Small object detection is one of the most challenging and non-negligible fields in computer vision. Inspired by the location–focus–identification process of the avian visual system, we present our location-focused small-object-detection algorithm for video or image sequence, OTNet. The model contains three modules corresponding to the forms of saliency, which drive the strongest response of OT to calculate the saliency map. The three modules are responsible for temporal–spatial feature extraction, spatial feature extraction and memory matching, respectively. We tested our model on the AU-AIR dataset and achieved up to 97.95% recall rate, 85.73% precision rate and 89.94 F1 score with a lower computational complexity. Our model is also able to work as a plugin module for other object detection models to improve their performance in bird-view images, especially for detecting smaller objects. We managed to improve the detection performance by up to 40.01%. The results show that our model performs well on the common metrics on detection, while simulating visual information processing for object localization of the avian brain.
Animals detect targets using a variety of visual cues, with the visual salience of these cues determining which environmental features receive priority attention and further processing. Surround modulation plays a crucial role in generating visual saliency, which has been extensively studied in avian tectal neurons. Recent work has reported that the suppression of tectal neurons induced by motion contrasting stimulus is stronger than that by luminance contrasting stimulus. However, the underlying mechanism remains poorly understood. In this study, we built a computational model (called Generalized Linear-Dynamic Modulation) which incorporates independent nonlinear tuning mechanisms for excitatory and inhibitory inputs. This model aims to describe how tectal neurons encode contrasting stimuli. The results showed that: (1) The dynamic nonlinear integration structure substantially improved the accuracy (significant difference (p < 0.001, paired t-test) in the goodness of fit between the two models) of the predicted responses to contrasting stimuli, verifying the nonlinear processing performed by tectal neurons. (2) The modulation difference between luminance and motion contrasting stimuli emerged from the predicted response by the full model but not by that with only excitatory synaptic input (spatial luminance: 89 ± 2.8% (GL_DM) vs. 87 ± 2.1% (GL_DMexc); motion contrasting stimuli: 87 ± 1.7% (GL_DM) vs. 83 ± 2.2% (GL_DMexc)). These results validate the proposed model and further suggest the role of dynamic nonlinear spatial integrations in contextual visual information processing, especially in spatial integration, which is important for object detection performed by birds.
In the research area of anomaly detection, novel and promising methods are frequently developed. However, most existing studies exclusively focus on the detection task only and ignore the interpretability of the underlying models as well as their detection results. Nevertheless, anomaly interpretation, which aims to provide explanation of why specific data instances are identified as anomalies, is an equally important task in many real-world applications. In this work, we propose a novel framework which synergizes several machine learning and data mining techniques to automatically learn invariant rules that are consistently satisfied in a given dataset. The learned invariant rules can provide explicit explanation of anomaly detection results in the inference phase and thus are extremely useful for subsequent decision-making regarding reported anomalies. Furthermore, our empirical evaluation shows that the proposed method can also achieve comparable or even better performance in terms of AUC and partial AUC on public benchmark datasets across various application domains compared with start-of-the-art anomaly detection models.
Most neural decoding methods prioritize predictive accuracy but provide limited means to test hypotheses about latent representational geometry. Guided by projection geometry, we recorded neural activity from the ENTO–MVL pathway of five pigeons performing matched 2D rotation, 2D stretch/compression, and 3D view-rotation tasks. Preliminary response analysis identified two groups with different 2D transformation preferences, both responsive during 3D view rotation, motivating a hypothesis of joint engagement. We introduce ProMaC, a product-manifold contrastive framework with specialist encoders and a prespecified S1 × coordinate prior, to test this hypothesis through cross-transformation generalization. Trained exclusively on 2D tasks, frozen branches recovered silhouette-derived projected rotation and log-scale descriptors from held-out 3D trials and aligned 2D and 3D centroids, outperforming matched controls (Holm-adjusted p < 0.001). Fixed-time-bin analyses supported correspondence beyond shared temporal progress, and leave-one-pigeon-out sensitivity analyses retained positive model advantages. Auxiliary analyses showed factor selectivity and greater projected-descriptor accuracy for direct ProMaC-Joint predictions than for angle-mediated predictions, without establishing information beyond physical angle. Persistent homology supported a stable loop but did not establish a complete product manifold. These findings support the generalization of factor-specific, 2D-derived coordinates along the tested one-parameter 3D view-rotation trajectories.