ilMassy/advertising-panel-segmentation

Advertising Panel Segmentation — SegFormer

Semantic segmentation models for detecting advertising LED boards in sports broadcasting footage.

Dataset

• 967 images at 1920x1080 from sports highlights
• Manually annotated via CVAT.ai with polygon annotations
• Binary segmentation: background (0) vs advertising board (1)
• Split: Train 672 / Val 149 / Test 146

Models

Experiment	Model	Augmentation	mIoU	Board IoU	Dice	Precision	Recall
Exp0	SegFormer-B0	Standard	87.15%	76.28%	86.55%	84.13%	89.10%
Exp1	SegFormer-B1	Standard	84.29%	71.12%	83.12%	79.26%	87.39%
Exp2	SegFormer-B1	Sport-specific	87.26%	76.45%	86.66%	85.76%	87.57%
Exp3-opt1	SegFormer-B1	Sport-specific reduced + LR 3e-5 + channels=512	84.58%	71.68%	83.51%	78.66%	88.99%
Exp3-opt2	SegFormer-B1	Sport-specific reduced + LR 3e-5 + channels=256	84.92%	72.26%	83.90%	80.40%	87.71%
Exp3-opt3	SegFormer-B1	Sport-specific + drop_path=0.15	86.42%	74.99%	85.71%	82.25%	89.47%

Best model: Exp2 - SegFormer-B1 Augmented (models/exp2_segformer_b1_augmented/best_mIoU_iter_14000.pth)

Detailed Results

Exp0 - SegFormer-B0 Baseline

• Best checkpoint: best_mIoU_iter_18000.pth
• mIoU: 87.15% | Board IoU: 76.28% | Dice: 86.55% | Precision: 84.13% | Recall: 89.10%

Exp1 - SegFormer-B1 Standard

• Best checkpoint: best_mIoU_iter_10000.pth
• mIoU: 84.29% | Board IoU: 71.12% | Dice: 83.12% | Precision: 79.26% | Recall: 87.39%

Exp2 - SegFormer-B1 Augmented ⭐ Best

• Best checkpoint: best_mIoU_iter_14000.pth
• mIoU: 87.26% | Board IoU: 76.45% | Dice: 86.66% | Precision: 85.76% | Recall: 87.57%

Exp3 - SegFormer-B1 Optimized1

• Best checkpoint: best_mIoU_iter_14000.pth
• mIoU: 84.58% | Board IoU: 71.68% | Dice: 83.51% | Precision: 78.66% | Recall: 88.99%

Exp3 - SegFormer-B1 Optimized2

• Best checkpoint: best_mIoU_iter_12000.pth
• mIoU: 84.92% | Board IoU: 72.26% | Dice: 83.90% | Precision: 80.40% | Recall: 87.71%

Exp3 - SegFormer-B1 Optimized3

• Best checkpoint: best_mIoU_iter_18000.pth
• mIoU: 86.42% | Board IoU: 74.99% | Dice: 85.71% | Precision: 82.25% | Recall: 89.47%

Framework

• PyTorch 2.4.1
• MMSegmentation 1.2.2
• mmcv 2.2.0

Repository Structure

models/
  exp0_segformer_b0_baseline/best_mIoU_iter_18000.pth
  exp1_segformer_b1_standard/best_mIoU_iter_10000.pth
  exp2_segformer_b1_augmented/best_mIoU_iter_14000.pth
  exp3_segformer_b1_optimized/best_mIoU_iter_14000.pth
  exp3_segformer_b1_optimized2/best_mIoU_iter_12000.pth
  exp3_segformer_b1_optimized3/best_mIoU_iter_18000.pth
dataset/
  processed.zip  (967 images + masks, train/val/test split)

GitHub

https://github.com/ilMassy/advertising-panel-segmentation