Table 3 Compare with methods on the pickleball dataset. In this experiment, we re-implement the SOTA methods on this dataset.

From: Curvelet-enhanced transformer architecture for blurred action fine-grained detection

Methods

Backbone

Frame size

AP

AP75

APM

APL

SwinT

SwinT

256 × 192

0.755

0.795

0.723

0.801

SimpleBaseline

ResNet-50

256 × 192

0.748

0.768

0.704

0.781

DERK

HRNet-W32

512 × 512

0.751

0.776

0.743

0.802

HigherHRNet + SWAHR

HRNet-W32

512 × 512

0.786

0.799

0.771

0.793

AECA

ResNet-18

384 × 288

0.769

0.787

0.762

0.810

EBA

ResNet-18

256 × 255

0.799

0.821

0.756

0.809

TokenPose

TokenPose-L/D24

256 × 192

0.796

0.819

0.751

0.810

RIFormer

HRFormer-B

256 × 192

0.801

0.818

0.764

0.813

MCTN

DETR

256 × 192

0.813

0.837

0.775

0.822

MCTN

RT-DETRv3

256 × 192

0.822

0.841

0.778

0.846

MCTN

DETR

384 × 288

0.816

0.837

0.777

0.827

MCTN

RT-DETRv3

384 × 288

0.822

0.846

0.780

0.844