Back to Projects

Soccer Vision Research

4 SOTA models. 1 modular pipeline. Zero-shot player identification.

View on GitHub
RF-DETRSAM2ByteTrackSigLIP

Drag to Compare

Raw Input
#1
#7
#3
#4
#8
#6
#2
#10
#9
#11
#1
Pipeline Output

Pipeline Flow

VideoInput
DetectionRF-DETR
SegmentationSAM2
TrackingByteTrack
ClassificationSigLIP

The Models

RF-DETR

Real-time detection transformer for player localization

97%+precision

SAM2

Segment Anything Model for pixel-perfect player masks

30 FPSreal-time

ByteTrack

Multi-object tracking with occlusion recovery

847+frames tracked

SigLIP

Vision-language model for zero-shot team classification

0-shotno training needed
97%+
Detection Precision
30
FPS Processing
10+
Configs Tested
500+
Video Clips