Multimodal Instance Segmentation

Multimodal instance segmentation extends classical instance segmentation — which assigns a per-pixel mask and a class label to every individual object in an image — by incorporating complementary sensor streams such as depth maps, LiDAR point clouds, or infrared frames. Fusing these modalities helps the model handle ambiguous appearances, low light, and occlusion that trip up RGB-only systems.