BlackfireZZZ/lenta-tech-life-2026-detector

Lenta Tech Life 2026 — Price-Tag Detector

Fine-tuned YOLO11x that detects supermarket price tags in frames captured by a shelf-scanning robot. It is the detection stage of an end-to-end price-tag recognition pipeline (detect → track → dedup → read fields → CSV) built for the Lenta Tech Life 2026 hackathon.

Model

• Architecture: Ultralytics YOLO11x (single class: price_tag).
• Base: openfoodfacts/price-tag-detection (itself a YOLO11 model), fine-tuned for the Lenta robot camera.
• Trained upright. The robot camera is mounted rotated 90° CCW; the detector is trained on de-rotated (upright) frames, so the runtime must feed it upright frames (the pipeline config sets frame_rotation: ccw for offline clips; uploaded clips are already upright).

Training data

Fine-tuned on the provided Lenta dataset (494 frames) plus external open-license datasets auto-labeled with open-source models, under strong camera-matched augmentation (motion blur, glare, perspective, wide-angle distortion) to match the robot's ultra-wide 2.8 mm lens. No cloud or online services are used at inference — the whole pipeline runs locally.

Usage

The pipeline resolves this model automatically from its config via an hf:// reference — no manual download needed for docker compose up:

detector:
  model_path: hf://BlackfireZZZ/lenta-tech-life-2026-detector/lenta_price_tag_detector_full494_off_aug_best.pt

Or load it directly with Ultralytics:

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

p = hf_hub_download(
    "BlackfireZZZ/lenta-tech-life-2026-detector",
    "lenta_price_tag_detector_full494_off_aug_best.pt",
)
model = YOLO(p)

License

AGPL-3.0, inherited from Ultralytics YOLO11 and the OpenFoodFacts base model. Any deployment must comply with AGPL-3.0.