In this report, we present our champion solution for the NTIRE 2026 Challenge on Video Saliency Prediction held in conjunction with CVPR 2026. To exploit complementary inductive biases for video saliency, we propose Video Saliency with Adaptive Gated Experts (ViSAGE), a multi-expert ensemble framework. Each specialized decoder performs adaptive gating and modulation to refine spatio-temporal features. The complementary predictions from different experts are then fused at inference. ViSAGE thereby aggregates diverse inductive biases to capture complex spatio-temporal saliency cues in videos. On the Private Test set, ViSAGE ranked first on two out of four evaluation metrics, and outperformed most competing solutions on the other two metrics, demonstrating its effectiveness and generalization ability. Our code has been released at https://github.com/iLearn-Lab/CVPRW26-ViSAGE.
Figures & tables
Figure 1: Overview of the proposed ViSAGE framework for video saliency prediction. Our method utilizes a shared InternVideo2 backbone to extract multi-scale spatio-temporal features, which are then processed by two complementary expert decoders: (1) a temporal-modulated decoder with spatial priors, and (2) a multi-scale decoder with deep auxiliary supervision. The outputs of both experts are fused at inference to generate the final saliency map, fully leveraging diverse inductive biases for robust and generalizable video saliency estimation.
Method
Venue
CC
SIM
AUC Judd
NSS
ACLNet [ 27 ]
TPAMI
0.522
0.427
0.869
2.020
TASED-Net [ 16 ]
ICCV
0.557
0.461
0.881
2.160
STAViS [ 21 ]
CVPR
0.579
0.482
0.883
2.260
ViNet [ 11 ]
IROS
0.632
0.498
0.899
2.530
TSFP-Net [ 2 ]
arXiv
0.651
0.527
0.906
2.620
CASP-Net [ 32 ]
CVPR
0.655
0.543
0.906
2.610
Table 1: Cross-dataset evaluation on the public DIEM dataset. To evaluate broader generalization, our ViSAGE model is tested in a zero-shot setting (trained exclusively on the NTIRE split without any fine-tuning on DIEM).
Team
CC
SIM
AUC Judd
NSS
Baseline
0.410
0.408
0.691
1.305
NTR
0.693
0.579
0.845
2.989
SHU-MIPLab
0.717
0.593
0.880
2.701
AAM
0.748
0.579
0.815
3.100
Vertex
0.796
0.664
0.886
3.154
ARK_MMLAB
0.790
0.660
0.891
3.456
Table 2: Performance comparison on the NTIRE 2026 Video Saliency Prediction challenge using the official private test results. We report CC, SIM, AUC Judd, and NSS. Best results are in bold ; second-best are underlined .
Model Variant
CC
SIM
AUC Judd
NSS
Expert 1 Variations
Full Expert 1
0.794
0.634
0.916
3.225
w/o center bias
0.790
0.632
0.915
3.215
w/o FiLM
0.790
0.633
0.908
3.221
w/o temporal modulation
0.790
0.626
0.904
3.221
w/o LoRA (Stage 1 only)
0.783
0.624
0.914
3.180
Table 3: Component ablation study evaluated on a held-out 10% subset of the ECCV 2024 Video Saliency Prediction training data. We isolate the impact of core modules in each expert as well as the LoRA adaptation.
Variant
CC
SIM
AUC Judd
NSS
Logit fusion (two experts)
0.837
0.699
0.897
3.438
Saliency fusion (two experts)
0.837
0.698
0.898
3.420
Expert 2 only
0.830
0.693
0.900
3.390
Expert 1 only
0.832
0.693
0.897
3.412
Table 4: Ablation on ensemble design evaluated on a held-out 10% subset of the ECCV 2024 Video Saliency Prediction training data. Best results are in bold ; second-best are underlined .
Figure 2: Visualization of predicted saliency under different types of variations.
Department of Mathematics, University of Maryland, College Park, MD 20742, USA · Robert H. Smith School of Business, University of Maryland, College Park, MD 20742, USA