Smart Fridge YOLO11n
这是一个面向智能冰箱场景训练的 YOLO11n 食材目标检测模型。模型用于边缘端预识别、区域定位、入库变化提醒和重复候选标记；食物名称确认、可见状态分析及综合建议由后续 VLM 和规则系统处理。
This YOLO11n detector is trained for smart-fridge food pre-detection. It is intended for object localization, inventory-change triggers, and duplicate-candidate marking on edge devices. Semantic confirmation and visible-condition assessment are delegated to a downstream VLM and rule layer.
Files
| File | Description |
|---|---|
best.pt |
Best Ultralytics checkpoint from the 80-epoch run |
model.onnx |
ONNX export used by the RK3399 edge runtime |
classes.txt |
Ordered list of 30 output classes |
training/results.csv |
Per-epoch training and validation metrics |
training/*.png |
Training curves, precision-recall curve, and confusion matrices |
training_config.yaml |
Sanitized training configuration without local paths |
SHA256SUMS |
SHA-256 checksums for all uploaded artifacts |
Training Data
- Dataset: fridge food images v14
- Source: Roboflow Universe
- Images reported by the exported dataset: 5,139
- License reported by the dataset export: CC BY 4.0
- Classes: 30 food categories, listed in
classes.txt
The class labels preserve the spelling used by the source dataset for compatibility with the deployed pipeline.
Training
| Item | Value |
|---|---|
| Base model | YOLO11n pretrained checkpoint |
| Framework | Ultralytics 8.4.84 / PyTorch 2.12.1 |
| Epochs | 80 |
| Image size | 640 x 640 |
| Batch size | 8 |
| Device | Apple MPS |
| Seed | 0, deterministic mode enabled |
Validation Results
| Checkpoint | Epoch | Precision | Recall | mAP50 | mAP50-95 |
|---|---|---|---|---|---|
| Best | 66 | 0.81796 | 0.74272 | 0.81128 | 0.58557 |
| Final | 80 | 0.81968 | 0.73055 | 0.80675 | 0.58047 |
Metrics come from the validation split configured in the exported dataset.
Usage
from ultralytics import YOLO
model = YOLO("best.pt")
results = model.predict("fridge.jpg", imgsz=640, conf=0.65)
The ONNX file is intended for CPU inference with ONNX Runtime. The deployed smart-fridge pipeline uses a confidence threshold of 0.65 before triggering downstream VLM analysis.
Limitations
- The model only detects the 30 classes in
classes.txtand may not generalize to unseen foods, packaging, lighting, occlusion, or camera angles. - A detection is not proof of identity, freshness, edibility, expiry, or food safety.
- The model was trained primarily on a public image dataset and requires evaluation or adaptation for each physical refrigerator.
- Dataset class names include source spelling such as
blue berryandstawberry; applications may map them to corrected display names without changing output indices. - This model should be used as a pre-detection component, with human review or downstream multimodal analysis for consequential decisions.
License
Ultralytics states that models trained with its YOLO training stack are covered by AGPL-3.0 by default. Commercial or proprietary deployment may require an Ultralytics Enterprise license. The training dataset is separately attributed as CC BY 4.0 by its Roboflow export; users are responsible for complying with both the model and dataset terms.
- Downloads last month
- 20