Initial FactoryVision AI model package (private review)

README.md

@@ -0,0 +1,66 @@
+---
+language:
+- en
+pipeline_tag: object-detection
+library_name: ultralytics
+tags:
+- industrial-vision
+- factoryvision-ai
+- ppe-detection
+- education
+- yolo11
+---
+# FactoryVision AI — PPE YOLO11n (fine-tuned)
+
+**FactoryVision AI — Credits and Attribution**
+
+**Original Developer & Model Author:** Thevanth Rajamohan  
+**Technical Review, Code Contributions & Model Publication Support:** Swapin Vidya  
+**Publishing Organisation:** TechSphere Consultancy Services Private Limited  
+**Hugging Face:** https://huggingface.co/TechsphereAI  
+**Official Website:** https://techsphereconsultancy.com/
+
+**Project Provenance:** FactoryVision AI was originally developed by Thevanth Rajamohan. Swapin Vidya contributed technical review, detection-metrics improvements, testing, model publication preparation and documentation. The models are being prepared for publication under TechSphere Consultancy Services Private Limited's Hugging Face namespace, with the original developer's permission.
+
+**Original GitHub Repository:** https://github.com/Theivanth/Factory-Vision-AI  
+**Merged contribution by Swapin Vidya:** https://github.com/Theivanth/Factory-Vision-AI/pull/1
+
+## Purpose and limitations
+Educational object detection for people and PPE-related objects in factory/construction images. **Not a certified industrial safety system.** Detected PPE status is only a suggested alert requiring human review; it cannot establish a worker's actual PPE compliance or substitute for a trained safety inspector. In particular, missed detections mean unsafe situations may go unflagged.
+
+## Architecture and files
+- Fine-tuned Ultralytics YOLO11n; original starting checkpoint `yolo11n.pt`
+- Weights: `model.pt` (archive `checkpoints/ppe_yolo.pt`)
+- 10 classes: see `labels.json`
+- Training: 30 configured epochs, 640px image size, batch size 16; the training script uses a held-out PPE test split
+
+## Training data
+Construction Site Safety Image Dataset, attributed to Snehil Sanyal / Roboflow Universe, referenced by the source project: https://www.kaggle.com/datasets/snehilsanyal/construction-site-safety-image-dataset-roboflow . Source project lists CC BY 4.0 for this dataset; verify provenance and attribution requirements before public release. Dataset images are not included in this model package.
+
+## Held-out test performance (from the supplied Kaggle output)
+- Precision (bounding boxes): **0.8400**
+- Recall (bounding boxes): **0.6398**
+- mAP@50: **0.7053**
+- mAP@50:95: **0.3927**
+
+The numbers above are reported by the training artefact `ppe_test_metrics.json`, not independently reproduced. The reported test split has 82 images; class-specific breakdown is not included in this package. Do not confuse the separate pretrained YOLO COCO person baseline with this PPE result.
+
+## Example inference
+```python
+from ultralytics import YOLO
+model = YOLO('model.pt')
+results = model.predict('your_image.jpg', conf=0.25)
+for result in results:
+    print(result.boxes)
+```
+
+Note: object detections alone do not perform the project's person-to-helmet association. The original application implements an additional alert logic layer; consult its source and use human review.
+
+## Licensing and redistribution
+- Ultralytics YOLO11 and its fine-tuned checkpoints have AGPL-3.0 obligations by default, or require a suitable enterprise licence for uses incompatible with AGPL-3.0. See https://www.ultralytics.com/license
+- The source repository's MIT licence does **not** automatically license this trained checkpoint, pretrained weights, or source datasets.
+- The original developer has reportedly authorised publication; the publisher must still confirm the specific publication terms and applicable third-party rights.
+- **Do not assign an MIT/Apache model licence or switch to public visibility until licensing has been checked and recorded.**
+
+## Known limitations
+Domain shift, lighting, small objects, occlusion, crowded scenes, cameras and inference thresholds may reduce performance. False negatives are especially concerning for PPE alerting.

SOURCE.md

@@ -0,0 +1,8 @@
+Source: https://github.com/Theivanth/Factory-Vision-AI
+Original Project Developer and Model Author: Thevanth Rajamohan
+Technical Reviewer, Code Contributor and Hugging Face Publication Support: Swapin Vidya
+Publishing Organisation: TechSphere Consultancy Services Private Limited
+Hugging Face: https://huggingface.co/TechsphereAI
+Official Website: https://techsphereconsultancy.com/
+Swapin Vidya's merged contribution: https://github.com/Theivanth/Factory-Vision-AI/pull/1
+Model checkpoint SHA-256: f5e40e04ea7819a0b6cb9579b28c4e4f4227c7b6941d924c8b2beeee39f127c8

labels.json

@@ -0,0 +1,12 @@
+[
+  "Hardhat",
+  "Mask",
+  "NO-Hardhat",
+  "NO-Mask",
+  "NO-Safety Vest",
+  "Person",
+  "Safety Cone",
+  "Safety Vest",
+  "machinery",
+  "vehicle"
+]

metrics.json

@@ -0,0 +1,7 @@
+{
+  "metrics/precision(B)": 0.8400341147197301,
+  "metrics/recall(B)": 0.6397617577867072,
+  "metrics/mAP50(B)": 0.7053377987370262,
+  "metrics/mAP50-95(B)": 0.3926973285767887,
+  "fitness": 0.3926973285767887
+}

model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f5e40e04ea7819a0b6cb9579b28c4e4f4227c7b6941d924c8b2beeee39f127c8
+size 5507546