Model families¶

All five public models take RGB and projected LiDAR image tensors and produce dense semantic class logits. They differ in where the two streams meet and how features become predictions.

Diagram: comparison of the five models by encoder, fusion point and prediction mechanism.

Model	Core distinction	Read more
CLFT	Global ViT features are reassembled and fused during decoding.	CLFT
CLFTv2	Hierarchical Swin features feed a top-down fusion pyramid.	CLFTv2
MaskFormerFusion	Fused features become class-labelled masks through learned queries.	MaskFormerFusion
Mask2FormerFusion	A multi-scale pixel decoder feeds masked-attention queries.	Mask2FormerFusion
DeepLabV3+	Two convolutional branches fuse after ASPP and decoding.	DeepLabV3+

The figures describe the implementations in this library. Each model page links to the relevant research paper and calls out local fusion adaptations. For constructors, inputs and training, start with the library API.