Livoa LogoLivoa
πŸš€ YOLOv8 Instance Segmentation Custom Architecture
Configuration: 10 classes | 128 prototypes | depthΓ—0.67 | widthΓ—0.75 | Input: 1024Γ—1024
πŸ“₯ Input


Image

1024Γ—1024Γ—3

🧩 Backbone (CSPDarknet)
Conv P1


64ch, s=2 β†’ 512Γ—512

Conv P2


128ch, s=2 β†’ 256Γ—256

C2f


128ch, n=3

Conv P3


256ch, s=2 β†’ 128Γ—128

C2f


256ch, n=6

Conv P4


512ch, s=2 β†’ 64Γ—64

C2f


512ch, n=6

Conv P5


1024ch, s=2 β†’ 32Γ—32

C2f


1024ch, n=3

SPPF


1024ch, k=5

🧡 Neck (PANet)
Top-down Fusion
Upsample Γ—2


P5β†’P4

Concat + C2f


512ch

Upsample Γ—2


P4β†’P3

Concat + C2f


256ch

Upsample Γ—2


P3β†’P2

Concat + C2f


128ch β†’ P2 Out

Bottom-up Re-aggregation
Conv + Concat


P2β†’P3 β†’ 256ch

Conv + Concat


P3β†’P4 β†’ 512ch

Conv + Concat


P4β†’P5 β†’ 1024ch

🧩 Seg Head

Multi-Scale Head


nc=10, nm=128

P2 P3 P4 P5

Prototype Gen


128 prototypes

Mask Coeffs


Per detection

πŸ—ƒ Output
Segmentation


1024Γ—1024

10 classes

Boxes + Masks

cnn yolo

by mahizha

0
0 uses