SMB Inventory Vision DETR
Production DETR-ResNet-50 detectors for SMB Inventory Vision. This is the default backend (INVENTORY_DETECTOR_BACKEND=detr) — better on the merchant shelf scenes this project was tuned for than the optional YOLO extra.
| Hub folder | Local path after download | Role |
|---|---|---|
shelf/ | models/detr_sku_merchant_v2_neg | Dense shelf / SKU facings (num_queries=400) |
table/ | models/detr_informal_v1 | Table / counter / single item |
These checkpoints are Apache-2.0 (facebook/detr-resnet-50 lineage). YOLO weights in a separate Hub repo are AGPL-3.0.
Draws bounding boxes around product facings. Naming SKUs is a separate catalog-match / OCR / Gemini stage in the server.
Files (and nothing else)
Each folder is a transformers DetrForObjectDetection directory:
config.jsonpreprocessor_config.jsonmodel.safetensors
Training metadata that listed private merchant photo stems is not included.
SHA256: see SHA256SUMS.
Download
python scripts/download-detr-weights.py
Or from this repo:
from huggingface_hub import snapshot_download
snapshot_download("jz67/localfirst-vision-detr", repo_type="model")
Intended use
Small-business inventory photos (shelf, counter, one item). Not a general COCO detector. Not for identifying people.