CLFTv2¶

Hierarchical features, fused through a top-down pyramid. CLFTv2 uses a shared Swin V2 backbone to encode RGB and projected LiDAR separately. At each scale it projects and combines their feature maps with the previous, coarser fusion stage.

Diagram: CLFTv2 takes RGB and projected LiDAR through a shared Swin V2 hierarchy, per-scale projection and residual fusion, and a segmentation head.

The highlighted card marks per-scale fusion. The backbone already returns spatial feature maps, so this path has no ViT token reassembly step.

Why choose it¶

Compared with CLFT, the main change is the hierarchical shifted-window backbone and its native multi-scale maps. The default library strategy is residual_average; the underlying fusion module also implements other strategies for experiments. The public class returns dense class logits.

from visin_fusion.models import CLFTv2

model = CLFTv2(num_classes=4, mode="cross_fusion", fusion_strategy="residual_average")
logits = model(rgb, lidar)

Use mode="rgb" or mode="lidar" to isolate one stream. See the library API for tensor shapes and training setup.

Research references¶

Tahves, Bellone & Sell, CLFTv2: Efficient Camera-LiDAR Fusion for Semantic Segmentation via Hierarchical Feature Pyramids, 2026. The CLFTv2 research architecture.
Liu et al., Swin Transformer V2: Scaling Up Capacity and Resolution, 2021. The backbone family used by the default library model.