FeatureExtractionResult

spatialhub.structures.FeatureExtractionResult contains dense patch tokens or global CLS embedding vectors produced by feature extraction adapters such as DINOv2Adapter.

from spatialhub.structures import FeatureExtractionResult

---

Fields

Field	Type	Shape	Description
images	np.ndarray	(N, H, W, 3) uint8	Preprocessed input image batch.
features	np.ndarray	(N, D) or (N, C, H, W) float32	Extracted feature embedding tensor (\(D=1024\) for ViT-L/14).
embedding_type	str	"global"	Embedding resolution: "global" (single CLS token per image) or "dense" (patch tokens).
l2_normalized	bool	False	Boolean indicating whether embedding vectors satisfy \(\lVert v \rVert_2 = 1.0\).

---

Usage Example

from spatialhub import DINOv2

extractor = DINOv2(model_variant="dinov2_vitl14")
result = extractor.extract_features(["frame_a.png", "frame_b.png"])

print(f"Features shape: {result.features.shape}")  # (2, 1024)
print(f"L2 Normalized: {result.l2_normalized}")    # True

# Compute pairwise cosine similarity
similarity = result.features[0] @ result.features[1].T
print(f"Cosine Similarity: {similarity:.4f}")