cubec commited on
Commit
d13e577
·
verified ·
1 Parent(s): 073988c

Finalize verified FoundationPose release and pinned restoration guide

Browse files
README.md CHANGED
@@ -61,6 +61,7 @@ artifacts are in the archives.
61
  | sage | True | True | 475.80 | 475.80 |
62
  | 3dfront | True | True | 273.53 | 273.53 |
63
  | ycbv | True | True | 3.60 | 3.60 |
 
64
 
65
  SAGE-10k keeps every room and view of a source layout in the same split,
66
  using the saved SHA-256 layout split with a 10% test fraction and no validation split.
@@ -69,14 +70,16 @@ The completed extent dataset retains 19,408 training/evaluation views;
69
 
70
  Archive inventories and SHA-256 checksums are in `bundles/<dataset>/index.json`.
71
  The extent archives exclude diagnostic subsets, raw downloads, model weights, and model-specific feature caches.
72
- FoundationPose archive publication is in progress. The complete inventory appears after verification of every remote archive.
73
 
74
 
75
  ## FoundationPose on another server
76
 
 
77
  The FoundationPose bundle preserves original RGB, depth, visible masks, cameras, object mapping, and complete per-object voxel occupancy at 64³.
78
  Its source manifest has SHA-256 `2c4dcd1cd1e2e56a0eb08f2f92da320f299729c8ce7a2f52828c272cbc1fa0ad`.
79
  The extracted files occupy 87.82 GB before filesystem overhead.
 
80
  The export excludes raw archives, meshes, model weights, image features, SLAT latents, and GT occlusion relations.
81
 
82
  The loaders resolve input paths relative to the restored dataset.
@@ -98,7 +101,8 @@ Run these commands from the repository root.
98
 
99
  ```bash
100
  python -m preprocess.huggingface download \
101
- --repo cubec/PartProbe3D-preprocessed --dataset foundationpose --cleanup
 
102
  ```
103
 
104
  3. Initialize the pinned Pixal3D source.
@@ -151,9 +155,11 @@ DINOv3 requires approved access to its [gated model repository](https://huggingf
151
 
152
  ### Prepare features and train SS
153
 
154
- The [FoundationPose configuration](https://github.com/cube-c/PartProbe3D/blob/slat-shape/ss_multiobj_depth/configs/foundationpose.json) uses GT depth, `depth_embedding` voxel conditioning, global cross-attention, and rank-8 NoRA.
155
  It creates features under `outputs/ss_multiobj_depth/foundationpose/features` and preserves the saved training/test splits.
156
  It starts a new run with W&B online by default.
 
 
157
 
158
  1. Prepare local features and encoded SS targets.
159
 
@@ -170,6 +176,7 @@ It starts a new run with W&B online by default.
170
  This configuration needs no source archives, meshes, or additional dataset downloads.
171
  The `occlusion_object` variant additionally requires a matching GT occlusion cache.
172
  That cache requires mesh geometry during preparation and is absent from this FoundationPose bundle.
 
173
  New SLAT shape encoding and mesh-based evaluation also require geometry or separately transferred targets.
174
  Exact resume of an existing run requires its matching checkpoint, feature manifests, source hashes, and settings.
175
 
@@ -190,6 +197,8 @@ These previews show saved inputs and GT voxel targets. They use no learned check
190
 
191
  [GT occupancy, test (GLB)](previews/foundationpose/test-occupancy.glb)
192
 
 
 
193
  ## GT occlusion relations
194
 
195
  Separate archives under `precomputed_occlusion/` preserve the completed GT occlusion caches.
 
61
  | sage | True | True | 475.80 | 475.80 |
62
  | 3dfront | True | True | 273.53 | 273.53 |
63
  | ycbv | True | True | 3.60 | 3.60 |
64
+ | foundationpose | True | True | 83.19 | 83.19 |
65
 
66
  SAGE-10k keeps every room and view of a source layout in the same split,
67
  using the saved SHA-256 layout split with a 10% test fraction and no validation split.
 
70
 
71
  Archive inventories and SHA-256 checksums are in `bundles/<dataset>/index.json`.
72
  The extent archives exclude diagnostic subsets, raw downloads, model weights, and model-specific feature caches.
73
+ The FoundationPose export contains 46,948 views in 86 verified archives.
74
 
75
 
76
  ## FoundationPose on another server
77
 
78
+ Use [code revision `92ca3ee`](https://github.com/cube-c/PartProbe3D/commit/92ca3eea1dbec6e162e246858eb7bdb7f7fd0701) or a compatible later revision on `slat-shape`.
79
  The FoundationPose bundle preserves original RGB, depth, visible masks, cameras, object mapping, and complete per-object voxel occupancy at 64³.
80
  Its source manifest has SHA-256 `2c4dcd1cd1e2e56a0eb08f2f92da320f299729c8ce7a2f52828c272cbc1fa0ad`.
81
  The extracted files occupy 87.82 GB before filesystem overhead.
82
+ The [verified inventory](https://huggingface.co/datasets/cubec/PartProbe3D-preprocessed/blob/073988c5b5e206570da67f57aada4a2f0ad94e95/bundles/foundationpose/index.json) contains 86 archives totaling 89.33 GB.
83
  The export excludes raw archives, meshes, model weights, image features, SLAT latents, and GT occlusion relations.
84
 
85
  The loaders resolve input paths relative to the restored dataset.
 
101
 
102
  ```bash
103
  python -m preprocess.huggingface download \
104
+ --repo cubec/PartProbe3D-preprocessed --dataset foundationpose --cleanup \
105
+ --revision 073988c5b5e206570da67f57aada4a2f0ad94e95
106
  ```
107
 
108
  3. Initialize the pinned Pixal3D source.
 
155
 
156
  ### Prepare features and train SS
157
 
158
+ The [FoundationPose configuration](https://github.com/cube-c/PartProbe3D/blob/92ca3eea1dbec6e162e246858eb7bdb7f7fd0701/ss_multiobj_depth/configs/foundationpose.json) uses GT depth, `depth_embedding` voxel conditioning, global cross-attention, and rank-8 NoRA.
159
  It creates features under `outputs/ss_multiobj_depth/foundationpose/features` and preserves the saved training/test splits.
160
  It starts a new run with W&B online by default.
161
+ Online logging requires W&B authentication.
162
+ The setting `training.wandb_mode="disabled"` permits a local run without W&B.
163
 
164
  1. Prepare local features and encoded SS targets.
165
 
 
176
  This configuration needs no source archives, meshes, or additional dataset downloads.
177
  The `occlusion_object` variant additionally requires a matching GT occlusion cache.
178
  That cache requires mesh geometry during preparation and is absent from this FoundationPose bundle.
179
+ A matching GT occlusion overlay can transfer separately without another download of the observation archives.
180
  New SLAT shape encoding and mesh-based evaluation also require geometry or separately transferred targets.
181
  Exact resume of an existing run requires its matching checkpoint, feature manifests, source hashes, and settings.
182
 
 
197
 
198
  [GT occupancy, test (GLB)](previews/foundationpose/test-occupancy.glb)
199
 
200
+ The [publication report](https://github.com/cube-c/PartProbe3D/blob/139c539/reports/huggingface_foundationpose_training_20261019.md) records checks, dataset identities, and limitations.
201
+
202
  ## GT occlusion relations
203
 
204
  Separate archives under `precomputed_occlusion/` preserve the completed GT occlusion caches.
catalog.json CHANGED
@@ -80,11 +80,12 @@
80
  }
81
  },
82
  "source_manifest_sha256": "2c4dcd1cd1e2e56a0eb08f2f92da320f299729c8ce7a2f52828c272cbc1fa0ad",
83
- "packaged": false,
84
- "bytes": 33531084800,
85
- "uploaded_bytes": 68556800,
86
- "uploaded_archives": 1,
87
- "uploaded": false
 
88
  }
89
  ]
90
  }
 
80
  }
81
  },
82
  "source_manifest_sha256": "2c4dcd1cd1e2e56a0eb08f2f92da320f299729c8ce7a2f52828c272cbc1fa0ad",
83
+ "packaged": true,
84
+ "bytes": 89326489600,
85
+ "uploaded_bytes": 89326489600,
86
+ "uploaded_archives": 86,
87
+ "uploaded": true,
88
+ "revision": "073988c5b5e206570da67f57aada4a2f0ad94e95"
89
  }
90
  ]
91
  }
metadata/foundationpose/publication.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "complete": true,
3
+ "dataset_revision": "073988c5b5e206570da67f57aada4a2f0ad94e95",
4
+ "workflow_commit": "92ca3eea1dbec6e162e246858eb7bdb7f7fd0701",
5
+ "documentation_commit": "139c539ec3a04bac62115409ea563975cbb6c017",
6
+ "archives": 86,
7
+ "bytes": 89326489600,
8
+ "views": 46948,
9
+ "occlusion_included": false,
10
+ "feature_caches_included": false,
11
+ "geometry_included": false,
12
+ "verified_utc": "2026-10-19T11:27:35.589394+00:00"
13
+ }
metadata/foundationpose/verification.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "complete": true,
3
+ "completed_utc": "2026-10-19T11:27:35.589394+00:00",
4
+ "repo": "cubec/PartProbe3D-preprocessed",
5
+ "revision": "073988c5b5e206570da67f57aada4a2f0ad94e95",
6
+ "archives": 86,
7
+ "bytes": 89326489600,
8
+ "views": 46948,
9
+ "counts": {
10
+ "test": {
11
+ "views": 4692,
12
+ "scenes": 4692
13
+ },
14
+ "train": {
15
+ "views": 42256,
16
+ "scenes": 42256
17
+ }
18
+ },
19
+ "artifact_kind": "extent-training",
20
+ "geometry_included": false,
21
+ "feature_caches_included": false,
22
+ "occlusion_included": false,
23
+ "source_manifest_sha256": "2c4dcd1cd1e2e56a0eb08f2f92da320f299729c8ce7a2f52828c272cbc1fa0ad",
24
+ "index_sha256": "357f8245629cb7b0c0384f3352f084f7dc6c24eed7c3cd146a1c985a0544974b",
25
+ "archive_members": 1919212,
26
+ "all_remote_sizes_match": true,
27
+ "all_remote_sha256_match": true,
28
+ "remote_inventory_matches_local_bytes": true,
29
+ "workflow_commit": "92ca3eea1dbec6e162e246858eb7bdb7f7fd0701",
30
+ "portability_preparation_views": 2,
31
+ "hf_roundtrip_archives": 1
32
+ }