Eval: YOLO26 as primary detection/pose backend #1

New issue

Open

opened 2026-06-23 09:30:12 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-06-23 09:30:12 -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 covering detection, instance segmentation, pose
estimation, oriented detection, and classification in a single pipeline. NMS-free
end-to-end inference. 40.9–57.5 mAP on COCO at 1.7–11.8ms T4 TensorRT latency.

YOLOE-26 variant: open-vocabulary (zero-shot, text-prompted) detection. No
retraining required for new categories — "find all food items" works out of the box.

Export targets: TensorRT, ONNX, CoreML, TFLite, OpenVINO, NCNN, ExecuTorch.

cf-vision integration

YOLO26 is a strong candidate for the primary backend in cf-vision, covering the
needs of all current planned consumers:

Consumer	Task head
Kiwi	detect (barcodes, food items), YOLOE-26 zero-shot
Nuthatch	detect (obstacles), pose (human proximity)
Raven	pose (gesture/hands-free input)

License constraint (CRITICAL)

Code/weights: AGPL-3.0 or Ultralytics commercial license.

Free/local tier: AGPL-3.0 acceptable (user runs locally, no SaaS distribution issue).
Cloud/paid tier: AGPL requires source disclosure for network services — hard blocker
without Ultralytics Enterprise license OR inference-only wrapper around raw weights.

cf-vision's backend abstraction must accommodate a swap to a non-AGPL runtime for
cloud tier (e.g., bare ONNX weights + custom inference, no Ultralytics training code).
The interface should be identical; only the backend differs between free and cloud.

Recommended eval steps

Install ultralytics in cf conda env, run YOLO26n on sample Kiwi images
Test YOLOE-26 zero-shot food detection ("cereal box", "apple", "can of soup")
Benchmark pose estimation latency on Heimdall GPU vs CPU fallback
Assess Jetson TensorRT export for Nuthatch path
Decision: AGPL acceptable for free tier? Budget Ultralytics Enterprise for cloud?

roadmap#51: cf-vision module RFC
kiwi#153: YOLO26 barcode/food detection eval
raven#26: YOLO26 pose for gesture input
roadmap#52: Nuthatch robot vision eval

## 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 covering detection, instance segmentation, pose estimation, oriented detection, and classification in a single pipeline. NMS-free end-to-end inference. 40.9–57.5 mAP on COCO at 1.7–11.8ms T4 TensorRT latency. YOLOE-26 variant: open-vocabulary (zero-shot, text-prompted) detection. No retraining required for new categories — "find all food items" works out of the box. Export targets: TensorRT, ONNX, CoreML, TFLite, OpenVINO, NCNN, ExecuTorch. ## cf-vision integration YOLO26 is a strong candidate for the primary backend in cf-vision, covering the needs of all current planned consumers: | Consumer | Task head | |----------|-----------| | Kiwi | detect (barcodes, food items), YOLOE-26 zero-shot | | Nuthatch | detect (obstacles), pose (human proximity) | | Raven | pose (gesture/hands-free input) | ## License constraint (CRITICAL) Code/weights: **AGPL-3.0** or Ultralytics commercial license. - Free/local tier: AGPL-3.0 acceptable (user runs locally, no SaaS distribution issue). - Cloud/paid tier: AGPL requires source disclosure for network services — hard blocker without Ultralytics Enterprise license OR inference-only wrapper around raw weights. cf-vision's backend abstraction must accommodate a swap to a non-AGPL runtime for cloud tier (e.g., bare ONNX weights + custom inference, no Ultralytics training code). The interface should be identical; only the backend differs between free and cloud. ## Recommended eval steps 1. Install `ultralytics` in cf conda env, run YOLO26n on sample Kiwi images 2. Test YOLOE-26 zero-shot food detection ("cereal box", "apple", "can of soup") 3. Benchmark pose estimation latency on Heimdall GPU vs CPU fallback 4. Assess Jetson TensorRT export for Nuthatch path 5. Decision: AGPL acceptable for free tier? Budget Ultralytics Enterprise for cloud? ## Related - roadmap#51: cf-vision module RFC - kiwi#153: YOLO26 barcode/food detection eval - raven#26: YOLO26 pose for gesture input - roadmap#52: Nuthatch robot vision eval

No labels

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference: Circuit-Forge/cf-vision#1

No description provided.