Release 2026/2/4

Feb Update

This month's update focuses on making DataVerse faster and more intuitive.Key updates include a new Home Page for quick project access, a simpler training setup interface, and the ability to edit tags without switching screens. We've also added Class-Agnostic NMS for better detection results.

New Features

New Home Page

Your new project command center. The Home Page now displays project status, recent updates, workflow progress, and activity notifications in one unified dashboard. Quickly access datasets, data slices, and training jobs without navigating through multiple pages.

Smart Training Pipeline

We've redesigned the training pipeline configuration interface to make it more intuitive and user-friendly. The new design focuses on three core training decisions with clear explanations to help users make informed choices.

Simplified Training Configuration

• Training Type: Clear choice between two training workflows with distinct purposes.

  • New Model: Training for new object categories or new tasks

  • Fine-Tune Model: Update existing model with new data using fewer epochs

Model Structure: Clear size options with explicit model architectures

• Small - YOLOv9-s: Fastest inference, lower accuracy

• Medium - YOLOv9-c: Balanced performance (Default)

• Large - YOLOv9-e: Highest accuracy, slower inference

Data Augmentation Interface: Data augmentation made simple. Four intuitive presets for color/brightness and position/size adjustments eliminate the need to tune individual parameters

• Photometric (Color & Brightness)

  • Off: No color changes (Use when color is critical, e.g., clothing color classification)

  • Conservative (Default): Light color/brightness variation, stays close to original appearance

  • Aggressive: Heavy variation for diverse lighting conditions

  • Custom: Manual control of hue, saturation, and brightness parameters

• Geometric (Position & Size)

  • Off: No spatial changes (Use when orientation matters, e.g., directional traffic signs)

  • Conservative: Moderate position/size variation without flipping

  • Aggressive: Comprehensive transformations including mosaic, mixup, copy-paste

  • Custom: Manual control of translate, scale, flip, mosaic, mixup, and copy-paste

Class-Agnostic NMS

When enabled, NMS filters overlapping boxes across all classes instead of per class. Available during model training, prediction, and model conversion.

Without Class-Agnostic NMS: Bounding boxes for person and motorcycle are displayed separately.

With Class-Agnostic NMS: Overlapping detections in nearby regions are merged into single bounding boxes.

Edit Tags Without Leaving Your Review Flow

When inspecting images in Data Visualization, you can now edit tags directly in the preview window. No more switching to the annotation panel—update tags on the spot and continue your review seamlessly. Perfect for QA workflows and rapid dataset curation.

Coming Soon

Batch Tag

Apply tag changes to multiple images at once instead of editing one by one. Select images in Data Visualization, then add, remove, or replace tags across your entire selection in a single action—dramatically speeding up dataset curation and quality review workflows.

VQA Prompt Testing Tool

Test and compare different VQA prompts before running full-scale inference. Preview model responses, iterate on prompt wording, and validate output quality on sample images more efficiently.