🌟 Summary
v8.4.214 makes image orientation consistent across reads, helping ensure labels, predictions, and displayed images all use the same upright view. It also includes platform, hardware-compatibility, and reliability fixes, plus a substantially refreshed OpenVINO guide.
📊 Key Changes
- 🖼️ EXIF orientation is now applied consistently when reading images, across supported formats, read flags, and OpenCV versions. This is the release’s main functional change: images should appear in the same orientation that viewers display, reducing the risk of labels or predictions being rotated relative to the image.
- 📚 OpenVINO documentation was reorganized and expanded with clearer export and deployment steps, INT8 quantization guidance, device behavior and limitations, and updated benchmarks for YOLO26 on server CPUs and Intel Core Ultra hardware. The guide documents deployment options rather than introducing a new model.
- 🧠 Segmentation mask shape checks avoid building full mask arrays when only the shape is requested, reducing unnecessary memory use.
- 🖥️ GPU detection now uses PyTorch’s device list, allowing AMD GPUs to be detected even when NVIDIA tools such as
nvidia-smior NVML are unavailable. Idle-GPU checks are skipped when utilization data cannot be read. - 🧰 Several edge cases were fixed: thin images no longer resize to zero pixels during compression; Windows can download weights from HTTP(S) URLs; and bounding-box-to-segmentation conversion keeps labels separate for matching image names in different subfolders.
- 🧹 RT-DETR training no longer reports unused class weights. This cleans up logging; training losses are unchanged.
- 📝 Platform and usage docs were updated to explain the 1,000-prediction limit for most Platform YOLO workflows (300 for RT-DETR), image rotation controls, cloud-training project naming, and prediction examples.
🎯 Purpose & Impact
- ✅ More reliable image handling: applying EXIF orientation consistently helps prevent annotation and inference mismatches, especially for images with embedded rotation metadata.
- ⚡ Less unnecessary memory work: querying segmentation mask dimensions no longer expands box-local masks into full-image masks.
- 🌐 Broader hardware and platform support: AMD GPU detection and Windows URL downloads work more reliably, while the documentation better explains how to deploy YOLO26 with OpenVINO.
- 🛠️ Safer data workflows: compression and label conversion now handle thin images and nested train/validation directories without the associated failures or overwritten labels.
What's Changed
- Platform docs: 1,000 detections per image, local save directory project naming, cloud training title by @pderrenger in #27204
- Update YouTube video link in
predictdocs by @RizwanMunawar in #27205 - Fix zero-sized dimensions in image compression by @AhmetMert01 in #27201
- Download URL weights on Windows in
attempt_download_assetby @MohammadHijjawi97 in #27202 - Keep train and val labels separate in yolo_bbox2segment by @cainiao33 in #27209
- Read segmentation mask shapes without building the full mask stack by @aswanth-07 in #27207
- Restructure the OpenVINO integration page and add server CPU benchmarks by @raimbekovm in #27208
- Detect AMD GPUs when nvidia-smi and NVML are unavailable by @AIwork4me in #27203
- Stop RT-DETR training from logging unused cls_pw class weights by @cainiao33 in #27210
- Read every image upright: apply EXIF orientation for all formats and flags by @sergiuwaxmann in #27168
New Contributors
- @AIwork4me made their first contribution in #27203
Full Changelog: v8.4.213...v8.4.214