Xinyi Wang, Wei Dai, Kyle Qiao, Ke Wang, Peng Chen, Gang Cao, Kangqin, Zhongpu Wang, Xiaode Zhang, Yanming Liu, Jihao Gu, Jingtao Xu, Gong Zhi. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Role-playing agents (RPAs) have attracted growing interest for their ability to simulate immersive and interactive characters. However, existing approaches primarily focus on static role profiles, overlooking the dynamic perceptual abilities inherent to humans. To bridge this gap, we introduce the concept of dynamic role profiles by incorporating video modality into RPAs. To support this, we construct Role-playing-Video60k, a large-scale, high-quality dataset comprising 60k videos and 700k corresponding dialogues. Based on this dataset, we develop a comprehensive RPA framework that combines adaptive temporal sampling with both dynamic and static role profile representations. Specifically, the dynamic profile is created by adaptively sampling video frames and feeding them to the LLM in temporal order, while the static profile consists of (1) character dialogues from training videos during fine-tuning, and (2) a summary context from the input video during inference. This joint integration enables RPAs to generate greater responses. Furthermore, we propose a robust evaluation method covering eight metrics. Experimental results demonstrate the effectiveness of our framework, highlighting the importance of dynamic role profiles in developing RPAs.

Detection in large scenes is a challenging issue due to small objects and extreme scale variation. It is difficult for the deep-learning-based detector to extract features of small objects with only a few pixels. Most existing methods employ image pyramid and feature pyramid for multi-scale inference to alleviate this issue. However, they lack scale awareness to adapt to objects with different scales. In this paper, we propose a novel Adaptive Zoom (AdaZoom) network for scale-aware large scene object detection. There are three main contributions. First, an Adaptive Zoom network is proposed to actively focus and adaptively zoom the focused regions for high-performance object detection in large scenes. Second, to tackle the problem of missing annotations for focused regions, we train AdaZoom with the reward which measures the quality of generated regions, based on the paradigm of deep reinforcement learning. At last, we propose collaborative training to iteratively promote the joint performance of AdaZoom and the detector. To validate the effectiveness, we conduct extensive experiments on VisDrone2019, UAVDT and DOTA datasets. The experiments show AdaZoom brings consistent and significant improvement over different detection networks, achieving state-of-the-art performance on these datasets, especially outperforming the existing methods by AP of 4.64% on VisDrone2019.

Intersection over union (IoU) score, also named Jaccard Index, is one of the most fundamental evaluation methods in machine learning. The original IoU computation cannot provide non-zero gradients and thus cannot be directly optimized by nowadays deep learning methods. Several recent works generalized IoU for bounding box regression, but they are not straightforward to adapt for pixelwise prediction. In particular, the original IoU fails to provide effective gradients for the non-overlapping and location-deviation cases, which results in performance plateau. In this paper, we propose PixIoU, a generalized IoU for pixelwise prediction that is sensitive to the distance for non-overlapping cases and the locations in prediction. We provide proofs that PixIoU holds many nice properties as the original IoU. To optimize the PixIoU, we also propose a loss function that is proved to be submodular, hence we can apply the Lovász functions, the efficient surrogates for submodular functions for learning this loss. Experimental results show consistent performance improvements by learning PixIoU over the original IoU for several different pixelwise prediction tasks on Pascal VOC, VOT-2020 and Cityscapes.

Based on extended phase shift (EPS) modulation with voltage matching control and the DC blocking capacitor, an improved modulation method for dual-active-bridge (DAB) DC-DC converters with the blocking capacitor is proposed in this paper, which can realize full load range zero voltage switching (ZVS) under wide voltage range with low root-mean-square (RMS) current. In the proposed scheme, the volt-second products of all windings are kept equal to achieve the full load range ZVS, and the secondary side of the converter is switched between the full bridge mode and the half bridge mode to switch the optimal matching point of the voltage and then reduces RMS current and enlarges the ZVS region. The operation principle of the proposed method is analyzed, and ZVS region is illustrated and compared with conventional modulation to demonstrate the performance. A 1kW experimental prototype with 200V input and 100 - 400V output was built to verify the feasibility and effectiveness of the proposed method.

Face Parsing assigns every pixel in a facial image with a semantic label, which could be applied in various applications including face recognition, facial beautification, affective computing and animation. While lots of progress have been made in this field, current state-of-the-art methods still fail to extract real effective feature and restore accurate score map, especially for those facial parts which have large variations of deformation and fairly similar appearance, e.g. mouth, eyes and thin eyebrows. In this paper, we propose a novel pixel-wise face parsing method called Residual Encoder Decoder Network (RED-Net), which combines a feature-rich encoder-decoder framework with adaptive prior mechanism. Our encoder-decoder framework extracts feature with ResNet and decodes the feature by elaborately fusing the residual architectures in to deconvolution. This framework learns more effective feature comparing to that learnt by decoding with interpolation or classic deconvolution operations. To overcome the appearance ambiguity between facial parts, an adaptive prior mechanism is proposed in term of the decoder prediction confidence, allowing refining the final result. The experimental results on two public datasets demonstrate that our method outperforms the state-of-the-arts significantly, achieving improvements of F-measure from 0.854 to 0.905 on Helen dataset, and pixel accuracy from 95.12% to 97.59% on the LFW dataset. In particular, convincing qualitative examples show that our method parses eye, eyebrow, and lip regins more accurately.

Document image quality assessment (DIQA) aims to predict the visual quality of degraded document images. Although the definition of “visual quality” can change based on the specific applications, in this paper, we use OCR accuracy as a metric for quality and develop a novel no-reference DIQA method based on high order image statistics for OCR accuracy prediction. The proposed method consists of three steps. First, normalized local image patches are extracted with regular grid and a comprehensive document image codebook is constructed by K-means clustering. Second, local features are softly assigned to several nearest codewords, and the direct differences between high order statistics of local features and codewords are calculated as global quality aware features. Finally, support vector regression (SVR) is utilized to learn the mapping between extracted image features and OCR accuracies. Experimental results on two document image databases show that the proposed method can accurately predict OCR accuracy and outperforms previous algorithms.

Almost all present successful general purpose blind image quality assessment (BIQA) models are designed only towards singly and synthetically distorted images. However, real world images generally contain multiply types of distortions other than single type of distortion. Even some state of the art BIQA methods cannot work well on authentically distorted images. In this paper we propose a novel effective BIQA method for authentically distorted images to remedy the shortage of the popular BIQA models. First, except for traditional natural scene statistics (NSS) based features, we introduce a series of quality aware features based on human visual system (HVS), such as image dynamic range, color information, blurriness. Then support vector regression (SVR) is utilized to learn the mapping between the combined effective features and human opinion scores. Experimental results demonstrate the promising performance of the proposed method.

Blind image quality assessment (BIQA) research aims to develop a perceptual model to evaluate the quality of distorted images automatically and accurately without access to the non-distorted reference images. The state-of-the-art general purpose BIQA methods can be classified into two categories according to the types of features used. The first includes handcrafted features which rely on the statistical regularities of natural images. These, however, are not suitable for images containing text and artificial graphics. The second includes learning-based features which invariably require large codebook or supervised codebook updating procedures to obtain satisfactory performance. These are time-consuming and not applicable in practice. In this paper, we propose a novel general purpose BIQA method based on high order statistics aggregation (HOSA), requiring only a small codebook. HOSA consists of three steps. First, local normalized image patches are extracted as local features through a regular grid, and a codebook containing 100 codewords is constructed by K-means clustering. In addition to the mean of each cluster, the diagonal covariance and coskewness (i.e., dimension-wise variance and skewness) of clusters are also calculated. Second, each local feature is softly assigned to several nearest clusters and the differences of high order statistics (mean, variance and skewness) between local features and corresponding clusters are softly aggregated to build the global quality aware image representation. Finally, support vector regression is adopted to learn the mapping between perceptual features and subjective opinion scores. The proposed method has been extensively evaluated on ten image databases with both simulated and realistic image distortions, and shows highly competitive performance to the state-of-the-art BIQA methods.

Blind image quality assessment (BIQA) aims to develop quantitative measures to automatically and accurately estimate perceptual image quality without any prior information about the reference image. In this paper, we introduce a novel BIQA metric by structural and luminance information, based on the characteristics of human visual perception for distorted image. We extract the perceptual structural features of distorted image by the local binary pattern distribution. Besides, the distribution of normalized luminance magnitudes is extracted to represent the luminance changes in distorted image. After extracting the features for structures and luminance, support vector regression is adopted to model the complex nonlinear relationship from feature space to quality measure. The proposed BIQA model is called no-reference quality assessment using statistical structural and luminance features (NRSL). Extensive experiments conducted on four synthetically distorted image databases and three naturally distorted image databases have demonstrated that the proposed NRSL metric compares favorably with the relevant state-of-the-art BIQA models in terms of high correlation with human subjective ratings. The MATLAB source code and validation results of NRSL are publicly online at http://www.ntu.edu.sg/home/wslin/Publications.htm.

Previous feature learning based blind image quality assessment (BIQA) methods invariably require large codebook or codebook updating procedure to obtain satisfying performance. In this paper, we propose a novel general purpose BIQA method, local feature aggregation (LFA) model, which requires only a much smaller codebook without the need for codebook updating. The proposed model consists of three steps. Firstly, normalized local raw image patches are extracted as local features through a regular grid and a 100 codeword codebook is constructed by K-means clustering. Secondly, the soft weighted differences between local features and codewords are aggregated to the global quality aware representation. Finally, support vector regression (SVR) is utilized to learn the mapping between features and subjective opinion scores. The proposed method is evaluated on two large image databases and achieves comparable performance to state-of-the-art BIQA methods.

In this paper, we propose a novel “Opinion Free” (OF) No-Reference Video Quality Assessment (NR-VQA) algorithm based on frame-level unsupervised feature learning and hysteresis temporal pooling. The system consists of three components: feature extraction with max-min pooling, frame quality prediction and temporal pooling. Frame level features are first extracted by unsupervised feature learning and used to train a linear Support Vector Regressor (SVR) for predicting quality scores frame by frame. Frame-level quality scores are then combined by temporal pooling to obtain a single video quality score. We tested the proposed method on the LIVE video quality database and experimental results show that without training on human opinion scores the proposed method is comparable to state-of-the-art NR-VQA algorithms.

In this paper, we propose two novel Statistical Metric Fusion (SMF) methods for Image Quality Assessment (IQA) metric enhancement. First, local quality map is constructed from existing state-of-the-art IQA algorithm. After that several statistical indices are extracted from local quality map. Finally, the extracted statistical indices are fused by Supervised Statistical Metric Fusion (SMF-S) based on Support Vector Regression (SVR) and Unsupervised Statistical Metric Fusion (SMF-U) based on Reciprocal Rank Fusion (RRF) to obtain the final quality score, respectively. Experimental results on the largest public IQA database TID2013 have demonstrated that the two proposed SMF methods can generally enhance the quality prediction performance of the fused IQA metric in terms of high correlation with human opinion scores.