New Release: Ultralytics v8.4.122

Discussion
, ,
1

:rocket: Ultralytics v8.4.122: Compact, Portable PNG Depth Maps

Ultralytics v8.4.122 standardizes the YOLO26-Depth dataset pipeline around compact, self-describing 16-bit PNG depth maps. This update simplifies depth-data handling, improves portability, and substantially reduces dataset storage requirements. :framed_picture::chart_decreasing:

:sparkles: New Features

  • Canonical depth format: Depth targets now use 16-bit PNG files with embedded metadata describing their range in meters.
  • Explicit invalid pixels: Pixel value 0 represents missing or invalid depth, allowing training and evaluation to ignore these pixels safely.
  • New depth utilities:
    • save_depth_png() writes depth maps in the required format.
    • load_depth() reconstructs meter-valued depth maps for training and inference.
  • Consistent NDJSON encoding: Depth records now follow the linear-u16 PNG encoding contract.

:package: Dataset Improvements

Built-in configurations and generators now produce and consume canonical PNG depth maps for:

  • Depth8
  • NYU Depth
  • KITTI
  • SUN RGB-D
  • DIODE
  • Hypersim
  • TartanAir
  • Virtual KITTI 2
  • ARKitScenes

PNG compression also makes published archives significantly smaller:

  • Depth8: approximately 8.9% of its previous size
  • NYU Depth: approximately 39.6% of its previous size

That means faster downloads, lower storage costs, and more portable depth datasets. :rocket:

:hammer_and_wrench: Improvements

  • Unified depth loading around one format and one code path.
  • Refreshed depth dataset guides, task documentation, reference pages, and validation tests.
  • Improved portability through metadata embedded directly in each depth file.
  • Better behavior when 16-bit depth assets pass through browser or display systems that reduce images to 8-bit.
  • Simplified manifests, converters, and depth-path handling.

This release focuses on the YOLO26-Depth data representation and dataset pipeline. It does not introduce a major change to the model architecture or training objectives.

:warning: Breaking Change and Migration

Support for .npy depth maps has been removed, including compatibility fallbacks.

Custom datasets using .npy targets must be converted to the new metadata-enabled PNG format before training or evaluation with v8.4.122. Existing hosted depth assets also require migration because this release does not include a runtime compatibility layer.

:bug: Bug Fixes

No standalone bug fixes are highlighted in this release. The primary change is the comprehensive PNG-only standardization of depth datasets, utilities, conversion paths, documentation, and tests.

:raising_hands: Contributor Spotlight

This release includes PR #25858: Use canonical PNG depth maps by @glenn-jocher. Thank you to the entire YOLO community and Ultralytics team for continuing to improve the depth workflow.

:inbox_tray: Try v8.4.122

Upgrade with:

pip install -U ultralytics

You can review the complete release on the Ultralytics v8.4.122 release page and inspect every change in the full v8.4.121 to v8.4.122 changelog.

Please try the new PNG depth workflow and share your feedback, migration experiences, or issues with the community. :speech_balloon: