Table 4 Experimental Results of the Algorithm on our datasets.

From: Fast moving table tennis ball tracking algorithm based on graph neural network

Algorithm Name

Train&Test Set

Validation Set

Target

Relationship

Target

Relationship

Pre.

Rec.

F1.

Pre.

Rec.

F1.

Pre.

Rec.

F1.

Pre.

Rec.

F1.

\(\mathrm{YOLOv8(conf=0.01)}^{1}\)

0.853

0.744

0.795

N/A

N/A

N/A

0.551

0.738

0.631

N/A

N/A

N/A

\(\mathrm{YOLOv8(conf=0.50)}^{1}\)

0.924

0.647

0.761

N/A

N/A

N/A

0.838

0.576

0.683

N/A

N/A

N/A

\(\mathrm{YOLOv8(0.01-0.50)}^{2}\)

0.908

0.685

0.781

N/A

N/A

N/A

0.773

0.641

0.701

N/A

N/A

N/A

\(\mathrm{SOT(0.05-0.95)}^{3}\)

N/A

N/A

N/A

0.832

0.705

0.763

N/A

N/A

N/A

0.792

0.483

0.6

\(\mathrm{GMP(0.40-0.60)}^{4}\)

0.88

0.946

0.912

0.865

0.927

0.895

0.981

0.946

0.963

0.912

0.938

0.925

\(\mathrm{GAT(0.40-0.60)}^{4}\)

0.849

0.907

0.877

0.837

0.883

0.859

0.875

0.863

0.869

0.834

0.813

0.823

\(\mathrm{T-FORT}^{5}\)

0.782

0.856

0.818

0.767

0.838

0.801

N/A

N/A

N/A

N/A

N/A

N/A

\(\textrm{FMO}^{5}\)

0.711

0.592

0.646

N/A

N/A

N/A

N/A

N/A

N/A

N/A

N/A

N/A

  1. Pre., Rec., and F1. are abbreviations for Precision, Recall, and F1-Score, respectively.
  2. [1] conf=0.01 Means we set the confidence threshold of detection in YOLOv8 as 0.01.
  3. [2] The average result of YOLOv8 when threshold are set as 0.01,0.05,0.10,...,0.45,0.50. We select these thersholds because YOLOv8’s F1-Score decrease sharply when threshold < 0.5 while keep consistent between 0.01 and 0.50.
  4. [3] The average result of SOT when threshold are set as 0.05,0.10,...,0.90,0.95.
  5. [4] The average result of GAT or GMP when threshold are set as 0.40,0.45,...,0.60.
  6. [5] When considering these algorithms, we only included the videos with algorithm results in their paper.
Back to article page