Link precomputed SLAT targets from dataset card
Browse files
README.md
CHANGED
|
@@ -60,8 +60,8 @@ The completed extent dataset retains 19,408 training/evaluation views;
|
|
| 60 |
592 source captures were rejected by target preparation.
|
| 61 |
|
| 62 |
Archive inventories and SHA-256 checksums are in `bundles/<dataset>/index.json`.
|
| 63 |
-
|
| 64 |
-
|
| 65 |
|
| 66 |
## Restore on another server
|
| 67 |
|
|
@@ -103,7 +103,8 @@ python -m slat_segmentation.train --config slat_segmentation/configs/front3d_sag
|
|
| 103 |
Original manifests, artifact completion markers, checksums, identities and split
|
| 104 |
policies are retained byte-for-byte inside the archives. Generation settings and
|
| 105 |
code hashes are also available under `metadata/<dataset>/conversion_policy.json`.
|
| 106 |
-
|
|
|
|
| 107 |
Source datasets retain their original licenses and access terms.
|
| 108 |
|
| 109 |
## Input and geometry previews
|
|
@@ -158,3 +159,20 @@ View: `003c09039641a2b2efa1`.
|
|
| 158 |
|
| 159 |
[Posed prompted GT geometry (GLB)](previews/ycbv/test-geometry.glb)
|
| 160 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 60 |
592 source captures were rejected by target preparation.
|
| 61 |
|
| 62 |
Archive inventories and SHA-256 checksums are in `bundles/<dataset>/index.json`.
|
| 63 |
+
The extent archives exclude diagnostic subsets, raw downloads, model weights, and model-specific feature caches.
|
| 64 |
+
FoundationPose currently has raw archives only.
|
| 65 |
|
| 66 |
## Restore on another server
|
| 67 |
|
|
|
|
| 103 |
Original manifests, artifact completion markers, checksums, identities and split
|
| 104 |
policies are retained byte-for-byte inside the archives. Generation settings and
|
| 105 |
code hashes are also available under `metadata/<dataset>/conversion_policy.json`.
|
| 106 |
+
The extent upload does not include model weights.
|
| 107 |
+
SLAT target inventories record the frozen encoder provenance.
|
| 108 |
Source datasets retain their original licenses and access terms.
|
| 109 |
|
| 110 |
## Input and geometry previews
|
|
|
|
| 159 |
|
| 160 |
[Posed prompted GT geometry (GLB)](previews/ycbv/test-geometry.glb)
|
| 161 |
|
| 162 |
+
## Precomputed SLAT-shape targets
|
| 163 |
+
|
| 164 |
+
The [SLAT target snapshots](precomputed_slat/README.md) contain available 32³ shape latents for both training datasets.
|
| 165 |
+
The frozen TRELLIS.2 shape encoder produced these latents.
|
| 166 |
+
|
| 167 |
+
| Dataset | Training views | Encoded objects | Archive GiB |
|
| 168 |
+
| --- | ---: | ---: | ---: |
|
| 169 |
+
| 3D-FRONT | 39,930 | 240,952 | 5.19 |
|
| 170 |
+
| SAGE-10k | 17,456 | 397,203 | 1.18 |
|
| 171 |
+
|
| 172 |
+
Three SAGE-10k objects retain encoding failures.
|
| 173 |
+
Scene encoding remains partial.
|
| 174 |
+
Frozen DINOv3 and NAF feature caches still require preparation.
|
| 175 |
+
The snapshots preserve the recorded completion states.
|
| 176 |
+
|
| 177 |
+
The [SLAT instructions](precomputed_slat/README.md#restore-the-caches) describe restoration with `--bundle-root precomputed_slat`.
|
| 178 |
+
The extent archives remain under `bundles/`.
|