Reports and images

Examples and report reading

See what each workflow produces and which files to inspect.

stable

Instance-segmentation report and overlays

A checked demo that pairs a JSON metric report with visible mask overlays.

Instance-segmentation report and overlays

Expected companion files or fields

  • instance_seg_demo_report.json
  • overlays/*.png
  • demo_config.json

Source artifact

stable

First COCO evaluation and release gate

Generate eight labeled images, validate the inputs, and run COCOeval. The example also includes a release check. Its synthetic predictions test the workflow, not model accuracy.

Expected companion files or fields

  • labeled_preview.png
  • labeled_sample_coco.json
  • labeled_sample_qualification/qualification.json
  • labeled_sample_qualification/checksums.json

Source artifact

bridge

External framework prediction artifacts

Convert Detectron2, MMDetection, YOLOv8, or YOLOX outputs to the shared prediction format before scoring.

Expected companion files or fields

  • wrapped predictions JSON
  • meta.export_settings
  • adapter/source identity

Source artifact

benchmark

Backend support and skipped runs

The support matrix lists implemented, conditional, and unsupported backends. A skipped run has no passing result.

Expected companion files or fields

  • benchmark support metadata
  • explicit support status
  • skip or failure reason where applicable

Source artifact

research

TTT Research evidence requirements

This synthetic image illustrates the required files and source records. It does not measure TTT effectiveness; TTT remains a Research feature.

TTT Research evidence requirements

Expected companion files or fields

  • baseline_predictions.json
  • method_predictions.json
  • before_after_compare.json
  • TTT log and plan

Source artifact · Stable baseline artifact

Before comparing two reports

  • Same task and metric family
  • Same dataset identity and split
  • Validated predictions interface contract
  • Recorded producer and export settings
  • Pinned protocol settings
  • Class and coordinate semantics aligned
  • Skipped, conditional, or unsupported lanes kept explicit
  • Visible evidence linked where the task provides it
Return to stable evaluation Resolve a mismatch