| --- |
| license: apache-2.0 |
| pipeline_tag: image-segmentation |
| --- |
| # Model Card for Model ID |
|
|
| <!-- YOLO‑based single‑class photocard detector trained on 156 hand‑annotated images. Designed to detect selfie‑style photocards (~55×85 mm) in real‑world scenes. Achieved 97.4% mAP@50, 94.4% precision, 94.8% recall --> |
|
|
| This modelcard aims to be a base template for new models. It has been generated using [this raw template](https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/templates/modelcard_template.md?plain=1). |
|
|
| ## Model Details |
| Architecture: YOLO object detector (Roboflow hosted training) |
|
|
| Task: Single‑class object detection |
|
|
| Output: Bounding boxes around photocards |
|
|
| Intended use: Stage 1 of a photocard processing pipeline (cataloging, classification, pricing) |
|
|
| https://app.roboflow.com/priyas-workspace-nntro/bdata-497-final-project/models/bdata-497-final-project/1 |
| ### Model Description |
|
|
| - **Developed by:** Priya Rasal |
| - **Model type:** YOLO object detector (bounding box detection) |
| - **Finetuned from model:** Pretrained YOLO backbone provided by Roboflow |
|
|
| ## Uses |
| ### Direct Use |
| Detecting photocards in images |
|
|
| Cropping photocards for downstream tasks |
|
|
| Preprocessing for classification, cataloging, or pricing systems |
|
|
| Identifying photocards in cluttered or real‑world scenes |
| ### Downstream Use |
| Photocard classification (e.g., identifying the idol or version) |
|
|
| Marketplace automation (auto‑detecting cards in listings) |
|
|
| Inventory management for collectors |
|
|
| Dataset creation for future models |
| ### Out-of-Scope Use |
| Detecting people or faces |
|
|
| Identifying which idol is on the photocard |
|
|
| Detecting non‑photocard rectangular objects (phones, books, receipts) |
|
|
| Any biometric or identity recognition |
| ## Bias, Risks, and Limitations |
| May detect rectangular objects as photocards in rare cases |
|
|
| May miss photocards with extreme glare, reflections, or occlusion |
|
|
| Performance depends on lighting and background diversity |
|
|
| Only trained on 156 images — limited exposure to rare edge cases |
|
|
| Not suitable for identity recognition or personal data analysis |
| ### Recommendations |
| Use a confidence threshold appropriate for your application |
|
|
| Validate predictions manually in high‑stakes use cases |
|
|
| Retrain or fine‑tune with more diverse data for improved robustness |
| ## How to Get Started with the Model |
| Use the code below to get started with the model. |
|
|
| from ultralytics import YOLO |
|
|
| model = YOLO("path/to/your/model.pt") |
| results = model("your_image.jpg") |
| results.show() |
| ## Training Details |
| ### Training Data |
| Total images: 156 |
| |
| Source: Personal photocard collection + marketplace images |
| |
| Annotation method: Foundation model auto‑labeling + manual correction |
| |
| Class definition: Photocard = selfie‑style, rectangular, ~55×85 mm |
| |
| Excluded: postcards, album inclusions, sleeves without cards, binder pockets |
| ### Training Procedure |
| Training platform: Roboflow Hosted Training |
| |
| Training approach: Transfer learning from pretrained YOLO backbone |
| |
| Image size: 640×640 |
| |
| Epochs: ~50–100 (auto‑selected) |
| |
| Batch size: Auto‑selected |
| |
| Learning rate: Warmup + cosine decay |
| |
| Optimizer: AdamW or SGD (Roboflow default) |
| |
| Precision: Mixed precision (fp16) |
| #### Speeds, Sizes, Times [optional] |
| Training time: ~10–20 minutes (depending on GPU) |
| |
| Model size: Depends on YOLO variant exported (typically 10–40 MB) |
| ## Evaluation |
| ### Testing Data, Factors & Metrics |
| #### Testing Data |
| Validation and test splits generated automatically by Roboflow (70/20/10) |
| #### Factors |
| Lighting variation |
| |
| Background clutter |
| |
| Card orientation |
| |
| Sleeve reflections |
| |
| Flash glare |
| |
| Occluding shadows |
| |
| Overlaps |
| #### Metrics |
| mAP@50: 97.4% |
| |
| Precision: 94.4% |
| |
| Recall: 94.8% |
| ### Results |
| #### Summary |
| The model demonstrates strong localization accuracy and generalization across diverse real‑world scenes. |
| |
| High precision and recall indicate low false positives and low missed detections. |
| |
| Photocard class annotated/identified 1023 times |
| ## Model Card Author |
| |
| Priya Rasal |
