Add prepared soybean PointTransformer dataset (67 train, 11 test)
Browse files
point_transformer/README.md
ADDED
|
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Prepared soybean PointTransformer data
|
| 2 |
+
|
| 3 |
+
`soybean.tar.gz` contains 78 prepared samples: **67 train / 11 test**, no
|
| 4 |
+
validation set, with 76,362,794 points. The archive is 2,239,573,703 bytes;
|
| 5 |
+
extracted files occupy 4,276,590,506 bytes.
|
| 6 |
+
|
| 7 |
+
Derived from the segmented soybean point clouds in `sample_params.tar.gz`
|
| 8 |
+
in TianhangCheng7/DemeterData using Demeter's
|
| 9 |
+
`script_point_transformer/prepare_data.py`. This is a reproducible split for
|
| 10 |
+
the preparation recipe, not a claim about the original paper's partition.
|
| 11 |
+
Paired `_i` / `_o` scans stay in the same split.
|
| 12 |
+
|
| 13 |
+
From a clean Demeter checkout:
|
| 14 |
+
|
| 15 |
+
```bash
|
| 16 |
+
hf download TianhangCheng7/DemeterData point_transformer/soybean.tar.gz point_transformer/soybean.tar.gz.sha256 --repo-type dataset --local-dir outputs/downloads
|
| 17 |
+
(cd outputs/downloads/point_transformer && sha256sum -c soybean.tar.gz.sha256)
|
| 18 |
+
tar -xzf outputs/downloads/point_transformer/soybean.tar.gz
|
| 19 |
+
```
|
| 20 |
+
|
| 21 |
+
This restores `data/point_transformer/soybean/{train,test}/*.pth` and
|
| 22 |
+
`manifest.json`. Existing data should be moved aside before extraction.
|
| 23 |
+
The manifest preserves coordinate transforms, thresholds, labels, split
|
| 24 |
+
lists and source-path provenance. Source paths need not exist on the new machine.
|
| 25 |
+
|
| 26 |
+
Each sample stores `coord`, `color`, `normal`, `semantic_gt5`, `instance_gt`,
|
| 27 |
+
`inv_dists`, and `scene_id`. Load with
|
| 28 |
+
`torch.load(path, map_location="cpu", weights_only=True)`.
|
| 29 |
+
Coordinates are centered, normalized and aligned with the stem along +X;
|
| 30 |
+
RGB is in [0, 1]. Semantic labels are 0 leaf, 1 other stem, 2 main stem,
|
| 31 |
+
3 flower, 4 fruit.
|
| 32 |
+
|
| 33 |
+
`inv_dists = T / max(d, T)`, where `d` is distance to the nearest point in
|
| 34 |
+
another organ instance and `T` is twice the mean within-instance tenth-neighbor
|
| 35 |
+
distance (including the point itself). See the repository's
|
| 36 |
+
`script_point_transformer/readme.md` for full preprocessing details and
|
| 37 |
+
training commands. Local CPU data checks passed; no training accuracy is claimed.
|
point_transformer/soybean.tar.gz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f8c34f2464a65721bbe195c86077cf12f914af4f8a0ab9e35bd7b5f9ac990a59
|
| 3 |
+
size 2239573703
|
point_transformer/soybean.tar.gz.sha256
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:894404e6c74c0884c419880a8f4a15b76bac6844c300c7f974eb6549a469005e
|
| 3 |
+
size 82
|