Frame Importance Voting for Video Scene Classification

Multi-Modal Video Action Recognition with Learnable Frame Pruning via Temporal Token Scoring

VideoSEMA: a scalable and efficient Mamba-like attention for video understanding

It is proved that the computationally cheaper split space-time attention is equivalent to full space-time attention and is promising to extend VideoSEMA to longer videos with a dilated/sparse temporal attention.

Towards an Efficient and Unified Strategy for Video Understanding Applications

Understanding long-range videos remains a key challenge in computer vision due to high temporal redundancy and computational burden. Despite strong performance of recent models, they are constrained in terms of scalability and generalization when applied to longer video sequences. In this work, we present Keyframe-based Spatio-Temporal Adaptive Representation (K-STAR), a redundancy-aware video summarization framework designed to generate compact and semantically rich representations that are effective in downstream tasks. The proposed method jointly models appearance and motion cues while filtering redundant frames. Importantly, it preserves critical temporal transitions while significantly reducing the number of processed frames. Additionally, each key frame is encoded using object, scene, and background-aware prompts, enabling richer semantic representation. Evaluated on the UCF-101 dataset, K-STAR achieves Top-1 accuracy of 93.06% and Top-5 accuracy of $\mathbf{9 8. 7 3 \%}$, with $\mathbf{5 6} \times$ frame reduction and $\mathbf{1 1. 5} \times$ faster inference, demonstrating competitive performance with substantially improved efficiency.

CNN-ViT Hybrid Architecture With Time-Color Coding for Classifying Behavioral Events in Video Sequences

Classifying behavioral events in video sequences requires modeling morphology and motion under severe data constraints. This paper presents a hybrid two-stream CNN-ViT architecture combining EfficientNetB4 for local spatial features with a Vision Transformer branch for global context. Time-color coding (TCC) transforms video frame sequences into a single RGB image encoding object trajectory, ensuring compatibility with standard image classifiers. A three-stage training strategy with successive unfreezing of EfficientNetB4 blocks enables stable transfer learning on small datasets. The approach is evaluated on two domains: laboratory rat behaviors (1 710 training images, 13 classes) and jet aircraft maneuvers (2 819 images, 8 classes), using a frame-level held-out split with 1 000-iteration bootstrapping. The full-capacity model (49.7M parameters) achieved 94.62% median accuracy on the biological domain. On the aeronautical domain, reducing model capacity by 62.5% improved accuracy from 18.8% to 70.0% on the full 8-class task by preventing memorization of stabilized-camera recording artifacts (TCC decoherence); removing one structurally anomalous class raised accuracy to 86.43% on the remaining 7 classes, with a 34.9% reduction in classification errors. All pairwise differences between the proposed architecture and ablated configurations were statistically significant, except for the RGB-only ablation in Domain 2 (p = 0.538-0.604 across accuracy and F1-score; bootstrap empirical test, B $=1~000$ ; p $\le 0.006$ for all remaining comparisons). These results confirm that the CNN-ViT architecture with TCC provides a practical framework for small-scale behavioral video analysis, offering guidelines for model capacity selection and dataset curation under resource constraints.

Development of a method for redundant frame removal in video sequences based on assessment of the black-and-white frames similarity

Video sequences redundant frame removal currently plays an important role in various video preprocessing applications related to video classification, change detection, and object recognition (e.g., defects, faults, human actions, etc.). This paper proposes a video sequence subsampling method based on estimating the similarity of two consecutive black-and white frames using a "sign-sign" correlation function and associativity coefficients, verifying the compliance of the calculated metric values with threshold intervals, and making decisions regarding frame removal. A generalized scheme of the proposed method is considered with a detailed description of the stages, and an algorithm for finding similarities and differences between two consecutive black-and-white frames of video sequences is described. This method reduces the size of video sequences by removing redundant or repeated frames, thereby reducing the amount of stored information and increasing the efficiency of detecting key moments in video images obtained in uncontrolled environments. Research results are presented, demonstrating that the proposed method can halve frame redundancy without losing important information or improving the quality of video sequences.

Frame-Level Evaluation in Weakly Supervised Video Anomaly Detection Mostly Measures Video-Level Ranking

Weakly supervised video anomaly detectors are trained with video-level labels but are commonly evaluated as temporal localizers using Micro-AUROC or AP over pooled test frames. Because these metrics compare frames from different videos, a detector can score well by separating videos without accurately ordering moments within them. We exactly decompose Micro-AUROC by video identity into Within-AUROC for temporal ordering within videos and Cross-AUROC for comparisons across videos. Across ShanghaiTech, XD-Violence, and UCF-Crime, only 0.071-0.388% of comparisons between anomalous and normal frames occur within the same video. When both classes remain distributed across V videos, this share decreases as O(1/V), a benchmark property we call temporal dilution. We train anomaly video binary classifiers under the same video-level supervision and repeat each video score across all frames. These video-constant outputs reach 81.40-97.18 Micro-AUROC despite having no within-video variation. Across 72 controlled runs, replacing every frame score with its video mean preserves a median 98.6% of the Micro-AUROC margin above chance. The same empirical pattern holds for author-released outputs and for XD-Violence under its official AP evaluation. A detector can therefore achieve a high pooled score even when it assigns the same score to every moment within each video.