论文检索

Zixin Wang, Yadan Luo, Zhi Chen 0010, Sen Wang 0001, Zi Huang

The prevalence of domain adaptive semantic segmentation has prompted concerns regarding source domain data leakage, where private information from the source domain could inadvertently be exposed in the target domain. To circumvent the requirement for source data, source-free domain adaptation has emerged as a viable solution that leverages self-training methods to pseudo-label high-confidence regions and adapt the model to the target data. However, the confidence scores obtained are often highly biased due to overconfidence and class-imbalance issues, which render both model selection and optimization problematic. In this paper, we propose a novel calibration-guided source-free domain adaptive semantic segmentation (Cal-SFDA) framework. The core idea is to estimate the expected calibration error (ECE) from the segmentation predictions, serving as a strong indicator of the model's generalization capability to the unlabeled target domain. The estimated ECE scores, in turn, assist the model training and fair selection in both source training and target adaptation stages. During model pre-training on the source domain, we ensure the differentiability of the ECE objective by leveraging the LogSumExp trick and using ECE scores to select the best source checkpoints for adaptation. To enable ECE estimation on the target domain without requiring labels, we train a value net for ECE estimation and apply statistic warm-up on its BatchNorm layers for stability. The estimated ECE scores assist in determining the reliability of prediction and enable class-balanced pseudo-labeling by positively guiding the adaptation progress and inhibiting potential error accumulation. Extensive experiments on two widely-used synthetic-to-real transfer tasks show that the proposed approach surpasses previous state-of-the-art by up to 5.25% of mIoU with fair model selection criteria.

Guangming Zhu 0001, Siyuan Wang 0004, Qing Cheng 0004, Kelong Wu, Hao Li 0179, Liang Zhang 0010

With the recent surge in the use of touchscreen devices, free-hand sketching has emerged as a promising modality for human-computer interaction. While previous research has focused on tasks such as recognition, retrieval, and generation of familiar everyday objects, this study aims to create a Sketch Input Method Editor (SketchIME) specifically designed for a professional Command, Control, Communications, Computer, and Intelligence (C4I) system. Within this system, sketches are utilized as low-fidelity prototypes for recommending standardized symbols in the creation of comprehensive situation maps. This paper also presents a systematic dataset comprising 374 specialized sketch types, and proposes a simultaneous recognition and segmentation architecture with multilevel supervision between recognition and segmentation to improve performance and enhance interpretability. By incorporating few-shot domain adaptation and class-incremental learning, the network's ability to adapt to new users and extend to new task-specific classes is significantly enhanced. Results from experiments conducted on both the proposed dataset and the SPG dataset illustrate the superior performance of the proposed architecture. Our dataset and code are publicly available at https://github.com/GuangmingZhu/SketchIME.

Fan Jiang 0009, Zilei Wang

Efficient and accurate perception system is critical for autonomous driving, including traffic object detection, drivable area segmentation, and lane detection. Most previous works do not consider the spatial and semantic cues in traffic scenes. In this paper, we propose a novel multi-task learning network to exploit these priors. Specifically, to model the co-occurrence and spatial relationships of traffic objects, we propose to use a Graph Convolutional Network (GCN) block operating on the patches of feature maps. It enables adaptive discovery and incorporation of semantic and spatial relationships in the feature space. Furthermore, we propose a sub-feature sharing method to mitigate negative transfer in multi-task learning. On the basis of a fully shared base network, we split the feature space of different tasks along the channel dimension, resulting in the shared and private features for each task. It allows the network parameters to be selectively updated by different tasks during training. Experimental results on the challenging BDD100K dataset demonstrate that our proposed approach gets consistent improvement with fewer parameters, and achieves new state-of-the-art performance in terms of accuracy and speed.

Sun'ao Liu, Yiheng Zhang, Zhaofan Qiu, Hongtao Xie 0001, Yongdong Zhang 0001, Ting Yao 0003

Referring image segmentation aims to segment the target object described by a natural-language utterance. Recent approaches typically distinguish pixels by aligning pixel-wise visual features with linguistic features extracted from the referring description. Nevertheless, such a free-form description only specifies certain discriminative attributes of the target object or its relations to a limited number of objects, which fails to represent the rich visual context adequately. The stand-alone linguistic features are therefore unable to align with all visual concepts, resulting in inaccurate segmentation. In this paper, we propose to address this issue by incorporating rich visual context into linguistic features for sufficient vision-language alignment. Specifically, we present Context-Aware Referring Image Segmentation (CARIS), a novel architecture that enhances the contextual awareness of linguistic features via sequential vision-language attention and learnable prompts. Technically, CARIS develops a context-aware mask decoder with sequential bidirectional cross-modal attention to integrate the linguistic features with visual context, which are then aligned with pixel-wise visual features. Furthermore, two groups of learnable prompts are employed to delve into additional contextual information from the input image and facilitate the alignment with non-target pixels, respectively. Extensive experiments demonstrate that CARIS achieves new state-of-the-art performances on three public benchmarks. Code is available at https://github.com/lsa1997/CARIS.

Yao Wu, Mingwei Xing, Yachao Zhang 0001, Yuan Xie 0006, Jianping Fan 0007, Zhongchao Shi, Yanyun Qu

Cross-modal Unsupervised Domain Adaptation (UDA) becomes a research hotspot because it reduces the laborious annotation of target domain samples. Existing methods only mutually mimic the outputs of cross-modality in each domain, which enforces the class probability distribution agreeable in different domains. However, these methods ignore the complementarity brought by the modality fusion representation in cross-modal learning. In this paper, we propose a cross-modal UDA method for 3D semantic segmentation via Bidirectional Fusion-then-Distillation, named BFtD-xMUDA, which explores cross-modal fusion in UDA and realizes distribution consistency between outputs of two domains not only for 2D image and 3D point cloud but also for 2D/3D and fusion. Our method contains three significant components: Model-agnostic Feature Fusion Module (MFFM), Bidirectional Distillation (B-Distill), and Cross-modal Debiased Pseudo-Labeling (xDPL). MFFM is employed to generate cross-modal fusion features for establishing a latent space, which enforces maximum correlation and complementarity between two heterogeneous modalities. B-Distill is introduced to exploit bidirectional knowledge distillation which includes cross-modality and cross-domain fusion distillation, and well-achieving domain-modality alignment. xDPL is designed to model the uncertainty of pseudo-labels by self-training scheme. Extensive experimental results demonstrate that our method outperforms state-of-the-art competitors in several adaptation scenarios.

Yixuan Wu, Jintai Chen, Jiahuan Yan, Yiheng Zhu 0002, Danny Z. Chen, Jian Wu 0001

Since annotating medical images for segmentation tasks commonly incurs expensive costs, it is highly desirable to design an annotation-efficient method to alleviate the annotation burden. Recently, contrastive learning has exhibited a great potential in learning robust representations to boost downstream tasks with limited labels. In medical imaging scenarios, ready-made meta labels (i.e., specific attribute information of medical images) inherently reveal semantic relationships among images, which have been used to define positive pairs in previous work. However, the multi-perspective semantics revealed by various meta labels are usually incompatible and can incur intractable "semantic contradiction" when combining different meta labels. In this paper, we tackle the issue of "semantic contradiction" in a gradient-guided manner using our proposed Gradient Mitigator method, which systematically unifies multi-perspective meta labels to enable a pre-trained model to attain a better high-level semantic recognition ability. Moreover, we emphasize that the fine-grained discrimination ability is vital for segmentation-oriented pre-training, and develop a novel method called Gradient Filter to dynamically screen pixel pairs with the most discriminating power based on the magnitude of gradients. Comprehensive experiments on four medical image segmentation datasets verify that our new method GCL: (1) learns informative image representations and considerably boosts segmentation performance with limited labels, and (2) shows promising generalizability on out-of-distribution datasets.

Yushen Wei, Yang Liu 0084, Hong Yan 0004, Guanbin Li, Liang Lin 0004

Existing methods for video question answering (VideoQA) often suffer from spurious correlations between different modalities, leading to a failure in identifying the dominant visual evidence and the intended question. Moreover, these methods function as black boxes, making it difficult to interpret the visual scene during the QA process. In this paper, to discover critical video segments and frames that serve as the visual causal scene for generating reliable answers, we present a causal analysis of VideoQA and propose a framework for cross-modal causal relational reasoning, named Visual Causal Scene Refinement (VCSR). Particularly, a set of causal front-door intervention operations is introduced to explicitly find the visual causal scenes at both segment and frame levels. Our VCSR involves two essential modules: i) the Question-Guided Refiner (QGR) module, which refines consecutive video frames guided by the question semantics to obtain more representative segment features for causal front-door intervention; ii) the Causal Scene Separator (CSS) module, which discovers a collection of visual causal and non-causal scenes based on the visual-linguistic causal relevance and estimates the causal effect of the scene-separating intervention in a contrastive learning manner. Extensive experiments on the NExT-QA, Causal-VidQA, and MSRVTT-QA datasets demonstrate the superiority of our VCSR in discovering visual causal scene and achieving robust video question answering.

Xingran Xie, Ting Jin, Boxiang Yun, Qingli Li, Yan Wang 0033

Hyperspectral images (HSIs) offer great potential for computational pathology. However, limited by the spectral redundancy and the lack of spectral prior in popular 2D networks, previous HSI based techniques do not perform well. To address these problems, we propose to segment HSIs from a deformable perspective, which processes different spectral bands independently and fuses spatiospectral features of interest via deformable attention mechanisms. In addition, we propose Deformable Self-Supervised Spectral Regression (DF-S3R), which introduces two self-supervised pre-text tasks based on the low rank prior of HSIs enabling the network learning with spectrum-related features. During pre-training, DF-S3R learns both spectral structures and spatial morphology, and the jointly pre-trained architectures help alleviate the transfer risk to downstream fine-tuning. Compared to previous works, experiments show that our deformable architecture and pre-training method perform much better than other competitive methods on pathological semantic segmentation tasks, and the visualizations indicate that our method can trace the critical spectral characteristics from subtle spectral disparities. Code will be released at https://github.com/Ayakax/DFS3R.

Shiqin Wang, Xin Xu 0007, Xianzheng Ma, Kui Jiang, Zheng Wang 0007

Unsupervised Domain Adaptive Nighttime Semantic Segmentation (UDA-NSS) aims to adapt a robust model from a labeled daytime domain to an unlabeled nighttime domain. However, current advanced segmentation methods ignore the illumination effect and class discrepancies of different semantic classes during domain adaptation, showing an uneven prediction phenomenon. It is the completely ignored and underexplored issues of ''hard-to-adapt'' classes that some classes have a large performance gap between existing UDA-NSS methods and supervised learning counterparts while others have a very low performance gap. To realize ''hard-to-adapt'' classes' more sufficient learning and facilitate the UDA-NSS task, we present an Online Informative Class Sampling (OICS) strategy to adaptively mine informative classes from the target nighttime domain according to the corresponding spectrogram mean and the class frequency via our Informative Mixture of Experts. Furthermore, an Informativeness-based cross-domain Mixed Sampling (InforMS) framework is designed to focus on informative classes from the target nighttime domain by vesting their higher sampling probabilities when cross-domain mixing sampling and achieves better performance in UDA-NSS tasks. Consequently, our method outperforms state-of-the-art UDA-NSS methods by large margins on three widely-used benchmarks (e.g., ACDC, Dark Zurich, and Nighttime Driving). Notably, our method achieves state-of-the-art performance with 65.1% mIoU on ACDC-night-test and 55.4% mIoU on ACDC-night-val.

Jinkai Zheng, Xinchen Liu, Shuai Wang 0003, Lihao Wang, Chenggang Yan 0001, Wu Liu 0005

Binary silhouettes and keypoint-based skeletons have dominated human gait recognition studies for decades since they are easy to extract from video frames. Despite their success in gait recognition for in-the-lab environments, they usually fail in real-world scenarios due to their low information entropy for gait representations. To achieve accurate gait recognition in the wild, this paper presents a novel gait representation, named Gait Parsing Sequence (GPS). GPSs are sequences of fine-grained human segmentation, i.e., human parsing, extracted from video frames, so they have much higher information entropy to encode the shapes and dynamics of fine-grained human parts during walking. Moreover, to effectively explore the capability of the GPS representation, we propose a novel human parsing-based gait recognition framework, named ParsingGait. ParsingGait contains a Convolutional Neural Network (CNN)-based backbone and two light-weighted heads. The first head extracts global semantic features from GPSs, while the other one learns mutual information of part-level features through Graph Convolutional Networks to model the detailed dynamics of human walking. Furthermore, due to the lack of suitable datasets, we build the first parsing-based dataset for gait recognition in the wild, named Gait3D-Parsing, by extending the large-scale and challenging Gait3D dataset. Based on Gait3D-Parsing, we comprehensively evaluate our method and existing gait recognition methods. Specifically, ParsingGait achieves a 17.5% Rank-1 increase compared with the state-of-the-art silhouette-based method. In addition, by replacing silhouettes with GPSs, current gait recognition methods achieve about 12.5% ~ 19.2% improvements in Rank-1 accuracy. The experimental results show a significant improvement in accuracy brought by the GPS representation and the superiority of ParsingGait.

Xinyu Hu, Xunjian Yin, Xiaojun Wan

Previous studies on machine translation evaluation mostly focused on the quality of individual sentences, while overlooking the important role of contextual information. Although WMT Metrics Shared Tasks have introduced context content into the human annotations of translation evaluation since 2019, the relevant metrics and methods still did not take advantage of the corresponding context. In this paper, we propose a context-aware machine translation evaluation metric called Cont-COMET, built upon the effective COMET framework. Our approach simultaneously considers the preceding and subsequent contexts of the sentence to be evaluated and trains our metric to be aligned with the setting during human annotation. We also introduce a content selection method to extract and utilize the most relevant information. The experiments and evaluation of Cont-COMET on the official test framework from WMT show improvements in both system-level and segment-level assessments.

Yiwei Qin, Weizhe Yuan, Graham Neubig, Pengfei Liu

Modern embedding-based metrics for evaluation of generated text generally fall into one of two paradigms: discriminative metrics that are trained to directly predict which outputs are of higher quality according to supervised human annotations, and generative metrics that are trained to evaluate text based on the probabilities of a generative model. Both have their advantages; discriminative metrics are able to directly optimize for the problem of distinguishing between good and bad outputs, while generative metrics can be trained using abundant raw text. In this paper, we present a framework that combines the best of both worlds, using both supervised and unsupervised signals from whatever data we have available. We operationalize this idea by training T5Score, a metric that uses these training signals with mT5 as backbone. We perform an extensive empirical comparison with other existing metrics on 5 datasets, 19 languages and 280 systems, demonstrating the utility of our method. Experimental results show that: T5Score achieves the best performance on all datasets against existing top-scoring metrics at the segment level.

Yanzhi Tian, Xiang Li, Zeming Liu, Yuhang Guo, Bin Wang

In-Image Machine Translation (IIMT) aims to convert images containing texts from one language to another. Traditional approaches for this task are cascade methods, which utilize optical character recognition (OCR) followed by neural machine translation (NMT) and text rendering. However, the cascade methods suffer from compounding errors of OCR and NMT, leading to a decrease in translation quality. In this paper, we propose an end-to-end model instead of the OCR, NMT and text rendering pipeline. Our neural architecture adopts encoder-decoder paradigm with segmented pixel sequences as inputs and outputs. Through end-to-end training, our model yields improvements across various dimensions, (i) it achieves higher translation quality by avoiding error propagation, (ii) it demonstrates robustness for out domain data, and (iii) it displays insensitivity to incomplete words. To validate the effectiveness of our method and support for future research, we construct our dataset containing 4M pairs of De-En images and train our end-to-end model. The experimental results show that our approach outperforms both cascade method and current end-to-end model.

Inderjeet Nair, Shwetha Somasundaram, Apoorv Saxena, Koustava Goswami

We address the task of evidence retrieval for long document question answering, which involves locating relevant paragraphs within a document to answer a question. We aim to assess the applicability of large language models (LLMs) in the task of zero-shot long document evidence retrieval, owing to their unprecedented performance across various NLP tasks. However, currently the LLMs can consume limited context lengths as input, thus providing document chunks as inputs might overlook the global context while missing out on capturing the inter-segment dependencies. Moreover, directly feeding the large input sets can incur significant computational costs, particularly when processing the entire document (and potentially incurring monetary expenses with enterprise APIs like OpenAI’s GPT variants). To address these challenges, we propose a suite of techniques that exploit the discourse structure commonly found in documents. By utilizing this structure, we create a condensed representation of the document, enabling a more comprehensive understanding and analysis of relationships between different parts. We retain 99.6% of the best zero-shot approach’s performance, while processing only 26% of the total tokens used by the best approach in the information seeking evidence retrieval setup. We also show how our approach can be combined with *self-ask* reasoning agent to achieve best zero-shot performance in complex multi-hop question answering, just \approx 4% short of zero-shot performance using gold evidence.

Chun Lin, Ying-Jia Lin, Chia-Jen Yeh, Yi-Ting Li, Ching-Wen Yang, Hung-Yu Kao

Recent Chinese word segmentation (CWS) models have shown competitive performance with pre-trained language models’ knowledge. However, these models tend to learn the segmentation knowledge through in-vocabulary words rather than understanding the meaning of the entire context. To address this issue, we introduce a context-aware approach that incorporates unsupervised sentence representation learning over different dropout masks into the multi-criteria training framework. We demonstrate that our approach reaches state-of-the-art (SoTA) performance on F1 scores for six of the nine CWS benchmark datasets and out-of-vocabulary (OOV) recalls for eight of nine. Further experiments discover that substantial improvements can be brought with various sentence representation objectives.

Amit Moryossef, Zifan Jiang, Mathias Müller, Sarah Ebling, Yoav Goldberg

Sign language segmentation is a crucial task in sign language processing systems. It enables downstream tasks such as sign recognition, transcription, and machine translation. In this work, we consider two kinds of segmentation: segmentation into individual signs and segmentation into phrases, larger units comprising several signs. We propose a novel approach to jointly model these two tasks. Our method is motivated by linguistic cues observed in sign language corpora. We replace the predominant IO tagging scheme with BIO tagging to account for continuous signing. Given that prosody plays a significant role in phrase boundaries, we explore the use of optical flow features. We also provide an extensive analysis of hand shapes and 3D hand normalization. We find that introducing BIO tagging is necessary to model sign boundaries. Explicitly encoding prosody by optical flow improves segmentation in shallow models, but its contribution is negligible in deeper models. Careful tuning of the decoding algorithm atop the models further improves the segmentation quality. We demonstrate that our final models generalize to out-of-domain video content in a different signed language, even under a zero-shot setting. We observe that including optical flow and 3D hand normalization enhances the robustness of the model in this context.

Robin Algayres, Pablo Diego-Simon, Benoît Sagot, Emmanuel Dupoux

Due to the absence of explicit word boundaries in the speech stream, the task of segmenting spoken sentences into word units without text supervision is particularly challenging. In this work, we leverage the most recent self-supervised speech models that have proved to quickly adapt to new tasks through fine-tuning, even in low resource conditions. Taking inspiration from semi-supervised learning, we fine-tune an XLS-R model to predict word boundaries themselves produced by top-tier speech segmentation systems: DPDP, VG-HuBERT and DP-Parse. Once XLS-R is fine-tuned, it is used to infer new word boundary labels that are used in turn for another fine-tuning step. Our method consistently improves the performance of each system and set a new state-of-the-art that is, on average 130% higher than the previous one as measured by the F1 score on correctly discovered word tokens on five corpora featuring different languages. Finally, our system can segment speech from languages unseen during fine-tuning in a zero-shot fashion.

Sathvik Nair, Philip Resnik

An important assumption that comes with using LLMs on psycholinguistic data has gone unverified. LLM-based predictions are based on subword tokenization, not decomposition of words into morphemes. Does that matter? We carefully test this by comparing surprisal estimates using orthographic, morphological, and BPE tokenization against reading time data. Our results replicate previous findings and provide evidence that *in the aggregate*, predictions using BPE tokenization do not suffer relative to morphological and orthographic segmentation. However, a finer-grained analysis points to potential issues with relying on BPE-based tokenization, as well as providing promising results involving morphologically-aware surprisal estimates and suggesting a new method for evaluating morphological prediction.

Tuhin Chakrabarty, Kanishk Singh, Arkadiy Saakyan, Smaranda Muresan

Natural language instructions are a powerful interface for editing the outputs of text-to-image diffusion models. However, several challenges need to be addressed: 1) underspecification (the need to model the implicit meaning of instructions) 2) grounding (the need to localize where the edit has to be performed), 3) faithfulness (the need to preserve the elements of the image not affected by the edit instruction). Current approaches focusing on image editing with natural language instructions rely on automatically generated paired data, which, as shown in our investigation, is noisy and sometimes nonsensical, exacerbating the above issues. Building on recent advances in segmentation, Chain-of-Thought prompting, and visual question answering, we significantly improve the quality of the paired data. In addition, we enhance the supervision signal by highlighting parts of the image that need to be changed by the instruction. The model fine-tuned on the improved data is capable of performing fine-grained object-centric edits better than state-of-the-art baselines, mitigating the problems outlined above, as shown by automatic and human evaluations. Moreover, our model is capable of generalizing to domains unseen during training, such as visual metaphors.

Jiho Kim, Yeonsu Kwon, Yohan Jo, Edward Choi

While large language models (LLMs) have made considerable advancements in understanding and generating unstructured text, their application in structured data remains underexplored. Particularly, using LLMs for complex reasoning tasks on knowledge graphs (KGs) remains largely untouched. To address this, we propose KG-GPT, a multi-purpose framework leveraging LLMs for tasks employing KGs. KG-GPT comprises three steps: Sentence Segmentation, Graph Retrieval, and Inference, each aimed at partitioning sentences, retrieving relevant graph components, and deriving logical conclusions, respectively. We evaluate KG-GPT using KG-based fact verification and KGQA benchmarks, with the model showing competitive and robust performance, even outperforming several fully-supervised models. Our work, therefore, marks a significant step in unifying structured and unstructured data processing within the realm of LLMs.