Changelog

Release notes for the Picsellia platform, most recent first.

v3.100

Added

• Chart on Deployment predictions reach the Shape of a Prediction (Label, text, Attribute values, area) and the metrics of its Review (Average Precision, Average Recall, True Positive, False Positive, and False Negative counts, shadow review included), with filtering and grouping on both.
• A Shape copied in the Annotation tool stays in the clipboard when you move to another Asset, and can be pasted there. Pasting remains unavailable on an Annotation you are not allowed to edit.
• The Simplify action already available on polygon is now usable for lines Shape . See Shape creation.

Improved

• Global improvements in terms of performances and stability for the Video Annotation tool.
• A tall tool card (SAM, keypoint, pencil, video propagation) no longer covers the side toolbar of the Annotation tool: it is stacked above it, with its extra options in an overflow menu.
• Faster loading of the Asset list, of Campaign assignments, and of the Organization pages.
• Splitting a Track leaves both halves exactly where interpolation placed them before the split. See Video tool.

Fixed

• Custom shortcuts set on the panels of the side rail and of the footer of the Annotation tool are triggered again. See Settings & Shortcuts.
• On macOS, shortcuts combining Option with a letter — move to first or last frame, draw with pencil — are recognized again.
• Cutting a polygon with the Scissors through one of its points no longer breaks the Shape, and a cut that does not split it in at least two parts leaves it untouched.
• Moving to the next frame right after pausing the video player works.
• The frame number stays above the timeline cursor once the timeline is scrolled, and zooming the track grid keeps the frame at its center in place.
• The footer of the video player keeps the height of its single-row bar.
• A video still being processed displays its thumbnail instead of an error message.
• The number of open and closed Issue of an assignment is displayed in the details panel of the Annotation tool.
• A Chart can be deleted whatever its computation status. See Charts.
• The Fast Training button is no longer offered on a DatasetVersion whose Version Type does not support it.

v3.99

Added

• Filter the Tasks of a Campaign on what they are about, beyond their Status, Assignee, and Step: the Asset behind the Task, the Metadata and Custom Metadata of its Data, the Shape of its Annotation, and Attribute values. See Campaign.
• The JSON Key inputs of the Chart creation form suggest the Custom Metadata keys declared in the schema of the Datalake. See Charts.
• A Metric of an Experiment can be opened in a modal, over the whole page. See Experiment overview.
• A seasonal theme, displayed around a few dates of the year, which you can turn off from the Appearance tab of your Personal Settings. See Setup your profile.

Improved

• The Annotation tool has been rebuilt on a single foundation shared by DatasetVersion annotation, Campaign Tasks, Issues, and Deployment prediction review, with a new layout. The Shortcuts panel lists the shortcuts of the context you are working in. See Annotation tool and Settings & Shortcuts.
• The undo history of the Annotation tool now records the exact change made at each step, instead of a full copy of the Annotation.
• The annotation export of a DatasetVersion runs as a Job, so a DatasetVersion of any size can be exported from the interface, and an export already prepared is downloaded right away. See Annotation import and export.

Fixed

• Video Asset: the video element is released between two videos, keyframes and sub-track frame states stay aligned, and the player footer reopens on the next video. See Video tool.
• In a Campaign, Submit, Skip and Rework act on the Task they were triggered on, the queue stays in place past the first 200 Tasks, a Review Task can be submitted before its first save, and a Task can be reassigned from the tool. See Campaign.
• Clicking a Chart applies the matching filter to the list displayed next to it, it also works for Custom Medata. See Charts.
• Processing parameters keep their type when they are loaded and edited, and boolean values are supported. See Processings.
• The item selector dropdown keeps the width of its content in Safari.

v3.98

Added

• Manage the Custom Metadata schema of a Datalake from the Metadata page of its Settings: declare keys with their type and options, search the schema, and choose whether Data may carry keys the schema does not declare. See Metadata.
• Index a Custom Metadata key for search, follow the indexation status of each key, and take a key out of the index again. On a Datalake over 100,000 Data, only indexed keys can be filtered on. See Metadata.
• Charts are available to every Organization, on the Datalake, DatasetVersion, Deployment predictions, and Experiment evaluations. See Charts.

Improved

• Custom Metadata filters suggest the keys declared in the schema, including on Deployment predictions, Evaluation, evaluation comparison, and Campaign task and Issue lists. See Metadata.
• String filters default to the equality operator.

Fixed

• The type and the options of a Custom Metadata key are reachable when declaring or updating it from the Datalake Settings.
• Chart of a Deployment receive their computation updates again.
• Chart cards and the query language auto-completion used to scope them.
• The video player timeline and Track selection in the Annotation tool.
• The classification Label selector resets to the Label of the Asset currently displayed in the Annotation tool.
• Number filters validate their Min and Max bounds.
• Date filters apply a consistent default time when a date carries none.
• A user already part of the Organization can no longer be selected when adding members. See Users management system.

v3.97

Added

• Filter Asset on Asset Attribute and Shape Attribute values, and on the Shape their Annotation contains, from the Add filter menu. See Attributes.
• The same Attribute and Shape filters on the PredictedAsset of a Deployment and the Evaluation of an Experiment. See Predictions overview and Evaluation.
• Embeddings and visual search on the PredictedAsset of a Deployment and the Evaluation of an Experiment, activated from their Data Embeddings settings page.
• Average precision and average recall are computed for each Prediction as soon as its review is submitted, for the Champion Prediction and the Shadow Prediction. See Prediction Review tool.
• A Custom Metadata schema can accept keys that it does not declare.

Improved

• The inference latency sent along with a Prediction is now stored with it. See Dashboard metrics.
• The Datalake and DatasetVersion visual search settings are now the Data Embeddings and Shape Embeddings pages. See Explore your Data with Embeddings.
• Filter menus display only the sections that apply to what you are browsing, with clearer labels and ordering.
• Experiment previews include the experiment status and creation date, and the name of the evaluated DatasetVersion.

Fixed

• Date filters no longer fail when your system locale is not US.
• The Label list no longer stays stuck loading in the Annotation tool when a request fails.
• The Custom Metadata schema form handles fields that allow several values, tracks unsaved changes, and no longer offers a Required option. See Metadata.
• Comparison bar charts handle Experiment with missing or partial logs. See Experiment comparison.
• The rights of the Organization Owner can no longer be modified. See Users management system.

v3.96

Added

• Metric-only charts: a chart can now be created without a group-by field. See Analytics and charts.
• SAM-assisted annotation now supports masks. See Shape creation.
• issue_count and has_open_issues are available in the assignment query language. See Query language.
• When copying assets with their annotations, an unconfigured destination DatasetVersion now takes its type from the source, and a destination with no Label receives the source labels.

Fixed

• Attribute values are now saved on Track objects. They are not yet included in COCO exports.
• COCO export now interpolates segmentation and keypoints correctly.
• Video assets with a dimension larger than 2048px are handled.
• Copy button text no longer overflows, and the datetime picker behaves correctly.
• Image tiles and multi-page documents are now deleted along with their data.
• Query language fields task.asset.data.datalake, task.asset.data.data_source, task.asset.nb_annotations and task.asset.nb_review return correct results.

v3.95

Added

• Configurable keyboard shortcuts for grid display and query modes, item selection, and data scroller actions. See Keyboard shortcuts.
• A global shortcut to copy the identifier of the resource you are viewing.
• Visibility, label, text, and label-control toggles in the detail view, each with its own shortcut.

Improved

• Action buttons, player controls, and context menus have been reworked around a single floating menu.

v3.94

Maintenance release. No user-facing changes.

v3.93

Added

• View and edit your keyboard shortcuts from your user settings. See Keyboard shortcuts.
• Custom metadata can be edited from the details panel, and its schema from a dedicated page. See Metadata.
• An evaluation comparison page, reachable from a project. See Evaluation.
• Tags in the task detail section. See Campaign.

Fixed

• Datetime query parameters are handled in filters.
• Number inputs always return a number.
• The video timeline no longer jumps when a keyframe is clicked. See Video tool.

v3.92

Maintenance release. No user-facing changes.

v3.91

Added

• Masks in evaluations. See Evaluation.
• An export_key option in annotation exports, allowing files to be keyed by asset identifier, data identifier, object name, or filename.
• A bulk data update API for custom metadata, metadata, tags, and source.
• GPU type selection when launching an experiment, available through the SDK. See Launch a training.
• Brightness, noise, and related fields in metadata charts. See Charts.
• Datalake administrators who are organization users can sync objects from the SDK.

Improved

• The query language has been rewritten on the backend; all writable queries are now supported. See Query language.
• Attributes are enabled for all organizations. See Attributes.
• Equality filters on custom metadata are faster.
• DatasetVersion statistics load faster in the dataset version list.
• Font sizes adapt better to small screens.

Fixed

• The last bin of integer charts is displayed correctly.
• Search no longer suffers from concurrency issues.
• Multipart upload initialization.
• Tile viewer errors.
• Uploading large lists of files, including large data sets and PVOC files.

v3.90

Added

• Keypoints and masks in deployments. See Deployments.
• Masks in evaluations. Shapes are displayed; comparison is not yet available. See Evaluation.
• Annotation file caching: requesting an export again for an unchanged dataset returns the same file.
• A webhook fired when an assignment is completed. See Configure webhooks.

Fixed

• Query language queries using task.annotation work on the assignment list.

v3.89

Added

• Mask drawing. See Shape creation.

v3.88

Added

• A MASK inference type, with mask storage and display in both the studio and the asset grid. See Shape creation.
• Shape lists in campaigns, and annotations displayed in the assignment grid and table. See Campaign.
• Attribute support in deployments, set on the label settings page and filled in on shapes from the studio. See Attributes.

Improved

• All Label pages have been reworked, including new pages for ModelVersion and Deployment. See Model version.

Fixed

• TIFF uploads now generate an RGB projection asynchronously. See Projections.
• Missing assets are synced when automatic addition of new assets is enabled.

v3.87

Added

• Shape lists for the MULTI inference type.
• A chart update form, a chart list in the data scroller, and query redirection from a chart. See Analytics and charts.
• Autocompletion for query language input in the chart creation and update forms. See Query language.
• An OVH storage connector. See Storages and datalakes.

Improved

• Chart forms now offer only the aggregate and group-by options valid for the current selection.

v3.86

Added

• Charts for experiment evaluations. See Evaluation.
• Keypoints in shape lists and in evaluations.

Fixed

• Null bytes in EXIF data no longer cause errors.

v3.85

Improved

• All attributes are shown at all times for classification shapes. See Attributes.

v3.84

Added

• Keypoints and skeletons in the studio, with reusable skeleton templates. See Annotation tool.
• Keypoints, split track selection, and a cut-to-cursor action for video. See Video tool.
• Shape text and label distribution charts for a DatasetVersion. See Analytics and charts.

v3.83

Maintenance release. No user-facing changes.

v3.82

Added

• Attributes. See Attributes.

v3.81

Added

• A new interface for the processing system, covering input creation, input update, and launching a processing with inputs. See Data processings.
• COCO annotation exports include attributes. See Attributes.

v3.80

Added

• Query language support in Datalake charts. See Charts.
• APIs for the attributes system. See Attributes.
• Visual search endpoints for deployments.

v3.79

Added

• Lock and unlock a DatasetVersion. See Dataset versioning system.
• Visual search for shapes. See Shapes view.
• Bucket object import records an error summary on the Internal Job page. See Track progress with jobs.
• An outlier_score alert on monitoring. See Set up alerts.

Improved

• Pill, badge, and capsule styles are unified across the application.
• The scan creation form matches the experiment creation form. See Experiment creation.
• Embedding projection computation. See Explore your data with embeddings.

Fixed

• Reassignment in campaigns. See Campaign.
• Deployment of non-PyTorch models on serving.

v3.78

Added

• Keypoint support across the application: visualize, draw, edit, and update skeletons. See Shape creation.
• A new experiment creation form. See Experiment creation.
• created_by for predicted assets in the details and table views. See Predictions.
• A queued status for jobs. See Track progress with jobs.

Improved

• The platform now runs on Python 3.14.

v3.77

Added

• The studio now has the same Details panel as the Details view in the grid, including its History and Issues panels. See Annotation tool.
• Switch the label of a shape with Ctrl and right-click, and hide or show selected shapes with a shortcut. See Shortcuts.
• nb_tags filtering in the assets, evaluations, and predicted assets query language. See Query language.

Improved

• Experiment creation has been reworked, covering base experiments, base models, attached datasets, and parameters. See Experiment creation.
• Importing and forking assets and annotations supports multiple origins. See Import annotations from another version.

v3.76

Added

• Points and lines in monitoring. See Deployments.
• A Projection Names field on the organization, defining which projections are available in the interface, together with an organization settings page to set them. See Projections.
• Visualize a selected projection for every image in the grid. See Projections.
• nb_tags in the Datalake and dataset query language. See Query language.
• Switch between review, champion, and shadow at grid level. See Shadow deployment.

Improved

• Data.filename coming from monitoring now uses the filename passed to monitor().
• Label colors are propagated on forks. See Fork a dataset.
• Names of models, experiments, and projects are validated more strictly.

Fixed

• The deployment global distribution chart. See Dashboard metrics.
• Heatmaps and AP and AR are no longer displayed for line and point deployments.
• Attaching a model to an experiment no longer overrides its label map or parameters.
• The position of the "Change label" button.