ComPtr

ComPtr: Towards Diverse Bi-source Dense Prediction Tasks via A Simple yet General Complementary Transformer

@ARTICLE{ComPtr,
    title={ComPtr: Towards Diverse Bi-source Dense Prediction Tasks via A Simple yet General Complementary Transformer},
    author={Youwei Pang and Xiaoqi Zhao and Lihe Zhang and Huchuan Lu},
    journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
    year    = {2025},
    doi     = {10.1109/TPAMI.2025.3578494},
}

Results

LEVIR-CD

Based on CAVER.

Methods	Backbone	F1 $\uparrow$	IoU $\uparrow$
TTP	SAM (ViT-L/16)	92.10	85.60
ComPtr-T	Swin-T	92.62	86.25
ComPtr-B	Swin-B	92.70	86.39

SYSU-CD

Based on CAVER.

Methods	Backbone	F1 $\uparrow$	IoU $\uparrow$
TransY-Net	Swin-B	82.84	70.71
ComPtr-T	Swin-T	83.43	71.57
ComPtr-B	Swin-B	83.18	71.21

RGB-T CC

Based on MMCC.

Methods	Backbone	GAME $_{0}\downarrow$	RMSE $\downarrow$
MMCC	PvTv2-B3	10.90	18.79
ComPtr-T	Swin-T	10.52	18.48
ComPtr-B	Swin-B	11.82	20.75

SUN-RGBD

Based on CMX.

Methods	Backbone	mIoU $\uparrow$
DFormer-L	DFormer-L	52.5
ComPtr-T	Swin-T	48.9
ComPtr-B	Swin-B	52.8

RGB-D/T SOD

Based on CAVER.

See Releases

(S $\_m$ & F $^{\omega}\_{\beta}$)

Methods	Backbone	DUTLF-Depth $\uparrow$	NJUD $\uparrow$	NLPR $\uparrow$	SIP $\uparrow$	STEREO1000 $\uparrow$
HRTransNet	HRFormer-B	95.07 & 94.93	93.33 & 92.00	94.18 & 91.70	90.90 & 89.56	92.12 & 89.37
ComPtr-T	Swin-T	94.37 & 92.17	92.89 & 89.48	93.56 & 88.66	91.34 & 87.96	91.99 & 87.45
ComPtr-B	Swin-B	95.87 & 95.14	93.96 & 91.95	94.29 & 90.73	91.53 & 89.48	93.29 & 90.29

Methods	Backbone	VT821 $\uparrow$	VT1000 $\uparrow$	VT5000 $\uparrow$
HRTransNet	HRFormer-B	90.59 & 84.92	93.79 & 91.27	91.23 & 86.99
ComPtr-T	Swin-T	90.52 & 83.32	94.17 & 89.82	90.71 & 83.10
ComPtr-B	Swin-B	92.37 & 87.48	95.12 & 92.66	92.84 & 88.34