---

name: scannet-semantic-scene-workflows description: "ScanNet semantic scene parsing workflows for the PointNet2 TensorFlow 1 repository: preprocessing, pickle layout, labels, training, and whole-scene evaluation." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

ScanNet Semantic Scene Workflows

Use this sub-skill when the task is about PointNet2 ScanNet semantic scene parsing: preparing ScanNet inputs, validating scannet_data_pointnet2 pickles, reasoning about label tables, constructing the legacy training command, or explaining whole-scene evaluation behavior.

Do not use this sub-skill for ModelNet classification or ShapeNetPart part segmentation except to mention that the same TensorFlow 1.x and PointNet++ custom-op constraints apply.

Read first

1. references/data-formats.md for the self-contained ScanNet pickle schema, raw preprocessing layout, class ids, and label-table rules.
2. references/workflows.md for preprocessing, training, random-block evaluation, virtual scans, and whole-scene evaluation recipes.
3. references/troubleshooting.md when a command, loader, label map, raw-data path, or whole-scene run fails.

Bundled helpers

• scripts/validate_scannet_layout.py: validates scannet_train.pickle / scannet_test.pickle, raw ScanNet scene prerequisites, generated preprocessing .npy files, optional demo outputs, and V1/V2 label-table column choices.
• scripts/smoke_scannet_loader.py: creates or loads a tiny ScanNet-style pickle fixture and exercises the adapted random-block and whole-scene loader behavior.
• scripts/build_scannet_command.py: emits shell commands for the legacy trainer and preprocessing/demo scripts without executing the original Python 2 code.

Operating rules

• Treat the original ScanNet trainer and preprocessing scripts as reference-only legacy scripts. They require Python 2 syntax, TensorFlow 1.x, large external datasets, and for PointNet++ model execution normally require compiled TensorFlow custom ops.
• Use the bundled validator before recommending a training run. The trainer expects data/scannet_data_pointnet2/scannet_train.pickle and scannet_test.pickle to contain two sequential pickle objects: a list of N x 3 XYZ arrays and a same-length list of N semantic-label arrays.
• Remember that ScanNet class id 0 is unannotated; the evaluation code excludes label 0 from accuracy metrics by checking label > 0 and positive sample weight.
• For ScanNetV2, do not reuse the V1 TSV column assumptions blindly. The repository note says the raw-class and NYU40-class columns are shifted by one relative to the V1 table; validate with explicit column overrides.
• Keep raw preprocessing and demo.py guidance path-aware: those scripts expect external ScanNet downloads, a scene list, and writable output directories. Validate missing raw-data paths and demo outputs independently.