---
title: "TAL-CGBR: Class-Guided Boundary Regression for Temporal Action Localization"
id: "plos-one-22-tal-cgbr-a-class-guided-boundary-regression-approach-for-temporal-action"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-22-tal-cgbr-a-class-guided-boundary-regression-approach-for-temporal-action"
content_type: "clinical_feed_article"
specialty: "Radiology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0360160"
published_at: "2026-10-12T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# TAL-CGBR: Class-Guided Boundary Regression for Temporal Action Localization
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-22-tal-cgbr-a-class-guided-boundary-regression-approach-for-temporal-action
- **Specialty:** [Radiology](https://medichelpline.com/clinical-feed/radiology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0360160)
- **Published At:** 2026-10-12T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- The paper proposes **TAL-CGBR**, a single-stage Temporal Action Localization (TAL) framework that fuses classification semantics into the regression branch to refine temporal boundaries. - Investigators designed a Class-Guided Boundary Regressor (CGBR) that transfers intermediate semantic features from the classification head into the boundary regressor, aiming to reduce task conflict between classification and localization. - The architecture uses a pre-trained video encoder, an Action Feature Extraction Unit (AFEU), Transformer-based encoding with windowed self-attention, and stacked Temporal Context Module (TCM) layers that produce an inverted pyramid multi-scale feature hierarchy. - A window-based attention mechanism restricts self-attention to a fixed temporal window, lowering theoretical complexity from full-sequence O(L^2D) to windowed O(W^2LD); the paper includes the attention equations but does not specify the exact window size W used in experiments. - The TCM layers combine convolutional projection and Transformer encoding to capture both short-term variations and stage-wise action evolution, with downsampling between layers to create multi-scale features. - During detection, multi-scale features feed into parallel classification and regression branches; during training, the model fuses classification features into regression to guide boundary prediction. - Experimental evaluation reported average mAP of **68.5%** on **THUMOS14** and **36.7%** on **ActivityNet v1.3**, placing the method second among recent methods cited by the authors and indicating a balance of accuracy and efficiency. - Datasets used included THUMOS14, ActivityNet1.3, and EPIC-Kitchens-100; these were publicly available and referenced by the authors. - The paper emphasizes improved boundary precision and reduced computational cost via a lightweight Temporal Context Module versus heavier transformer decoders; precise parameter counts and some hyperparameters were not reported in the source text. - The study team performed ablation experiments that, according to the authors, show consistent improvements from cross-branch feature fusion, but the source did not enumerate detailed ablation metrics or all training hyperparameters in the provided text.
## Clinical Analysis & Structured Key Points
TAL-CGBR: A class-guided boundary regression approach for Temporal Action Localization | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Reader Comments Figures Figures Abstract Temporal Action Localization (TAL) aims to identify all action categories along with their precise temporal boundaries in untrimmed videos. It serves as a fundamental task supporting various downstream applications, including intelligent video editing, automatic summarization, anomaly detection, and human–computer interaction. Despite significant progress, TAL in real-world scenarios remains challenging due to two critical factors: substantial intra-class variations in action duration, ranging from a few seconds to several minutes, and severe degradation of boundary localization accuracy caused by variations in illumination, background clutter, and camera viewpoints. To address these issues, this paper presents a Temporal Action Localization with Class-Guided Boundary Regression (TAL-CGBR) framework, which incorporates semantic information from the classification branch into the regression network to enhance boundary precision. In addition, a window-based attention mechanism and a temporal hierarchical feature structure are introduced to improve computational efficiency while maintaining high localization accuracy. Experimental evaluations on benchmark datasets demonstrate that the proposed TAL-CGBR achieves competitive performance, attaining 68.5% average mAP on THUMOS14 and 36.7% average mAP on ActivityNet v1.3, ranking second among recent methods and demonstrating a favorable balance between accuracy and computational efficiency. Citation: Wei J, Chai H, Hou W, Yuan H, Huang X, Huang S (2026) TAL-CGBR: A class-guided boundary regression approach for Temporal Action Localization. PLoS One 21(10): e0360160. https://doi.org/10.1371/journal.pone.0360160 Editor: Xian-Shi Zhang, University of Electronic Science and Technology of China, CHINA Received: April 1, 2026; Accepted: September 22, 2026; Published: October 12, 2026 Copyright: © 2026 Wei et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: The datasets used in this study are publicly available from the following sources: • THUMOS14 dataset: https://www.crcv.ucf.edu/THUMOS14/download.html • ActivityNet1.3 dataset: http://activity-net.org/download.html • EPIC-Kitchens-100 dataset: https://epic-kitchens.github.io/2020-100#downloadFiles These datasets were accessed at their respective official download pages and were used under the terms and licenses provided by the dataset creators. Funding: This work is partially supported by the Second Batch of Undergraduate Provincial Teaching Reform Project of the 14th Five-Year Plan of Zhejiang Province, China (JGBA2024676), the Fundamental Research Funds of Hangzhou City University (No. X-202306), and the Teaching Reform Project of Zhejiang Province, China (No. JGCG2025196). Competing interests: The authors have declared that no competing interests exist. 1. Introduction In recent years, substantial progress has been achieved in the field of video understanding, with Temporal Action Localization (TAL) emerging as one of its core tasks [ 1 , 2 ]. TAL aims to determine when and what actions occur in untrimmed videos [ 2 ]. This technology has been widely applied to various domains, including intelligent video editing, pedestrian detection for autonomous driving, security surveillance, and human–computer interaction. These applications impose stringent requirements on both the precision and robustness of action boundary localization [ 3 ]. Early research primarily focused on manually trimmed short video clips, which typically contain simple scenes and only a single action category. As a result, existing methods are able to To address these issues, this paper presents a Temporal Action Localization with Class-Guided Boundary Regression (TAL-CGBR) framework, which jointly detects action boundaries and categories in untrimmed long videos while maintaining a balance between localization precision and computational efficiency [ 4 – 10 ]. To mitigate the limitations of existing attention-based methods—particularly ActionFormer, which adopts local window attention and multi-scale features but trains classification and regression branches independently—we propose a Class-Guided Boundary Regressor (CGBR). Unlike prior works that treat classification and regression as separate tasks, CGBR explicitly transfers intermediate semantic features from the classification branch to the regression network, providing category-level guidance to refine boundary predictions. This cross-branch semantic fusion is, to the best of our knowledge, the first attempt to leverage classification semantics for boundary regression in TAL. Additionally, while ActionFormer employs a computationally heavy transformer decoder, we introduce a lightweight Temporal Context Module (TCM) that achieves superior performance with fewer parameters. The main contributions of this paper are summarized as follows: (1) A Class-Guided Boundary Regressor (CGBR) is designed to incorporate classification semantics into the regression process, effectively enhancing boundary localization accuracy; (2) A window-based attention mechanism and a temporal hierarchical feature structure are introduced to improve detection efficiency and robustness across varying action durations; (3) The proposed TAL-CGBR achieves 68.5% average mAP on THUMOS14 and 36.7% average mAP on ActivityNet v1.3, achieving competitive performance against recent state‑of‑the‑art methods, with average mAP of 68.5% on THUMOS14 and 36.7% on ActivityNet v1.3 In summary, while our framework shares certain high-level design choices with ActionFormer [ 4 ]—such as local window attention and multi-scale features—the core novelty of our work lies in the CGBR, which introduces semantic guidance into boundary regression. This cross-branch interaction addresses a fundamental limitation of prior works and, as demonstrated in our ablation study, leads to consistent improvements in boundary precision. Moreover, our lightweight TCM design achieves better performance with lower computational cost, and our extensive evaluation across third-person and first-person datasets further validates the robustness and generalization of our approach. 2. Materials and methods Related works: Current temporal action localization (TAL) methods can generally be categorized into two main classes according to their dependence on pre-generated temporal proposals. 2.1. Two-stage methods Two-stage methods, represented by VSGN are inspired by the proposal–classification paradigm in image object detection, following the workflow of “proposal generation followed by refined classification and regression” [ 11 ]. Specifically, these approaches first generate a set of coarse temporal candidate segments—time intervals that may contain actions—and subsequently perform two tasks for each segment: identifying the action category and refining the start and end boundaries. Although two-stage methods typically achieve high detection accuracy, they suffer from several inherent drawbacks. The generation and refinement of a large number of proposals substantially increase computational complexity, and maintaining high-quality proposals requires carefully designed components, such as specialized generation algorithms, task-specific loss functions, and complex two-branch architectures [ 12 ]. Furthermore, the discrepancy in feature distributions between the proposal generation and localization stages may introduce cumulative errors, ultimately degrading boundary precision. 2.2. Single-stage methods The second class, exemplified by Actionformer, employs a single-stage design that directly predicts both the action category and its temporal boundaries within a single inference process, thereby removing the dependency on proposal generation [ 4 ]. Single-stage approaches generally offer improved computational efficiency compared with two-stage methods. Recently, Transformer-based frameworks have achieved progress in this category by leveraging the strong sequence-modeling capability of self-attention mechanisms to capture long-range temporal dependencies and complex action dynamics [ 4 – 6 ]. Consequently, single-stage Transformer-based models currently dominate TAL performance benchmarks. The method proposed in this paper belongs to this category and extends prior Transformer-based approaches with improved efficiency and boundary precision. 2.3 Transformer-based TAL methods Although Transformer networks have demonstrated strong representational capacity, their direct adaptation to TAL remains non-trivial. Wang et al. employed a standard Transformer module to model long-term dependencies for boundary prediction, enhancing boundary confidence estimation by learning correlations among temporal segments [ 13 ]. However, their approach still relied on manually predefined anchor boxes, limiting adaptability to varying action durations. Tan et al. replaced the standard Transformer encoder with a boundary attention module to better capture temporal context [ 14 ]; however, their architecture required three separately trained branch heads, leading to increased model complexity. Moreover, these methods focused primarily on boundary regression while neglecting the joint modeling of action categories and boundaries. Liu et al., inspired by DETR, introduced an end-to-end Transformer for TAL, but its self-attention mechanism was less effective in capturing short-term dependencies, resulting in suboptimal recognition of short-duration actions [ 5 , 15 ]. Zhang et al. enhanced temporal dependency modeling via a temporal feature pyramid, yet applying self-attention at every feature level substantially increased computational costs [ 4 ]. In addition, their unified network design for both category classification and boundary regression ignored the task-specific differences between the two objectives, which limited the precision of boundary localization. Notably, ActionFormer [ 4 ] shares high-level components with our work—namely local window attention and multi-scale features—but treats classification and regression as independent tasks without semantic interaction. In contrast, our CGBR introduces a cross-branch feature fusion mechanism that leverages classification semantics to actively guide boundary refinement, addressing the task discrepancy that prior works overlooked. Beyond architectural variations, several recent studies have explicitly explored the interaction between classification and localization tasks in TAL. Zhu et al. [ 16 ] proposed a disentangled representation learning framework that separates category-specific and localization-specific features, aiming to reduce task conflict. However, their approach relies on additional adversarial training and does not actively transfer semantic information between branches. Shao et al. [ 17 ] introduced action sensitivity learning, which adaptively weights samples based on their difficulty for classification and localization, achieving improved performance at strict IoU thresholds. Nevertheless, their method adjusts training dynamics rather than modifying the network architecture to enable explicit semantic guidance. In contrast, our CGBR directly fuses intermediate classification features into the regression branch, providing a more direct and principled mechanism for leveraging category-level context to refine boundary predictions. Another line of research exploits graph-based structures to model temporal relationships and enhance localization. Zeng et al. [ 18 , 19 ] introduced graph convolutional modules that capture temporal dependencies and semantic relationships among action proposals, demonstrating improved boundary refinement through message passing across proposals. While these graph-based methods effectively model inter-proposal relationships, they typically operate at the proposal level and require explicit graph construction, which can be computationally expensive. Our approach, by contrast, operates at the feature level within a single-stage framework, using a simple concatenation-based fusion that introduces no additional graph construction overhead while still providing effective semantic guidance. 3. Results The overall architecture of the proposed TAL-CGBR model is depicted in Fig 1 . Initially, the input video is processed by a pre-trained encoder to produce an initial feature sequence, which serves as the input to the subsequent modules. This sequence is then passed through an Action Feature Extraction Unit (AFEU). The convolutional projection layer aligns the feature dimensions with the Transformer’s input requirements, whereas the Transformer module captures long-range spatiotemporal dependencies within the video sequence through its self-attention mechanism, thereby generating high-level action representations enriched with spatiotemporal context. Download: PNG larger image TIFF original image Fig 1. Overview of the proposed model (TAL-CGBR). https://doi.org/10.1371/journal.pone.0360160.g001 To effectively represent actions with diverse temporal durations, the spatiotemporally enhanced features are further processed through multiple Temporal Context Module (TCM) layers. Each TCM layer captures both transient variation patterns and stage-wise evolution characteristics of actions by mining temporal contextual dependencies. At the output of each TCM layer, a down sampling operation is applied to progressively shorten the feature sequence, thereby forming an inverted pyramid-shaped multi-scale feature hierarchy. This hierarchical structure enables the model to adaptively capture action information across varying temporal granularities while maintaining computational efficiency. During the action detection stage, the multi-scale features from all levels of the inverted pyramid are simultaneously fed into two parallel branches: a classification branch and a regression branch. The classification branch predicts the action category for each temporal segment and outputs the corresponding probability distribution, whereas the regression branch estimates the start and end temporal boundaries of each action instance. To mitigate the limitations arising from independent optimization of classification and regression tasks and to improve boundary regression accuracy, a feature fusion mechanism is employed during training. This mechanism transfers intermediate semantic features from the classification branch to the regression branch, enabling the boundary regressor to leverage both its inherent regression supervision and the semantic guidance from the classification branch. By incorporating category-level information in this manner, the regression branch achieves enhanced precision in temporal boundary prediction. The proposed model offers several structural advantages. First, the integration of multiple TCM layers enables adaptive capture of action features with diverse durations. By combining convolution and Transformer, TAL-CGBR reinforces spatiotemporal correlations in features early in the process, improving the modeling of complex actions. Second, the inverted pyramid-shaped feature sequences, combined with the synergistic effect of TCMs, allow effective action information extraction across different temporal scales. Finally, the feature fusion mechanism introduced during training enables the boundary regression branch to fully utilize category prediction information, enhancing boundary prediction accuracy. Collectively, these designs systematically address core challenges in temporal action localization—such as difficulties in modeling long-range dependencies and poor adaptability to dynamic action durations—from three dimensions: feature representation, multi-scale adaptation, and task interaction. In summary, our single-stage pipeline integrates video feature encoding, window-based self-attention, a multi-scale temporal feature hierarchy, and two parallel branches for action classification and boundary regression, forming a coherent and interpretable framework for temporal action localization. 3.1. Windowed attention mechanism The window‑based attention mechanism restricts the self‑attention receptive field to a fixed temporal window of size W, reducing computational complexity from O(L^2D) to O(W^2LD) while preserving the ability to model local temporal dependencies critical for action boundary detection. For the feature sequence output by the video encoder, in order to better extract action information, this paper introduces an action feature extraction module from [ 4 ], which includes a projection layer composed of convolutional neural networks and an encoding layer composed of Transformer networks. After passing through these modules, the input feature sequence F yields the action feature sequence . The projection layer consists of a series of convolutional neural networks and activation layers, designed to adjust the dimensions of the input feature sequence so that it matches the dimensions required by subsequent Transformer networks. Additionally, following the research works of Li et al. and Xiao et al., incorporating convolutional layers before Transformer networks for processing temporal sequences can better adapt to windowed attention [ 20 , 21 ]. Furthermore, the inductive bias introduced by these convolutional layers can make subsequent Transformer network training more straightforward. The encoding layer is composed of multiple stacked Transformer networks. Referencing existing research works, our model improves upon the original Transformer’s self-attention mechanism by limiting its temporal scope [ 22 – 24 ]. For temporal action localization tasks, an action typically occurs across a sequence of nearby frames; therefore, elements far away in the feature sequence contribute little to action information extraction but significantly increase computational complexity. (1) The output of the self-attention layer is: (2) Here, , , denote the learnable transformation matrices for queries, keys, and values, respec- tively. represents the input feature sequence, which undergoes the above transformations to compute the attention queries , keys , and values . The Softmax function is applied row-wise, and denotes the dimension of the input features. Our model employs the Multi-Head Self Attention (MHSA) mechanism, which involves multiple parallel self-attention operations. While the original attention mechanism computes across the entire sequence, our model restricts attention computation to a fixed window for computational efficiency. The original mechanism has time and space complexities of , whereas computing attention within a fixed window size W reduces complexity to . Here, typically scales with the total video duration in practical applications, while is a fixed window size, generally much smaller than ( ). This modification significantly reduces computational costs. Following the original Transformer network architecture, after the multi-head attention layer, our implementation includes operations such as Layer Normalization (LN) and Multi-Layer Perceptron (MLP). Specifically, in our model, Layer Normalization is applied before both the multi-head attention and the feed-forward network. The detailed computation process is shown in Equation 3 : (3) Here, , , represent multi-head self-attention, layer normalization, and multi-layer perceptron operations, respectively. denotes the input feature sequence, is an intermediate variable obtained by adding the residual of the attention
## Related Clinical Research

- [TCLNet: Thermal contrast-aware lightweight network improves infrared object detection](https://medichelpline.com/clinical-feed/plos-one-6-tclnet-a-thermal-contrast-aware-lightweight-network-for-infrared-object.md)
- [X-ray diffraction of hydrated macromolecular crystals using a laser-driven X-ray source](https://medichelpline.com/clinical-feed/biorxiv-23-x-ray-diffraction-from-hydrated-macromolecular-crystals-using-laser-driven-x.md)
- [Measuring Directional Disagreement Between AI and Radiologists: PNA, NNA, NENA and Fragility](https://medichelpline.com/clinical-feed/medrxiv-5-disagreement-between-ai-and-radiologists-quantifying-direction-and-fragility.md)
- [CPMsi for scalable Mass Spectrometry Imaging: fast, cohort-scale MSI data processing](https://medichelpline.com/clinical-feed/biorxiv-3-cpmsi-a-data-driven-computational-pipeline-for-fast-large-scale-mass.md)
- [Scaled m‑BonacciNet: Recurrence‑Guided Lightweight CNN for Medical Image Classification](https://medichelpline.com/clinical-feed/medrxiv-6-scaled-m-bonaccinet-a-recurrence-guided-lightweight-cnn-for-efficient-medical.md)

## Navigation
- [← Back to Radiology Feed](https://medichelpline.com/clinical-feed/radiology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.