Eval: YOLO26 for barcode detection and food item recognition #153

New issue

Open

opened 2026-06-23 09:29:09 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-06-23 09:29:09 -07:00

Owner

Source

Paper: https://arxiv.org/abs/2606.03748 — Ultralytics YOLO26 (June 2026)
Code/weights: https://github.com/ultralytics/ultralytics

What it is

Unified real-time vision model family. Single model covers object detection, instance
segmentation, pose estimation, oriented detection, and classification. NMS-free
end-to-end inference. 40.9–57.5 mAP on COCO at 1.7–11.8ms T4 TensorRT latency.

YOLOE-26 variant adds open-vocabulary (zero-shot text-prompted) detection — no
retraining needed for new categories.

Export targets: TensorRT, ONNX, CoreML, TFLite, OpenVINO, NCNN, ExecuTorch (edge-ready).

Kiwi use cases

Barcode detection: Replace or augment current barcode scanning with YOLO's oriented
detection head — handles tilted/partial barcodes in photos.
Food item recognition: YOLOE-26 open-vocabulary mode could identify pantry items
from photos without a food-specific training run ("what food is in this photo?").
Receipt OCR assist: Detect product label regions before passing to OCR pipeline,
improving accuracy on cluttered receipts.
Future mobile: CoreML/TFLite export paths support on-device inference for a future
Kiwi mobile app.

License situation

Code/weights: AGPL-3.0 (open source) or Ultralytics commercial license.

Free/local tier: AGPL-3.0 is fine — user runs inference locally, no SaaS distribution.
Cloud/paid tier: AGPL-3.0 requires source disclosure for network services — hard blocker.
Either purchase Ultralytics Enterprise license or wrap only the model weights (not
the Ultralytics training framework) in cf-core's inference layer.

Recommendation

Prototype YOLOE-26 zero-shot food detection for the photo-add flow (free tier, local
inference). Evaluate whether it reduces the need for barcode-only input.
If it ships to cloud tier, budget for Ultralytics commercial license or design around
inference-only (weights) usage.

cf-core vision module (see roadmap ticket)
Current barcode path in kiwi/app/

## Source Paper: https://arxiv.org/abs/2606.03748 — Ultralytics YOLO26 (June 2026) Code/weights: https://github.com/ultralytics/ultralytics ## What it is Unified real-time vision model family. Single model covers object detection, instance segmentation, pose estimation, oriented detection, and classification. NMS-free end-to-end inference. 40.9–57.5 mAP on COCO at 1.7–11.8ms T4 TensorRT latency. YOLOE-26 variant adds open-vocabulary (zero-shot text-prompted) detection — no retraining needed for new categories. Export targets: TensorRT, ONNX, CoreML, TFLite, OpenVINO, NCNN, ExecuTorch (edge-ready). ## Kiwi use cases - **Barcode detection:** Replace or augment current barcode scanning with YOLO's oriented detection head — handles tilted/partial barcodes in photos. - **Food item recognition:** YOLOE-26 open-vocabulary mode could identify pantry items from photos without a food-specific training run ("what food is in this photo?"). - **Receipt OCR assist:** Detect product label regions before passing to OCR pipeline, improving accuracy on cluttered receipts. - **Future mobile:** CoreML/TFLite export paths support on-device inference for a future Kiwi mobile app. ## License situation Code/weights: **AGPL-3.0** (open source) or Ultralytics commercial license. - Free/local tier: AGPL-3.0 is fine — user runs inference locally, no SaaS distribution. - Cloud/paid tier: AGPL-3.0 requires source disclosure for network services — hard blocker. Either purchase Ultralytics Enterprise license or wrap only the model weights (not the Ultralytics training framework) in cf-core's inference layer. ## Recommendation Prototype YOLOE-26 zero-shot food detection for the photo-add flow (free tier, local inference). Evaluate whether it reduces the need for barcode-only input. If it ships to cloud tier, budget for Ultralytics commercial license or design around inference-only (weights) usage. ## Related - cf-core vision module (see roadmap ticket) - Current barcode path in `kiwi/app/`