Skip to main content
Click any dataset in the Dataset Inventory to open the Dataset Exploration View, your primary workspace for analyzing that dataset. Once inside, you have tools for search, filtering, cluster navigation, and quality analysis that work together in a systematic exploration workflow.
Dataset Exploration View

​
Interface Layout

Once you are inside a dataset, the exploration interface loads with five main areas:
Three tabs organize your workflow when working with a dataset:

​
Exploration Workflow

Effective dataset curation typically involves a multi-step process rather than a single query. These steps explain how to combine Visual Layer’s tools to move from broad discovery to a precise, curated selection.
1

Find: Discover Relevant Content

Start by casting a wide net to locate potential candidates.
  • Use Semantic Search for conceptual queries (e.g., “damaged packaging” or “outdoor crowd”).
  • Use Cluster Navigation to browse high-level patterns without a specific query.
  • Use Visual Search with an external image upload to find matches for a reference asset.
Goal: Surface a broad set of results that includes your target data, accepting some initial false positives.
2

Narrow: Refine Your Results

Filter out irrelevant matches to focus on your specific criteria.
  • Apply Visual Similarity on a specific search result. Crop a region (e.g., just the defect, not the background) to narrow the search visually.
  • Add Metadata Filters (Folders, Labels, Dates) to restrict the search scope to relevant batches or sources.
Goal: Reduce noise and focus on content that matches your specific visual or metadata requirements.
3

Refine: Identify Distinctive Content

Ensure your selection represents the diversity of your data, not just the most common examples.
  • Apply the Uniques Filter to hide repetitive content. Increasing the threshold surfaces only visually distinct items, covering edge cases and varied angles rather than just the average.
  • Check Outliers to find rare examples or anomalies that might be missing from the main clusters.
Goal: Create a diverse selection that includes valuable edge cases.
4

Clean: Remove Redundancy and Issues

Polish the collection before finalizing.
  • Review Duplicates to select single representative frames from burst sequences or near-identical backups.
  • Filter Quality Issues (blur, dark, bright) to automatically exclude unusable or low-quality assets.
Goal: Ensure the final dataset is lean, high-quality, and free of technical flaws.
5

Organize: Save and Share

Preserve your work for the team.
  • Save as View to capture the full combination of search queries, filters, and thresholds. Team members can access the dynamic collection without rebuilding the logic.
  • Export the selection to JSON or CSV for downstream training or annotation workflows.

How to Search & Filter

Step-by-step guide to running searches and filters, plus the full list of available options.

How Search Works

Concepts behind semantic search and visual similarity

Visual Search

Find similar images using visual embeddings

Understanding Clusters

How Visual Layer organizes images