BIAS: A Biologically Inspired Algorithm for Video Saliency Detection

Zhao-ji Zhang; Ya-tang Li

cs.CVApr 10, 2026

BIAS: A Biologically Inspired Algorithm for Video Saliency Detection

Authors: Zhao-ji Zhang, Ya-tang Li

Organizations: Academy for Advanced Interdisciplinary Studies, Peking University, Beijing 100871, China · Chinese Institute for Brain Research, Beijing (CIBR), Beijing 102206, China

Abstract

We present BIAS, a fast, biologically inspired model for dynamic visual saliency detection in continuous video streams. Building on the Itti--Koch framework, BIAS incorporates a retina-inspired motion detector to extract temporal features, enabling the generation of saliency maps that integrate both static and motion information. Foci of attention (FOAs) are identified using a greedy multi-Gaussian peak-fitting algorithm that balances winner-take-all competition with information maximization. BIAS detects salient regions with millisecond-scale latency and outperforms heuristic-based approaches and several deep-learning models on the DHF1K dataset, particularly in videos dominated by bottom-up attention. Applied to traffic accident analysis, BIAS demonstrates strong real-world utility, achieving state-of-the-art performance in cause-effect recognition and anticipating accidents up to 0.72 seconds before manual annotation with reliable accuracy. Overall, BIAS bridges biological plausibility and computational efficiency to achieve interpretable, high-speed dynamic saliency detection.

Figures & tables

Panel of Figure 1 — Fig. 1: (a) General architecture of BIAS. (b) Comparison of center–surround computation between Itti’s original model (top) and BIAS (bottom). (c) Motion computation inspired by direction-selective cells in the fly retina.

Weight(a,b,c)	CC	SIM	s-AUC	NSS	AUC-J
1.0, 0.2, 0.2	0.297	0.171	0.578	1.57	0.822
1.0, 0.3, 0.2	0.300	0.193	0.578	1.59	0.819
1.0, 0.2, 0.3	0.301	0.178	0.576	1.60	0.821
1.0, 0.3, 0.3	0.307	0.183	0.581	1.63	0.828
1.0, 0.5, 0.3	0.299	0.190	0.579	1.58	0.818
1.0, 0.3, 0.5	0.297	0.173	0.574	1.60	0.820

TABLE I: Quantitative comparison of different weights on DHF1K

Fig. 2: Predicted saliency maps on an example video clip from the DHF1K dataset. From left to right: original frames, image-based saliency maps, motion-based saliency maps, combined master saliency maps, predicted fixations with GWTA, and human fixation ground truth. SM = saliency map, GWTA = Gaussian winner-take-all.

Method	AUC-J	SIM	s-AUC	CC	NSS	Size (MB)	DLM	Time ⋆ (s)
SalFoM	0.922	0.421	0.735	0.569	3.354	1574	✓	90
TMFI	0.915	0.407	0.731	0.546	3.146	234	✓	4.95
THTD-Net	0.915	0.406	0.730	0.548	3.139	220	✓	12
STSANet	0.912	0.383	0.723	0.529	3.010	643	✓	5.25
TSFP-Net	0.912	0.392	0.723	0.517	2.967	58.4	✓	1.65
VSFT	0.911	0.411	0.720	0.518	2.977	71.4	✓	6

TABLE II: Quantitative comparison of different methods on DHF1K

Fig. 3: Comparison of performance and runtime between BIAS and other models.

Method	AUC-J	SIM	sAUC	CC	NSS
BIAS (Bottom-Up) ( $c\in\{2\}$ , $\delta\in\{1\}$ )	0.869	0.237	0.577	0.358	1.851
BIAS ( $c\in\{2\}$ , $\delta\in\{4\}$ )	0.849	0.221	0.561	0.323	1.670
BIAS ( $c\in\{2\}$ , $\delta\in\{1\}$ )	0.849	0.223	0.560	0.323	1.670
w/ Lévy flight, w/o GWTA ( $c\in\{2\}$ , $\delta\in\{4\}$ )	0.846	0.237	0.559	0.311	1.626
w/ EWMA, w/o GWTA ( $c\in\{2\}$ , $\delta\in\{4\}$ )	0.835	0.184	0.582	0.310	1.632
w/ EWMA, w/o GWTA ( $c\in\{2\}$ , $\delta\in\{1\}$ )	0.828	0.184	0.582	0.307	1.625

TABLE III: Ablation study on DHF1K

Panel of Figure 7 — Fig. 4: (a) Correlation coefficients for different center–delta pairs. (b) Performance comparison with and without EWMA. (c) Performance comparison with and without GWTA across different center–delta pairs.

Panel of Figure 8 — Fig. 5: (a) From left to right: original frames, human fixation ground truth, master saliency maps, predicted fixations with GWTA, and GWTA predictions overlaid on the original frames. The top two rows show examples of well-predicted saliency maps; the middle two rows show examples of moderate performance; the bottom two rows show examples of low performance. (b) Performance metrics for videos with distinct COCO-labeled contents. *** $p<0.001$ , ** $p<0.01$ , * $p<0.05$ . (c) Correlation of mean performance metrics with the average numbers of YOLO-detected objects for each video. $r$ is the Pearson correlation coefficient. (d–e) Performance metrics of predicted saliency maps in different video categories. The number of objects and people, camera speed, content speed, and capture time were obtained from annotation data of the DHF1K dataset. The number of clusters was obtained using the DBSCAN clustering algorithm.

Fig. 6: The input is a sequence of predicted saliency maps. SPARK-ResNet extracts only spatial features, which are then processed via convolution and pooling before being fed into the MS-TCN for temporal segmentation.

Model	Kinetics-I3D	BIAS-SparK	Itti-SparK [ 8 ]	SalFoM-Spark [ 62 ]
Cause IoU $\geq 0.1$	0.538	0.513	0.172	0.455
Cause IoU $\geq 0.3$	0.415	0.323	0.068	0.293
Cause IoU $\geq 0.5$	0.201	0.143	0.007	0.111
Cause IoU $\geq 0.7$	0.061	0.054	0.003	0.021
Effect IoU $\geq 0.1$	0.638	0.796	0.462	0.648
Effect IoU $\geq 0.3$	0.462	0.606	0.293	0.483

TABLE IV: Performance on traffic accident causality recognition

Model	IoU $\geq$ 0.1	IoU $\geq$ 0.3	IoU $\geq$ 0.5	IoU $\geq$ 0.7	mTTA (s)	mLT (s)
BIAS-SparK	0.89	0.78	0.57	0.24	2.26	0.72
Kinetics-I3D	0.83	0.73	0.53	0.26	2.05	0.08
Itti-SparK	0.58	0.42	0.19	0.06	0.83	-0.62
SalFoM-SparK	0.77	0.68	0.44	0.16	1.58	-0.17
DRIVE ⋆	0.93	0.30	0.05	0.02	8.01	6.73
rand DRIVE †	0.93	0.37	0.11	0.03	7.04	5.73

TABLE V: Traffic Accident Prediction IoU and Anticipation

Panel of Figure 12 — Fig. 7: Comparison of predicted times across different models. (a) Predicted time relative to cause onset and effect offset. For cause onset, Kinetics-I3D: $-0.08\pm 3.05$ , $p=0.65$ (one-sample $t$ -test against 0); BIAS-SparK: $-0.72\pm 2.32$ , $p=1.1\times 10^{-6}$ ; Itti-SparK: $0.63\pm 3.24$ ; SalFoM-SparK: $0.17\pm 2.16$ . For effect offset, Kinetics-I3D: $0.10\pm 3.02$ ; BIAS-SparK: $0.61\pm 2.69$ ; Itti-SparK: $-0.69\pm 3.17$ ; SalFoM-SparK: $-0.60\pm 2.72$ . (b) Predicted TTAs by different methods.

Module	2-Layer Conv	Conv + AvgPool	AvgPool Only
Cause IoU $>0.1$	0.391	0.513	0.204
Cause IoU $>0.3$	0.237	0.323	0.086
Cause IoU $>0.5$	0.104	0.143	0.028
Cause IoU $>0.7$	0.025	0.054	0.007
Effect IoU $>0.1$	0.556	0.796	0.243
Effect IoU $>0.3$	0.394	0.606	0.129

TABLE A1: Ablation study on bottleneck module design. All hidden dimensions = 128.

Hidden Dim	32	64	128	256
Cause IoU $>0.1$	0.502	0.455	0.513	0.455
Cause IoU $>0.3$	0.333	0.304	0.323	0.287
Cause IoU $>0.5$	0.136	0.125	0.143	0.147
Cause IoU $>0.7$	0.035	0.025	0.054	0.039
Effect IoU $>0.1$	0.724	0.713	0.796	0.742
Effect IoU $>0.3$	0.584	0.559	0.606	0.563

TABLE A2: Effect of hidden dimension size in the 1-layer convolutional bottleneck.

Video Saliency Prediction

See all 12 papers

Moving Forward with Video Saliency: A New Dataset and Benchmark where Motion Matters
Oct 2, 2026Susmit Agrawal, Rebecca Wanner, Juliane Verwiebe +3Video Saliency Prediction Visual Saliency
Tübingen AI Center, University of Tübingen · IMPRS-IS · Optocycle +2
ViSAGE @ NTIRE 2026 Challenge on Video Saliency Prediction
Apr 9, 2026Kun Wang, Yupeng Hu, Zhiran Li +3Video Saliency Prediction Visual Saliency
Shandong University · Harbin Institute of Technology (Shenzhen) · City University of Hong Kong +1
TDSal: Task-Based Top-Down Saliency Prediction Model
Jul 10, 2026Can Mizrakli, Tolga K. CapinVisual Attention Salient Object Detection
Karlsruhe Institute of Technology, Germany · TED University, Türkiye

Oct 2, 2026J/K move · Enter open · S save

Moving Forward with Video Saliency: A New Dataset and Benchmark where Motion Matters

Susmit Agrawal, Rebecca Wanner, Juliane Verwiebe, Matthias Tangemann+2

Tübingen AI Center, University of Tübingen · IMPRS-IS · Optocycle+2

Video Saliency Prediction Visual Saliency Saliency Map Evaluation

Open paper AskPDF arXiv

Video saliency prediction is inherently harder to model than static image saliency due to the additional temporal dimension. Video saliency benchmarks rest on the premise that predicting gaze on video requires utilizing temporal activity distributed across frames. Prior work has challenged this, showing that static baselines recover a significant fraction of the explainable gaze information on LEDOV, a popular video saliency dataset, and that video saliency models fail in the same places as this static baseline. We verify that this diagnosis still stands: under a more capable gold standard than the original analysis, and an updated panel of recent architectures, the strongest temporal architecture in the panel still does not substantially improve over a fine-tuned static baseline. However, it remains unclear whether the marginal gain reflects limitations of current temporal architectures or a lack of temporal patterns in the benchmark itself. We introduce SalTempto, a video saliency benchmark with greater dynamism: 224 clips of highly dynamic content, sourced from the HACS-Segments dataset so that each clip contains an event together with its lead-up and aftermath, with gaze recordings from up to 16 subjects and a training split for adapting pretrained models. On SalTempto, the static baseline recovers only about 13% of the headroom above the centerbias, against more than half on LEDOV. A fine-tuned temporal architecture shows a substantial gain in performance over the static baseline, indicating that it does capture meaningfully more temporal information, which LEDOV fails to measure. Yet, even this SoTA model still leaves nearly half of SalTempto's headroom unexplained, indicating room for improvement in video saliency modelling. Examination of SalTempto also lets us describe human tendencies that models miss. SalTempto link: https://huggingface.co/datasets/bethgelab/video_saliency.

Abstract

Video Saliency Prediction

Explore similar work

Moving Forward with Video Saliency: A New Dataset and Benchmark where Motion Matters

ViSAGE @ NTIRE 2026 Challenge on Video Saliency Prediction

TDSal: Task-Based Top-Down Saliency Prediction Model