adskip_yolotune

Model Card for adskip_yolotune

This model is a fine-tuned version of the Ultralytics YOLOv11 model for computer vision tasks. It specializes in detecting the "Skip Ad" button that appears on YouTube video players, enabling automated ad skipping in custom applications.

Model Details

Model Description

• Developed by: github.com/calebtt
• Model type: Object detection (YOLOv11 fine-tune)
• License: AGPL-3.0 (same as the Ultralytics YOLO base model)
• Finetuned from model: Ultralytics/yolov11n.pt (default; supports resuming from previous best.pt if selected)

Model Sources

• Repository: https://github.com/calebtt/youtube_skip_button_fine_tuned_yolo
• Demo: See scripts and examples in the repository above

Uses

This model is intended for use in applications that require reliable detection of the YouTube "Skip Ad" button, such as automated ad-skipping tools, browser extensions, or screen automation scripts. Foreseeable users include developers building ad-management tools and end-users seeking enhanced YouTube viewing experiences. It primarily affects viewers by potentially reducing ad interruptions.

Bias, Risks, and Limitations

This model is fine-tuned specifically for detecting the semi-transparent pill-shaped "Skip Ad" button (labeled as "skip_button") in the bottom-right corner of YouTube video players.

• It may fail if YouTube changes the button's appearance, position, language, or style.
• Performance can degrade with varying screen resolutions, player themes (e.g., dark mode), overlays, or non-standard YouTube interfaces (e.g., embedded players).
• False positives may occur on similar UI elements (e.g., other buttons or overlays), leading to unintended actions.
• False negatives could miss the button during short ad windows.

Recommendations

Efforts were made during training to minimize false positives. For reliable use:

• Apply a high confidence threshold (e.g., >0.6; higher like >0.8 for further FP reduction).
• Use iou=0.70 and max_det=20 as in evaluation setup.
• Combine with temporal checks (e.g., button persistence across frames).
• Regularly retrain or update if YouTube UI changes.

How to Get Started with the Model

Load and use the model with the Ultralytics YOLO library.

The repository contains scripts for inference, further training, and testing: https://github.com/calebtt/youtube_skip_button_fine_tuned_yolo

Training Details

Training Data

• Base annotations from the Annotated UI Element Dataset for Desktop Environments: https://zenodo.org/records/10822752
• Augmented with local collections and iterative manual annotations of YouTube screenshots containing the "Skip Ad" button.

Training Procedure

Standard Ultralytics YOLO fine-tuning procedure was followed.

Preprocessing

Images were resized to 1280x1280 (script's default imgsz=1280). Standard YOLO augmentations applied (e.g., mosaic, flip, HSV adjustments).

Training Hyperparameters

• Training regime: Mixed precision (AMP enabled, typically fp16 on GPU)
• Epochs: 100
• Batch size: 8
• Patience (early stopping): 30 epochs
• Workers (data loaders): 8
• Save period (checkpoints): Every 5 epochs
• Image size: 1280
• Optimizer and learning rate: Ultralytics defaults (e.g., SGD with lr0=0.01)
• Supports resuming from the latest best.pt under runs/ if selected via --resume-best
• Specific details available in repository training scripts.

Speeds, Sizes, Times

Not specified in available information. Inference speed comparable to base YOLOv11 variant (e.g., real-time on modern GPUs). The training script's test inference mode measures time per image and total detections, emphasizing latency.

Evaluation

Testing Data, Factors & Metrics

Testing Data

Held-out split from custom YouTube screenshots and variations of the UI dataset.

Factors

• Variations in button appearance (language, theme, transparency)
• Screen resolutions and player sizes
• Presence of overlays or similar UI elements

Metrics

Standard YOLO object detection metrics:

• [email protected]
• [email protected]:0.95
• Precision, Recall (at various confidence thresholds) Post-training test inference (optional via --test) computes practical stats: images with ≥1 detection, total detections, avg confidence, inference time (using conf=0.60, iou=0.70, imgsz=1280, max_det=20).

Results

Specific quantitative results not publicly detailed. Qualitative performance shows reliable detection on standard YouTube ads with high-confidence thresholding.

Summary

The model achieves robust detection for its narrow task when used with appropriate thresholds, outperforming generic models on this UI element. Test mode emphasizes minimizing false positives via high conf thresholds.

Model Examination

No specific interpretability analysis provided.

Environmental Impact

Fine-tuning a pretrained YOLOv11 model typically requires modest compute.

• Hardware Type: GPU-accelerated (e.g., consumer NVIDIA GPU for training with batch=8)
• Hours used: Several hours to days, depending on dataset size and epochs
• Cloud Provider: Local or unspecified
• Compute Region: Not specified
• Carbon Emitted: Low (estimate <10 kg CO2eq using MLCO2 calculator for similar fine-tunes)

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

Technical Specifications

Model Architecture and Objective

YOLOv11 object detection architecture (single-stage detector with anchor-free head). Objective: Bounding box regression and classification for the single class "skip_button".

Compute Infrastructure

Hardware

• RTX 3060 12GB card
• GPU-accelerated (CUDA-compatible recommended for training/inference).

Software

Ultralytics YOLOv11 library (Python), PyTorch backend.

Citation

No formal paper published.

Model Card Authors

CalebTT

Model Card Contact

Contact via GitHub repository issues: https://github.com/calebtt/youtube_skip_button_fine_tuned_yolo