Learning Robust Anymodal Segmentor with Unimodal and Cross-modal Distillation

Update

[11/2024], pre-trained weights and evalutation code on MUSES are released.

[11/2024], draft paper is available at pdf.

git clone --
cd --
conda create -n anyseg python=3.9
conda activate anyseg
pip install -r requirements.txt

Used Datasets: MUSES / DELIVER

Method	F	E	L	FE	FL	EL	FEL	Mean	Weights
CMX	2.52	2.35	3.01	41.15	41.25	2.56	42.27	19.30	-
CMNeXt	3.50	2.77	2.64	6.63	10.28	3.14	46.66	10.80	-
MAGIC	43.22	2.68	22.95	43.51	49.05	22.98	49.02	33.34	-
Any2Seg	44.40	3.17	22.33	44.51	49.96	22.63	50.00	33.86	-
Ours	46.01	19.57	32.13	46.29	51.25	35.21	51.14	40.23	model

We appreciate the previous open-source works: DELIVER / SegFormer / MUSES