From 385be2eca72eb3833efa4be0a0088b34e764788a Mon Sep 17 00:00:00 2001
From: shuishen <1109946754@qq.com>
Date: Mon, 31 Aug 2026 09:04:45 +0800
Subject: [PATCH] feat(pointcloud): complete annotation and result lifecycle workflows

---
 capabilities/05-3d-pointcloud/apply_pointcloud_semantic_model.py |  100 ++++++++++++++++++++++++++++++--------------------
 1 files changed, 60 insertions(+), 40 deletions(-)

diff --git a/capabilities/05-3d-pointcloud/apply_pointcloud_semantic_model.py b/capabilities/05-3d-pointcloud/apply_pointcloud_semantic_model.py
index f50cd21..8a1ff26 100644
--- a/capabilities/05-3d-pointcloud/apply_pointcloud_semantic_model.py
+++ b/capabilities/05-3d-pointcloud/apply_pointcloud_semantic_model.py
@@ -25,9 +25,6 @@
 from train_pointcloud_semantic_model import CLASS_SCHEMA, PointWiseNet
 
 
-PREVIEW_POINT_LIMIT = 400_000
-
-
 def sha256(path: Path) -> str:
     digest = hashlib.sha256()
     with path.open("rb") as stream:
@@ -79,44 +76,30 @@
     return xyz, np.clip(rgb, 0.0, 1.0), None
 
 
-def preview_indices(predictions: np.ndarray, limit: int) -> np.ndarray:
-    """Deterministically retain every small predicted class before filling the budget."""
-    count = len(predictions)
-    if count <= limit:
-        return np.arange(count, dtype=np.int64)
-    rng = np.random.default_rng(42)
-    selected: list[np.ndarray] = []
-    remaining = limit
-    groups = sorted((np.flatnonzero(predictions == code) for code in np.unique(predictions)), key=len)
-    for group in groups:
-        keep = min(len(group), max(1, min(50_000, remaining // max(1, len(groups) - len(selected)))))
-        selected.append(group if keep == len(group) else np.sort(rng.choice(group, size=keep, replace=False)))
-        remaining -= keep
-    chosen = np.concatenate(selected)
-    if len(chosen) < limit:
-        mask = np.ones(count, dtype=bool)
-        mask[chosen] = False
-        fill = rng.choice(np.flatnonzero(mask), size=limit - len(chosen), replace=False)
-        chosen = np.concatenate((chosen, fill))
-    if len(chosen) > limit:
-        chosen = rng.choice(chosen, size=limit, replace=False)
-    return np.sort(chosen.astype(np.int64))
-
-
-def load_model(path: Path, device: torch.device) -> tuple[PointWiseNet, list[int]]:
+def load_model(path: Path, device: torch.device) -> tuple[PointWiseNet, list[int], dict[int, dict[str, Any]]]:
     payload = torch.load(path, map_location=device, weights_only=False)
-    if not isinstance(payload, dict) or payload.get("schema_version") != 1:
+    if not isinstance(payload, dict) or payload.get("schema_version") not in {1, 2}:
         raise ValueError("Unsupported model schema.")
     codes = payload.get("class_codes")
-    if not isinstance(codes, list) or len(codes) < 2 or any(not isinstance(code, int) or code not in CLASS_SCHEMA for code in codes):
+    if not isinstance(codes, list) or len(codes) < 2 or any(not isinstance(code, int) or not 1 <= code <= 255 for code in codes):
         raise ValueError("Model class schema is invalid.")
+    saved_classes = payload.get("classes")
+    classes: dict[int, dict[str, Any]] = {}
+    for code in codes:
+        raw = saved_classes.get(str(code)) if isinstance(saved_classes, dict) else CLASS_SCHEMA.get(code)
+        if not isinstance(raw, dict):
+            raise ValueError("Model is missing class metadata for a predicted class.")
+        key, label, color = raw.get("key"), raw.get("label"), raw.get("color")
+        if not isinstance(key, str) or not isinstance(label, str) or not isinstance(color, list) or len(color) != 3 or not all(isinstance(item, int) and 0 <= item <= 255 for item in color):
+            raise ValueError("Model class metadata is invalid.")
+        classes[code] = {"code": code, "key": key, "label": label, "color": color}
     model = PointWiseNet(len(codes)).to(device)
     try:
         model.load_state_dict(payload["state_dict"], strict=True)
     except (KeyError, RuntimeError) as exc:
         raise ValueError("Model weights do not match the supported PointWiseNet architecture.") from exc
     model.eval()
-    return model, codes
+    return model, codes, classes
 
 
 def main() -> int:
@@ -126,11 +109,15 @@
     parser.add_argument("--output", type=Path, required=True)
     parser.add_argument("--device", choices={"auto", "cpu", "cuda"}, default="auto")
     parser.add_argument("--batch-size", type=int, default=4096)
+    parser.add_argument("--annotation-source-id", type=str, default="")
+    parser.add_argument("--candidate-confidence", type=float, default=0.95)
     args = parser.parse_args()
     if args.device == "cuda" and not torch.cuda.is_available():
         raise SystemExit("CUDA was requested but is unavailable.")
     if args.batch_size < 1 or args.batch_size > 262_144:
         raise SystemExit("Batch size must be between 1 and 262144.")
+    if not 0.5 <= args.candidate_confidence < 1.0:
+        raise SystemExit("candidate-confidence must be between 0.5 and 1.0.")
     if not args.model.is_file() or not args.input.is_file():
         raise SystemExit("Model or input point cloud is unavailable.")
     if args.output.exists() and any(args.output.iterdir()):
@@ -139,24 +126,29 @@
     started = time.perf_counter()
     device_name = "cuda" if args.device == "cuda" or (args.device == "auto" and torch.cuda.is_available()) else "cpu"
     device = torch.device(device_name)
-    model, class_codes = load_model(args.model, device)
+    model, class_codes, class_schema = load_model(args.model, device)
     xyz, rgb, source_las = load_cloud(args.input)
     center = xyz.mean(axis=0)
     scale = float(max(np.abs(xyz - center).max(), 1e-6))
     features = np.column_stack(((xyz - center) / scale, rgb)).astype(np.float32)
-    pieces: list[np.ndarray] = []
+    prediction_pieces: list[np.ndarray] = []
+    confidence_pieces: list[np.ndarray] = []
     with torch.no_grad():
         for start in range(0, len(features), args.batch_size):
             values = torch.from_numpy(features[start:start + args.batch_size]).to(device)
-            pieces.append(model(values).argmax(dim=1).cpu().numpy())
-    predictions = np.asarray([class_codes[index] for index in np.concatenate(pieces)], dtype=np.uint8)
+            probabilities = torch.softmax(model(values), dim=1)
+            confidence, predicted = probabilities.max(dim=1)
+            prediction_pieces.append(predicted.cpu().numpy())
+            confidence_pieces.append(confidence.cpu().numpy())
+    predictions = np.asarray([class_codes[index] for index in np.concatenate(prediction_pieces)], dtype=np.uint8)
+    confidence = np.concatenate(confidence_pieces).astype(np.float32, copy=False)
     counts = {int(code): int((predictions == code).sum()) for code in class_codes}
 
     args.output.mkdir(parents=True, exist_ok=True)
-    preview_selection = preview_indices(predictions, PREVIEW_POINT_LIMIT)
+    preview_selection = np.arange(len(predictions), dtype=np.int64)
     preview = o3d.geometry.PointCloud()
     preview.points = o3d.utility.Vector3dVector(xyz[preview_selection])
-    preview.colors = o3d.utility.Vector3dVector(np.asarray([CLASS_SCHEMA[int(code)]["color"] for code in predictions[preview_selection]], dtype=np.float64) / 255.0)
+    preview.colors = o3d.utility.Vector3dVector(np.asarray([class_schema[int(code)]["color"] for code in predictions[preview_selection]], dtype=np.float64) / 255.0)
     preview_path = args.output / "predicted-semantic-preview.ply"
     if not o3d.io.write_point_cloud(str(preview_path), preview, write_ascii=False):
         raise RuntimeError("Could not write predicted PLY preview.")
@@ -182,11 +174,39 @@
         writer = csv.DictWriter(stream, fieldnames=["class_code", "class_key", "class_label", "point_count"])
         writer.writeheader()
         for code in class_codes:
-            writer.writerow({"class_code": code, "class_key": CLASS_SCHEMA[code]["key"], "class_label": CLASS_SCHEMA[code]["label"], "point_count": counts[code]})
-    summary = {"class_codes": class_codes, "class_counts": {str(code): counts[code] for code in class_codes}, "input_points": int(len(xyz)), "preview_points": int(len(preview_selection)), "preview_sampling": "deterministic class-aware cap; smaller predicted classes retained before the remaining budget is sampled", "input_has_rgb": True}
+            writer.writerow({"class_code": code, "class_key": class_schema[code]["key"], "class_label": class_schema[code]["label"], "point_count": counts[code]})
+    summary = {"class_codes": class_codes, "class_counts": {str(code): counts[code] for code in class_codes}, "input_points": int(len(xyz)), "preview_points": int(len(preview_selection)), "preview_sampling": "complete prediction point set without display sampling", "input_has_rgb": True}
+    candidate_path: Path | None = None
+    if args.annotation_source_id:
+        selected = np.flatnonzero(confidence >= args.candidate_confidence)
+        candidate_counts = {
+            int(code): int(np.count_nonzero(predictions[selected] == code))
+            for code in class_codes
+        }
+        candidate_path = args.output / "automatic-annotation-candidates.json"
+        candidate_path.write_text(json.dumps({
+            "schema_version": 1,
+            "source_id": args.annotation_source_id,
+            "source_sha256": sha256(args.input),
+            "model_sha256": sha256(args.model),
+            "candidate_confidence": args.candidate_confidence,
+            "input_points": int(len(xyz)),
+            "candidate_count": int(len(selected)),
+            "candidate_class_counts": {str(code): candidate_counts[code] for code in class_codes},
+            "labels": [[int(index), int(predictions[index]), round(float(confidence[index]), 6)] for index in selected],
+        }, ensure_ascii=False), encoding="utf-8")
+        summary["automatic_annotation"] = {
+            "candidate_confidence": args.candidate_confidence,
+            "candidate_count": int(len(selected)),
+            "candidate_class_counts": {str(code): candidate_counts[code] for code in class_codes},
+            "candidate_file": candidate_path.name,
+        }
     summary_path = args.output / "prediction-summary.json"
     summary_path.write_text(json.dumps(summary, ensure_ascii=False, indent=2), encoding="utf-8")
-    metadata = {"capability": "05-3d-pointcloud", "classification": "B", "created_at": datetime.now(UTC).isoformat(), "model": {"path": str(args.model), "sha256": sha256(args.model), "architecture": "PointWiseNet shared MLP"}, "input": {"path": str(args.input), "sha256": sha256(args.input), "bytes": args.input.stat().st_size, "points": int(len(xyz)), "has_rgb": True}, "classes": {str(code): CLASS_SCHEMA[code] for code in class_codes}, "prediction": summary, "processing": {"requested_device": args.device, "device": device_name, "batch_size": args.batch_size, "normalization": {"method": "source-local per input", "xyz_center": center.tolist(), "xyz_scale": scale}}, "versions": {"python": sys.version.split()[0], "torch": torch.__version__, "open3d": o3d.__version__, "laspy": laspy.__version__}, "artifacts": {"preview": preview_path.name, "classified_las": classified_las.name, "class_counts": csv_path.name, "summary": summary_path.name}, "elapsed_seconds": round(time.perf_counter() - started, 3), "limitations": ["Predictions are model candidates, not asset inventory or inspection conclusions.", "This model requires observed RGB; it cannot infer labels for XYZ-only point clouds.", "Model metrics apply only to the labelled source spatial blocks. The current pole/tower class has high false-positive risk and requires review.", "New inputs are normalized with their own XYZ centre and scale to match the training feature definition; this preserves their coordinates but does not prove cross-site generalization."]}
+    artifacts = {"preview": preview_path.name, "classified_las": classified_las.name, "class_counts": csv_path.name, "summary": summary_path.name}
+    if candidate_path:
+        artifacts["automatic_candidates"] = candidate_path.name
+    metadata = {"capability": "05-3d-pointcloud", "classification": "B", "created_at": datetime.now(UTC).isoformat(), "model": {"path": str(args.model), "sha256": sha256(args.model), "architecture": "PointWiseNet shared MLP"}, "input": {"path": str(args.input), "sha256": sha256(args.input), "bytes": args.input.stat().st_size, "points": int(len(xyz)), "has_rgb": True}, "classes": {str(code): class_schema[code] for code in class_codes}, "prediction": summary, "processing": {"requested_device": args.device, "device": device_name, "batch_size": args.batch_size, "normalization": {"method": "source-local per input", "xyz_center": center.tolist(), "xyz_scale": scale}}, "versions": {"python": sys.version.split()[0], "torch": torch.__version__, "open3d": o3d.__version__, "laspy": laspy.__version__}, "artifacts": artifacts, "elapsed_seconds": round(time.perf_counter() - started, 3), "limitations": ["Predictions are model candidates, not asset inventory or inspection conclusions.", "Automatic candidates retain only points at or above the recorded confidence threshold; user confirmation is required before merging them into a training revision.", "This model requires observed RGB; it cannot infer labels for XYZ-only point clouds.", "Model metrics apply only to the labelled source spatial blocks. The current pole/tower class has high false-positive risk and requires review.", "New inputs are normalized with their own XYZ centre and scale to match the training feature definition; this preserves their coordinates but does not prove cross-site generalization."]}
     (args.output / "run_metadata.json").write_text(json.dumps(metadata, ensure_ascii=False, indent=2), encoding="utf-8")
     print(json.dumps(metadata, ensure_ascii=False))
     return 0

--
Gitblit v1.9.3