pavanyadava07/makerspace-yolo11n

YOLO11n makerspace detector, v0-public

Detector of the Makerspace Reuse Scanner: a camera at the bench detects parts and offcuts, the API keeps a reuse inventory and a cited local knowledge base answers "how do I reuse or dispose of this?". Live demo: Space pavanyadava07/makerspace-reuse-scanner.

Files: yolo11n_makerspace.onnx (static 640×640, opset 17, output 1×19×8400, runs with ONNX Runtime and NumPy pre/post-processing, see edge/detector.py in the repo) · yolo11n_makerspace.pt (Ultralytics checkpoint) · classes.yaml (15 classes) · eval_2026-09-02.{md,json} (the only source of accuracy numbers) · bench_*.json (device-labelled throughput) · SOURCES.md (data + licences).

Measured on the held-out test split (2026-09-02)

overall	mAP50	mAP50-95	precision	recall
all classes	0.410	0.303	0.525	0.423

class	AP50	AP50-95
screw	0.995	0.872
nut_bolt	0.226	0.086
pcb	0.994	0.917
motor	0.441	0.243
battery	0.030	0.016
tool	0.049	0.019
plastic_container	0.482	0.265
cardboard	0.179	0.121
glass	0.295	0.189

Read before trusting these: screw and pcb test images are single-object studio shots, so their scores flatter the model; battery and tool have too little data to be usable; the dominant failure is missed small or distant objects, not class confusion. Six of the 15 classes (filament_spool, wood_offcut, cable, 3d_print_part, acrylic_sheet, metal_profile) have no training data yet and are never predicted. Throughput: 96.09 ms median inference on an x86 CPU execution provider (bench_*.json); no Raspberry Pi 5 number yet.

Training data and licence

Trained by training/build_public_dataset.py on public data only: LVIS on COCO images (CC BY 4.0), TACO (CC BY 4.0), TrashNet (research use), a cropped-PCB set from Hugging Face (no licence declared) and MVTec AD screws (CC BY-NC-SA 4.0), plus train-only copy-paste composites. Because of the MVTec share these weights are released non-commercial, share-alike (CC BY-NC-SA 4.0). Recipe: YOLO11n, frozen first 10 layers, 80 epochs, imgsz 640, seed 0 (training/train.py).

Use

import onnxruntime as ort, numpy as np
sess = ort.InferenceSession("yolo11n_makerspace.onnx", providers=["CPUExecutionProvider"])
# input: 1×3×640×640 float32 RGB in [0,1], letterboxed; output: 1×19×8400 = (cx, cy, w, h, 15 class scores) per anchor

Author: Pavan Yadav Annappa · MIT for the code, CC BY-NC-SA 4.0 for these weights.