ReefNetAI 0.4.0 Beta: More confident review, segmentation, and recovery

ReefNetAI 0.4.0 Beta is now available. This release focuses on making day-to-day annotation work clearer, safer, and easier to recover. Classification and segmentation now share a stronger review model, the segmentation workspace adds new editing tools and major performance improvements, and CoralNet imports provide more honest progress and recovery controls.

A clearer annotation and review workflow

Classification points and segmentation masks now use the same explicit outcomes: Confirm, Correct, Skip, and Uncertain. Review reasons explain why an item needs attention, including uncertainty, low confidence, quality concerns, specialist input, or a review-every-item policy.

• Review Queue links open the exact point or mask that needs attention.
• Independent-review policies prevent submitters from approving their own work when a second reviewer is required.
• Completion and reviewed-export eligibility are checked consistently by the platform.
• Compact status indicators keep pending, completed, unresolved, mandatory, and advisory work easy to scan.

Assistive Memory you can inspect and manage

A new job-level Assistive Memory view shows recent learning examples and their delivery state. Users can distinguish waiting, pending, confirmed, retrying, unavailable, and removed entries, so the interface never claims that a correction was learned before it was acknowledged. Authorized users can remove individual entries, while owners can preview and confirm a full reset.

Safer model choices and clearer evaluation

Job setup now uses task-focused language and safer defaults. New standalone jobs begin with General models, while an eligible Source model becomes selectable only after it has been trained, reviewed, and promoted. Model evaluation remains separate from annotation vocabulary, and selecting a model never silently rewrites existing predictions.

A major segmentation workspace upgrade

Version 0.4.0 Beta expands the segmentation toolkit with Polygon Edit, tolerance-based Magic Wand selection, and an on-demand Cleanup panel. Saved masks can now be edited from the preview dialog, while improved contour handling preserves smoother boundaries and holes throughout editing, saving, refinement, and cutout workflows.

• Create, delete, add, subtract, and marker changes now participate in undo and redo.
• Preview tools provide consistent working space, retain the current view, and unwind nested actions predictably with Escape.
• Unsaved masks can be recovered after a browser crash or closed tab when the server data is still compatible.
• Image-level edit protection prevents two participants from unknowingly changing the same segmentation image at once.

Faster work on high-resolution imagery

Mask saving, overlap checks, hover previews, cutouts, and automatic-mask geometry were redesigned to avoid unnecessary full-image processing. In one measured 53-megapixel example, saving a batch of 25 masks reduced peak memory from about 960 MB to 94 MB and processing time from about 15 seconds to 3 seconds.

More reliable CoralNet imports

CoralNet import runs now report their current stage, processed and total counts, and recent activity. Users can pause, resume, cancel, or restart eligible runs, inspect progress and failure details, and retry from the part of the product that owns the operation. Interrupted or abandoned work is handled more explicitly, and an import that cannot recover reports that outcome instead of returning to an ambiguous state.

Better tools for structured studies

For teams running structured annotation studies, facilitators can now create sessions, manage study blocks, record pauses and technical interruptions, complete sessions, and export results. Participant actions survive page reloads, timing excludes recorded pauses and interruptions, and repeated submission attempts do not create duplicate accepted results.

ReefNetAI 0.4.0 Beta is about confidence: confidence in what needs review, what the system has learned, what will be saved, and how interrupted work can continue.

Try 0.4.0 Beta

Explore the updated workflows and let us know how they perform with real monitoring data. You can read the documentation for guidance or join the ReefNetAI Discord to share feedback with the team.