Voxel51
Similarity seek for high quality management
As soon as you discover one problematic annotation, similarity search turns into a robust instrument to seek out all associated errors. Click on on a mislabeled pattern and immediately retrieve essentially the most comparable pictures to examine if they’ve the identical systematic labeling downside.
FiftyOne’s similarity search transforms “discover extra like this” from guide tedium into on the spot discovery. Index your information set as soon as, then immediately retrieve visually comparable samples via point-and-click or programmatic queries.
import fiftyone as fo
import fiftyone.mind as fob
import fiftyone.zoo as foz
# Load dataset
dataset = foz.load_zoo_dataset("quickstart")
# Index pictures by similarity
fob.compute_similarity(
dataset,
mannequin="clip-vit-base32-torch",
brain_key="img_sim"
)
# Type by most certainly to comprise annotation errors
mistake_view = dataset.sort_by("mistakenness", reverse=True)
# Question the primary pattern and discover 10 most comparable pictures
query_id = mistake_view.take(1).first().id
similar_view = dataset.sort_by_similarity(query_id, ok=10, brain_key="img_sim")
# Launch App to view comparable samples and for point-and-click similarity search
session = fo.launch_app(dataset)
Key capabilities embody on the spot visible search via the App interface, object-level similarity indexing for detection patches, and scalable again ends that change from sklearn to Qdrant, Pinecone, or different vector databases for manufacturing.
