SKU-110K YOLO11 Object Detector 📦
Two lightweight Ultralytics YOLO11 checkpoints trained on SKU-110K at imgsz=640.
The goal is simple: detect product-like objects on shelves and in real-world photos. It is a handy baseline for visual search prototypes, shelf analysis, dense object detection experiments, and other R&D work where a generic object detector is useful.
Try it in the browser: free Gradio demo 🚀
The models predict one class:
0: object
This is not a product-category classifier or a brand-recognition model. It just finds generic retail/product objects. That makes it flexible, but also means you should test it on your own images before using it in any real pipeline.
What's Inside ✨
| model | PyTorch | ONNX | note |
|---|---|---|---|
| YOLO11n 640 | weights/sku110k-yolo11-n640.pt | weights/sku110k-yolo11-n640.onnx | small, fast, easy to try |
| YOLO11s 640 | weights/sku110k-yolo11-s640.pt | weights/sku110k-yolo11-s640.onnx | stronger metrics, still lightweight |
The ONNX exports are fixed-shape exports with input shape [1, 3, 640, 640].
Basic Metrics 📊
Final test evaluation was run on the SKU-110K test split. Ultralytics skipped one corrupt test image, so the reported test run used 2,935 images and 431,419 instances.
| model | precision | recall | mAP50 | mAP50-95 | inference |
|---|---|---|---|---|---|
| YOLO11n 640 | 0.896 | 0.838 | 0.906 | 0.550 | 0.9 ms/img |
| YOLO11s 640 | 0.913 | 0.867 | 0.927 | 0.577 | 4.3 ms/img |
Validation checkpoints:
| model | best epoch | val precision | val recall | val mAP50 | val mAP50-95 |
|---|---|---|---|---|---|
| YOLO11n 640 | 31 | 0.8962 | 0.8380 | 0.9023 | 0.5440 |
| YOLO11s 640 | 41 | 0.9142 | 0.8529 | 0.8969 | 0.5600 |
Qualitative Example 🖼️
This real-world shelf photo was annotated with the published YOLO11s SKU-110K checkpoint. It is included for visual context only. It is not part of the SKU-110K benchmark and should not be read as category-specific validation for these weights.
Quick Start 🚀
from ultralytics import YOLO
model = YOLO("weights/sku110k-yolo11-s640.pt")
results = model.predict("image.jpg", imgsz=640, conf=0.25)
For the smaller model:
from ultralytics import YOLO
model = YOLO("weights/sku110k-yolo11-n640.pt")
results = model.predict("image.jpg", imgsz=640, conf=0.25)
R&D Notes ⚠️
These weights are research and development artifacts. They are useful for experiments, demos, and baseline comparisons, but they have not been validated for production use, safety-critical use, or automated business decisions.
The training dataset is SKU-110K, a dense retail shelf object dataset. Expect the model to detect many product-like things, and expect misses or extra boxes on photos that look very different from retail shelves. Use the included example as a quick visual check, then validate on your own target data before deployment.
The weights are derived from SKU-110K training data. Keep publication and downstream use aligned with the upstream SKU-110K dataset terms; this model card does not claim a more permissive license than the source data allows.
Files 📁
weights/sku110k-yolo11-n640.pt: YOLO11n PyTorch checkpointweights/sku110k-yolo11-n640.onnx: YOLO11n ONNX exportweights/sku110k-yolo11-s640.pt: YOLO11s PyTorch checkpointweights/sku110k-yolo11-s640.onnx: YOLO11s ONNX exportexamples/: qualitative example image generated with the published YOLO11s checkpointmetrics.json: structured metrics and dataset summarychecksums.sha256: SHA-256 checksums for published files
Verify local files with:
shasum -a 256 -c checksums.sha256