 @@
+DATASET: '3dmatch'
+NUM_POINTS: 4096
+SPARSE_ATTENTION: True
+# This threshold was optimised on the validation of 3dmatch
+THRESHOLD: .5
+DGR_OPTIM: True
+# Safeguard registration (as in DGR) when the confidence scores are below a threshold (WSUM_THRESHOLD)
+# This threshold was optimised to apply safeguard on the same number of scans as in DGR
+SAFEGUARD: True
+WSUM_THRESHOLD: 930

pcam/configs/kitti/soft.yaml

@@ -0,0 +1,3 @@
+DATASET: 'kitti'
+NUM_POINTS: 2048
+SPARSE_ATTENTION: False

pcam/configs/kitti/soft_icp.yaml

@@ -0,0 +1,4 @@
+DATASET: 'kitti'
+NUM_POINTS: 2048
+SPARSE_ATTENTION: False
+ICP: .6

pcam/configs/kitti/sparse.yaml

@@ -0,0 +1,3 @@
+DATASET: 'kitti'
+NUM_POINTS: 2048
+SPARSE_ATTENTION: True

pcam/configs/kitti/sparse_icp.yaml

@@ -0,0 +1,4 @@
+DATASET: 'kitti'
+NUM_POINTS: 2048
+SPARSE_ATTENTION: True
+ICP: .6

pcam/configs/modelnet/soft.yaml

@@ -0,0 +1,3 @@
+DATASET: 'modelnet'
+NUM_POINTS: 2048
+SPARSE_ATTENTION: False

pcam/configs/modelnet/soft_noise.yaml

@@ -0,0 +1,3 @@
+DATASET: 'modelnet_noise'
+NUM_POINTS: 2048
+SPARSE_ATTENTION: False

pcam/configs/modelnet/soft_unseen.yaml

@@ -0,0 +1,3 @@
+DATASET: 'modelnet_unseen'
+NUM_POINTS: 2048
+SPARSE_ATTENTION: False

pcam/configs/modelnet/sparse.yaml

@@ -0,0 +1,3 @@
+DATASET: 'modelnet'
+NUM_POINTS: 2048
+SPARSE_ATTENTION: True

pcam/configs/modelnet/sparse_noise.yaml

@@ -0,0 +1,3 @@
+DATASET: 'modelnet_noise'
+NUM_POINTS: 2048
+SPARSE_ATTENTION: True

pcam/configs/modelnet/sparse_unseen.yaml

@@ -0,0 +1,3 @@
+DATASET: 'modelnet_unseen'
+NUM_POINTS: 2048
+SPARSE_ATTENTION: True

pcam/data/3dmatch/3dmatch.md

@@ -0,0 +1,13 @@
+# Dataset: 3DMatch
+
+## Folder structure
+
+Place the 3DMatch dataset in this folder (`/path/to/pcam/data/3dmatch/`), which should contain the following subfolders:
+```
+./threedmatch/            # Contains the training and validation set
+./threedmatch/*.npz
+./threedmatch/*.txt
+./threedmatch_test/       # Contains the test set
+./threedmatch_test/*.npz
+./threedmatch_test/*.txt
+```

pcam/data/kitti/kitti.md

@@ -0,0 +1,14 @@
+# Dataset: KITTI
+
+## Folder structure
+
+Place the KITTI dataset in this folder (`/path/to/pcam/data/kitti/`), which should contain the following subfolders:
+```
+./dataset/
+./dataset/poses/
+./dataset/poses/*.txt
+./dataset/sequences/
+./dataset/sequences/00/
+./dataset/sequences/01/
+etc.
+```

pcam/data/modelnet/modelnet.md

@@ -0,0 +1,12 @@
+# Dataset: ModelNet
+
+## Folder structure
+
+Place the ModelNet dataset in this folder (`/path/to/pcam/data/modelnet/`), which should contain the following subfolders:
+```
+./modelnet40_ply_hdf5_2048/
+./modelnet40_ply_hdf5_2048/ply_data_test0.h5
+./modelnet40_ply_hdf5_2048/ply_data_test1.h5
+./modelnet40_ply_hdf5_2048/ply_data_test_0_id2file.json
+etc.
+```

pcam/datasets/collate.py

@@ -0,0 +1,68 @@
+# Part of the code in this file is taken from https://github.com/chrischoy/DeepGlobalRegistration/blob/46dd264580b4191accedc277f4ae434acdb4d380/dataloader/base_loader.py#L24
+
+from pcam.tool.transforms import decompose_rotation_translation
+import lightconvpoint.nn as lcp_nn
+import torch
+import numpy as np
+
+class CollateFunc:
+  def __init__(self):
+    self.collation_fn = self.collate_pair_fn
+
+  def __call__(self, list_data):
+    return self.collation_fn(list_data)
+
+  def collate_pair_fn(self, list_data):
+    N = len(list_data)
+
+    list_data = [data for data in list_data if data is not None]
+    if N != len(list_data):
+      logging.info(f"Retain {len(list_data)} from {N} data.")
+    if len(list_data) == 0:
+      raise ValueError('No data in the batch')
+
+    xyz0_full, xyz1_full, xyz0, xyz1, trans, inv_trans, one_one_attention, filename, p0_mean, p1_mean = list(
+      zip(*list_data)
+    )
+
+    trans_batch = torch.from_numpy(np.stack(trans)).float()
+    inv_trans_batch = torch.from_numpy(np.stack(inv_trans)).float()
+
+    pts1 = torch.from_numpy(np.stack(xyz0)).float()
+    pts2 = torch.from_numpy(np.stack(xyz1)).float()
+    raw_pts1 = torch.from_numpy(np.stack(xyz0_full)).float()
+    raw_pts2 = torch.from_numpy(np.stack(xyz1_full)).float()
+
+    p0_mean = torch.from_numpy(np.stack(p0_mean)).float().transpose(1, 2)
+    p1_mean = torch.from_numpy(np.stack(p1_mean)).float().transpose(1, 2)
+
+    Rs, ts = decompose_rotation_translation(trans_batch)
+    Rs_inv, ts_inv = decompose_rotation_translation(inv_trans_batch)
+
+    ts = ts.unsqueeze(-1)
+    ts_inv = ts_inv.unsqueeze(-1)
+
+    search = lcp_nn.SearchQuantized(K=32, stride=1)
+    pts1 = pts1.transpose(1, 2)
+    pts2 = pts2.transpose(1, 2)
+    indices1, pts1 = search(pts1)
+    indices2, pts2 = search(pts2)
+
+    one_one_attention = torch.from_numpy(np.stack(one_one_attention)).float()
+
+    return {
+      "raw_src": raw_pts1,
+      "raw_tgt": raw_pts2,
+      "src": pts1,
+      "tgt": pts2,
+      "filename": filename,
+      "attention": one_one_attention,
+      "indices1": indices1,
+      "indices2": indices2,
+      "rotation": Rs,
+      "translation": ts,
+      "inv_rotation": Rs_inv,
+      "inv_translation": ts_inv,
+      "p0_mean": p0_mean,
+      "p1_mean": p1_mean,
+    }