We present a new boundary sensitive framework for polyp segmentation, termed Polyper.Our method is motivated by a clinical approach that seasoned medical practitioners often leverage the inherent features of interior polyp regions to tackle blurred boundaries.Inspired by this, we propose to explicitly leverages boundary regions to bolster the model's boundary discrimination capability while minimizing computational resource wastage. Our approach first extracts low-confidence boundary regions and high-confidence prediction regions from an initial segmentation map through differentiable morphological operators.Then, we design the boundary sensitive attention that concentrates on augmenting the features near the boundary regions using the high-confidence prediction region's characteristics to generate good segmentation results.Our proposed method can be seamlessly integrated with classical encoder networks, like ResNet-50, MiT-B1, and Swin Transformer.To evaludate the effectiveness of Polyper, we conduct experiments on five publicly available challenging datasets, and receive state-of-the-art performance on all of them. Code is available at https://github.com/haoshao-nku/medical_seg.git.
论文检索
输入标题、作者或关键词，从 2,866 篇学术成果中精准定位
Diffusion models have achieved remarkable progress in low-light image enhancement. However, there remain two practical limitations: (1) existing methods mainly focus on the spatial domain for the diffusion process, while neglecting the essential features in the frequency domain; (2) conventional patch-based sampling strategy inevitably leads to severe checkerboard artifacts due to the uneven overlapping. To address these limitations in one go, we propose a Multi-Domain Multi-Scale (MDMS) diffusion model for low-light image enhancement. In particular, we introduce a spatial-frequency fusion module to seamlessly integrates spatial and frequency information. By leveraging the Multi-Domain Learning (MDL) paradigm, our proposed model is endowed with the capability to adaptively facilitate noise distribution learning, thereby enhancing the quality of the generated images. Meanwhile, we propose a Multi-Scale Sampling (MSS) strategy that follows a divide-ensemble manner by merging the restored patches under different resolutions. Such a multi-scale learning paradigm explicitly derives patch information from different granularities, thus leading to smoother boundaries. Furthermore, we empirically adopt the Bright Channel Prior (BCP) which indicates natural statistical regularity as an additional restoration guidance. Experimental results on LOL and LOLv2 datasets demonstrate that our method achieves state-of-the-art performance for the low-light image enhancement task. Codes are available at https://github.com/Oliiveralien/MDMS.
The dot product self-attention (DPSA) is a fundamental component of transformers. However, scaling them to long sequences, like documents or high-resolution images, becomes prohibitively expensive due to the quadratic time and memory complexities arising from the softmax operation. Kernel methods are employed to simplify computations by approximating softmax but often lead to performance drops compared to softmax attention. We propose SeTformer, a novel transformer where DPSA is purely replaced by Self-optimal Transport (SeT) for achieving better performance and computational efficiency. SeT is based on two essential softmax properties: maintaining a non-negative attention matrix and using a nonlinear reweighting mechanism to emphasize important tokens in input sequences. By introducing a kernel cost function for optimal transport, SeTformer effectively satisfies these properties. In particular, with small and base-sized models, SeTformer achieves impressive top-1 accuracies of 84.7% and 86.2% on ImageNet-1K. In object detection, SeTformer-base outperforms the FocalNet counterpart by +2.2 mAP, using 38% fewer parameters and 29% fewer FLOPs. In semantic segmentation, our base-size model surpasses NAT by +3.5 mIoU with 33% fewer parameters. SeTformer also achieves state-of-the-art results in language modeling on the GLUE benchmark. These findings highlight SeTformer applicability for vision and language tasks.
Training Generative Adversarial Networks (GAN) to generate high-quality images typically requires large datasets. Network pruning during training has recently emerged as a significant advancement for data-efficient GAN. However, simple and straightforward pruning can lead to the risk of losing key information, resulting in suboptimal results due to GAN’s competitive dynamics between generator (G) and discriminator (D). Addressing this, we present RG-GAN, a novel approach that marks the first incorporation of dynamic weight regeneration and pruning in GAN training to improve the quality of the generated samples, even with limited data. Specifically, RG-GAN initiates layer-wise dynamic pruning by removing less important weights to the quality of the generated images. While pruning enhances efficiency, excessive sparsity within layers can pose a risk of model collapse. To mitigate this issue, RG-GAN applies a dynamic regeneration method to reintroduce specific weights when they become important, ensuring a balance between sparsity and image quality. Though effective, the sparse network achieved through this process might eliminate some weights important to the combined G and D performance, a crucial aspect for achieving stable and effective GAN training. RG-GAN addresses this loss of weights by integrating learned sparse network weights back into the dense network at the previous stage during a follow-up regeneration step. Our results consistently demonstrate RG-GAN’s robust performance across a variety of scenarios, including different GAN architectures, datasets, and degrees of data scarcity, reinforcing its value as a generic training methodology. Results also show that data augmentation exhibits improved performance in conjunction with RG-GAN. Furthermore, RG-GAN can achieve fewer parameters without compromising, and even enhancing, the quality of the generated samples. Code can be found at this link: https://github.com/IntellicentAI-Lab/RG-GAN
Text-to-image diffusion models can synthesize high quality images, but they have various limitations. Here we highlight a common failure mode of these models, namely, generating uncommon concepts and structured concepts like hand palms. We show that their limitation is partly due to the long-tail nature of their training data: web-crawled data sets are strongly unbalanced, causing models to under-represent concepts from the tail of the distribution. We characterize the effect of unbalanced training data on text-to-image models and offer a remedy. We show that rare concepts can be correctly generated by carefully selecting suitable generation seeds in the noise space, using a small reference set of images, a technique that we call SeedSelect. SeedSelect does not require retraining or finetuning the diffusion model. We assess the faithfulness, quality and diversity of SeedSelect in creating rare objects and generating complex formations like hand images, and find it consistently achieves superior performance. We further show the advantage of SeedSelect in semantic data augmentation. Generating semantically appropriate images can successfully improve performance in few-shot recognition benchmarks, for classes from the head and from the tail of the training data of diffusion models.
Active Learning (AL) aims to enhance the performance of deep models by selecting the most informative samples for annotation from a pool of unlabeled data. Despite impressive performance in closed-set settings, most AL methods fail in real-world scenarios where the unlabeled data contains unknown categories. Recently, a few studies have attempted to tackle the AL problem for the open-set setting. However, these methods focus more on selecting known samples and do not efficiently utilize unknown samples obtained during AL rounds. In this work, we propose an Entropic Open-set AL (EOAL) framework which leverages both known and unknown distributions effectively to select informative samples during AL rounds. Specifically, our approach employs two different entropy scores. One measures the uncertainty of a sample with respect to the known-class distributions. The other measures the uncertainty of the sample with respect to the unknown-class distributions. By utilizing these two entropy scores we effectively separate the known and unknown samples from the unlabeled data resulting in better sampling. Through extensive experiments, we show that the proposed method outperforms existing state-of-the-art methods on CIFAR-10, CIFAR-100, and TinyImageNet datasets. Code is available at https://github.com/bardisafa/EOAL.
Point cloud completion is an indispensable task for recovering complete point clouds due to incompleteness caused by occlusion, limited sensor resolution, etc. The family of coarse-to-fine generation architectures has recently exhibited great success in point cloud completion and gradually became mainstream. In this work, we unveil one of the key ingredients behind these methods: meticulously devised feature extraction operations with explicit cross-resolution aggregation. We present Cross-Resolution Transformer that efficiently performs cross-resolution aggregation with local attention mechanisms. With the help of our recursive designs, the proposed operation can capture more scales of features than common aggregation operations, which is beneficial for capturing fine geometric characteristics. While prior methodologies have ventured into various manifestations of inter-level cross-resolution aggregation, the effectiveness of intra-level one and their combination has not been analyzed. With unified designs, Cross-Resolution Transformer can perform intra- or inter-level cross-resolution aggregation by switching inputs. We integrate two forms of Cross-Resolution Transformers into one up-sampling block for point generation, and following the coarse-to-fine manner, we construct CRA-PCN to incrementally predict complete shapes with stacked up-sampling blocks. Extensive experiments demonstrate that our method outperforms state-of-the-art methods by a large margin on several widely used benchmarks. Codes are available at https://github.com/EasyRy/CRA-PCN.
Multi-Step Denoising Scheduled Sampling: Towards Alleviating Exposure Bias for Diffusion Models
PDF ↗Denoising Diffusion Probabilistic Models (DDPMs) have achieved significant success in generation tasks. Nevertheless, the exposure bias issue, i.e., the natural discrepancy between the training (the output of each step is calculated individually by a given input) and inference (the output of each step is calculated based on the input iteratively obtained based on the model), harms the performance of DDPMs. To our knowledge, few works have tried to tackle this issue by modifying the training process for DDPMs, but they still perform unsatisfactorily due to 1) partially modeling the discrepancy and 2) ignoring the prediction error accumulation. To address the above issues, in this paper, we propose a multi-step denoising scheduled sampling (MDSS) strategy to alleviate the exposure bias for DDPMs. Analyzing the formulations of the training and inference of DDPMs, MDSS 1) comprehensively considers the discrepancy influence of prediction errors on the output of the model (the Gaussian noise) and the output of the step (the calculated input signal of the next step), and 2) efficiently models the prediction error accumulation by using multiple iterations of a mathematical formulation initialized from one-step prediction error obtained from the model. The experimental results, compared with previous works, demonstrate that our approach is more effective in mitigating exposure bias in DDPM, DDIM, and DPM-solver. In particular, MDSS achieves an FID score of 3.86 in 100 sample steps of DDIM on the CIFAR-10 dataset, whereas the second best obtains 4.78. The code will be available on GitHub.
Learning to forecast bimanual object manipulation sequences from unimanual observations has broad applications in assistive robots and augmented reality. This challenging task requires us to first infer motion from the missing arm and the object it would have been manipulating were the person bimanual, then forecast the human and object motion while maintaining hand-object contact during manipulation. Previous attempts model the hand-object interactions only implicitly, and thus tend to produce unrealistic motion where the objects float in air. We address this with a novel neural network that (i) identifies and forecasts the pose for only the objects undergoing motion through an object motion module and (ii) refines human pose predictions by encouraging hand-object contact during manipulation through an ensemble of human pose predictors. The components are also designed to be generic enough for use in both unimanual and bimanual contexts. Our approach outperforms the state-of-the-art pose forecasting methods on bimanual manipulation datasets.
Continuous sign language recognition (CSLR) aims to recognize gloss sequences from continuous sign videos. Recent works enhance the gloss representation consistency by mining correlations between visual and contextual modules within individual sentences. However, there still remain much richer correlations among glosses across different sentences. In this paper, we present a simple yet effective Cross-Sentence Gloss Consistency (CSGC), which enforces glosses belonging to a same category to be more consistent in representation than those belonging to different categories, across all training sentences. Specifically, in CSGC, a prototype is maintained for each gloss category and benefits the gloss discrimination in a contrastive way. Thanks to the well-distinguished gloss prototype, an auxiliary similarity classifier is devised to enhance the recognition clues, thus yielding more accurate results. Extensive experiments conducted on three CSLR datasets show that our proposed CSGC significantly boosts the performance of CSLR, surpassing existing state-of-the-art works by large margins (i.e., 1.6% on PHOENIX14, 2.4% on PHOENIX14-T, and 5.7% on CSL-Daily).
Inferring the 3D structure of a non-rigid dynamic scene from a single moving camera is an under-constrained problem. Inspired by the remarkable progress of neural radiance fields (NeRFs) in photo-realistic novel view synthesis of static scenes, it has also been extended to dynamic settings. Such methods heavily rely on implicit neural priors to regularize the problem. In this work, we take a step back and investigate how current implementations may entail deleterious effects including limited expressiveness, entanglement of light and density fields, and sub-optimal motion localization. Further, we devise a factorisation-based framework that represents the scene as a composition of bandlimited, high-dimensional signals. We demonstrate compelling results across complex dynamic scenes that involve changes in lighting, texture and long-range dynamics.
E2HQV: High-Quality Video Generation from Event Camera via Theory-Inspired Model-Aided Deep Learning
PDF ↗The bio-inspired event cameras or dynamic vision sensors are capable of asynchronously capturing per-pixel brightness changes (called event-streams) in high temporal resolution and high dynamic range. However, the non-structural spatial-temporal event-streams make it challenging for providing intuitive visualization with rich semantic information for human vision. It calls for events-to-video (E2V) solutions which take event-streams as input and generate high quality video frames for intuitive visualization. However, current solutions are predominantly data-driven without considering the prior knowledge of the underlying statistics relating event-streams and video frames. It highly relies on the non-linearity and generalization capability of the deep neural networks, thus, is struggling on reconstructing detailed textures when the scenes are complex. In this work, we propose E2HQV, a novel E2V paradigm designed to produce high-quality video frames from events. This approach leverages a model-aided deep learning framework, underpinned by a theory-inspired E2V model, which is meticulously derived from the fundamental imaging principles of event cameras. To deal with the issue of state-reset in the recurrent components of E2HQV, we also design a temporal shift embedding module to further improve the quality of the video frames. Comprehensive evaluations on the real world event camera datasets validate our approach, with E2HQV, notably outperforming state-of-the-art approaches, e.g., surpassing the second best by over 40% for some evaluation metrics.
S2CycleDiff: Spatial-Spectral-Bilateral Cycle-Diffusion Framework for Hyperspectral Image Super-resolution
PDF ↗Hyperspectral image super-resolution (HISR) is a technique that can break through the limitation of imaging mechanism to obtain the hyperspectral image (HSI) with high spatial resolution. Although some progress has been achieved by existing methods, most of them directly learn the spatial-spectral joint mapping between the observed images and the target high-resolution HSI (HrHSI), failing to fully reserve the spectral distribution of low-resolution HSI (LrHSI) and the spatial distribution of high-resolution multispectral imagery (HrMSI). To this end, we propose a spatial-spectral-bilateral cycle-diffusion framework (S2CycleDiff) for HISR, which can step-wise generate the HrHSI with high spatial-spectral fidelity by learning the conditional distribution of spatial and spectral super-resolution processes bilaterally. Specifically, a customized conditional cycle-diffusion framework is designed as the backbone to achieve the spatial-spectral-bilateral super-resolution by repeated refinement, wherein the spatial/spectral guided pyramid denoising (SGPD) module seperately takes HrMSI and LrHSI as the guiding factors to achieve the spatial details injection and spectral correction. The outputs of the conditional cycle-diffusion framework are fed into a complementary fusion block to integrate the spatial and spectral details to generate the desired HrHSI. Experiments have been conducted on three widely used datasets to demonstrate the superiority of the proposed method over state-of-the-art HISR methods. The code is available at https://github.com/Jiahuiqu/S2CycleDiff.
Existing unsupervised deep product quantization methods primarily aim for the increased similarity between different views of the identical image, whereas the delicate multi-level semantic similarities preserved between images are overlooked. Moreover, these methods predominantly focus on the Euclidean space for computational convenience, compromising their ability to map the multi-level semantic relationships between images effectively. To mitigate these shortcomings, we propose a novel unsupervised product quantization method dubbed Hierarchical Hyperbolic Product Quantization (HiHPQ), which learns quantized representations by incorporating hierarchical semantic similarity within hyperbolic geometry. Specifically, we propose a hyperbolic product quantizer, where the hyperbolic codebook attention mechanism and the quantized contrastive learning on the hyperbolic product manifold are introduced to expedite quantization. Furthermore, we propose a hierarchical semantics learning module, designed to enhance the distinction between similar and non-matching images for a query by utilizing the extracted hierarchical semantics as an additional training supervision. Experiments on benchmark image datasets show that our proposed method outperforms state-of-the-art baselines.
Image captioning aims at generating descriptive and meaningful textual descriptions of images, enabling a broad range of vision-language applications. Prior works have demonstrated that harnessing the power of Contrastive Image Language Pre-training (CLIP) offers a promising approach to achieving zero-shot captioning, eliminating the need for expensive caption annotations. However, the widely observed modality gap in the latent space of CLIP harms the performance of zero-shot captioning by breaking the alignment between paired image-text features. To address this issue, we conduct an analysis on the CLIP latent space which leads to two findings. Firstly, we observe that the CLIP's visual feature of image subregions can achieve closer proximity to the paired caption due to the inherent information loss in text descriptions. In addition, we show that the modality gap between a paired image-text can be empirically modeled as a zero-mean Gaussian distribution. Motivated by the findings, we propose a novel zero-shot image captioning framework with text-only training to reduce the modality gap. In particular, we introduce a subregion feature aggregation to leverage local region information, which produces a compact visual representation for matching text representation. Moreover, we incorporate a noise injection and CLIP reranking strategy to boost captioning performance. We also extend our framework to build a zero-shot VQA pipeline, demonstrating its generality. Through extensive experiments on common captioning and VQA datasets such as MSCOCO, Flickr30k and VQAV2, we show that our method achieves remarkable performance improvements. Code is available at https://github.com/Artanic30/MacCap.
High-Order Structure Based Middle-Feature Learning for Visible-Infrared Person Re-identification
PDF ↗Visible-infrared person re-identification (VI-ReID) aims to retrieve images of the same persons captured by visible (VIS) and infrared (IR) cameras. Existing VI-ReID methods ignore high-order structure information of features while being relatively difficult to learn a reasonable common feature space due to the large modality discrepancy between VIS and IR images. To address the above problems, we propose a novel high-order structure based middle-feature learning network (HOS-Net) for effective VI-ReID. Specifically, we first leverage a short- and long-range feature extraction (SLE) module to effectively exploit both short-range and long-range features. Then, we propose a high-order structure learning (HSL) module to successfully model the high-order relationship across different local features of each person image based on a whitened hypergraph network. This greatly alleviates model collapse and enhances feature representations. Finally, we develop a common feature space learning (CFL) module to learn a discriminative and reasonable common feature space based on middle features generated by aligning features from different modalities and ranges. In particular, a modality-range identity-center contrastive (MRIC) loss is proposed to reduce the distances between the VIS, IR, and middle features, smoothing the training process. Extensive experiments on the SYSU-MM01, RegDB, and LLCM datasets show that our HOS-Net achieves superior state-of-the-art performance. Our code is available at https://github.com/Jaulaucoeng/HOS-Net.
Empowering CAM-Based Methods with Capability to Generate Fine-Grained and High-Faithfulness Explanations
PDF ↗Recently, the explanation of neural network models has garnered considerable research attention. In computer vision, CAM (Class Activation Map)-based methods and LRP (Layer-wise Relevance Propagation) method are two common explanation methods. However, since most CAM-based methods can only generate global weights, they can only generate coarse-grained explanations at a deep layer. LRP and its variants, on the other hand, can generate fine-grained explanations. But the faithfulness of the explanations is too low. To address these challenges, in this paper, we propose FG-CAM (Fine-Grained CAM), which extends CAM-based methods to enable generating fine-grained and high-faithfulness explanations. FG-CAM uses the relationship between two adjacent layers of feature maps with resolution differences to gradually increase the explanation resolution, while finding the contributing pixels and filtering out the pixels that do not contribute. Our method not only solves the shortcoming of CAM-based methods without changing their characteristics, but also generates fine-grained explanations that have higher faithfulness than LRP and its variants. We also present FG-CAM with denoising, which is a variant of FG-CAM and is able to generate less noisy explanations with almost no change in explanation faithfulness. Experimental results show that the performance of FG-CAM is almost unaffected by the explanation resolution. FG-CAM outperforms existing CAM-based methods significantly in both shallow and intermediate layers, and outperforms LRP and its variants significantly in the input layer. Our code is available at https://github.com/dongmo-qcq/FG-CAM.
Regenerating urban layout is an essential process for urban regeneration. In this paper, we propose a new task called text-driven urban layout regeneration, which provides an intuitive input modal - text - for users to specify the regeneration, instead of designing complex rules. Given the target region to be regenerated, we propose a one-stage text-driven urban layout regeneration model, Text2City, to jointly and progressively regenerate the urban layout (i.e., road and building layouts) based on textual layout descriptions and surrounding context (i.e., urban layouts and functions of the surrounding regions). Text2City first extracts road and building attributes from the textual layout description to guide the regeneration. It includes a novel one-stage joint regenerator network based on the conditioned denoising diffusion probabilistic models (DDPMs) and prior knowledge exchange. To harmonize the regenerated layouts through joint optimization, we propose the interactive & enhanced guidance module for self-enhancement and prior knowledge exchange between road and building layouts during the regeneration. We also design a series of constraints from attribute-, geometry- and pixel-levels to ensure rational urban layout generation. To train our model, we build a large-scale dataset containing urban layouts and layout descriptions, covering 147K regions. Qualitative and quantitative evaluations show that our proposed method outperforms the baseline methods in regenerating desirable urban layouts that meet the textual descriptions.
High-Fidelity 3D Head Avatars Reconstruction through Spatially-Varying Expression Conditioned Neural Radiance Field
PDF ↗One crucial aspect of 3D head avatar reconstruction lies in the details of facial expressions. Although recent NeRF-based photo-realistic 3D head avatar methods achieve high-quality avatar rendering, they still encounter challenges retaining intricate facial expression details because they overlook the potential of specific expression variations at different spatial positions when conditioning the radiance field. Motivated by this observation, we introduce a novel Spatially-Varying Expression (SVE) conditioning. The SVE can be obtained by a simple MLP-based generation network, encompassing both spatial positional features and global expression information. Benefiting from rich and diverse information of the SVE at different positions, the proposed SVE-conditioned NeRF can deal with intricate facial expressions and achieve realistic rendering and geometry details of high-fidelity 3D head avatars. Additionally, to further elevate the geometric and rendering quality, we introduce a new coarse-to-fine training strategy, including a geometry initialization strategy at the coarse stage and an adaptive importance sampling strategy at the fine stage. Extensive experiments indicate that our method outperforms other state-of-the-art (SOTA) methods in rendering and geometry quality on mobile phone-collected and public datasets. Code and data can be found at https://github.com/minghanqin/AvatarSVE.
Image editing approaches with diffusion models have been rapidly developed, yet their applicability are subject to requirements such as specific editing types (e.g., foreground or background object editing, style transfer), multiple conditions (e.g., mask, sketch, caption), and time consuming fine-tuning of diffusion models. For alleviating these limitations and realizing efficient real image editing, we propose a novel editing technique that only requires an input image and target text for various editing types including non-rigid edits without fine-tuning diffusion model. Our method contains three novelties: (I) Target-text Inversion Schedule (TTIS) is designed to fine-tune the input target text embedding to achieve fast image reconstruction without image caption and acceleration of convergence. (II) Progressive Transition Scheme applies progressive linear interpolation between target text embedding and its fine-tuned version to generate transition embedding for maintaining non-rigid editing capability. (III) Balanced Attention Module (BAM) balances the tradeoff between textual description and image semantics. By the means of combining self-attention map from reconstruction process and cross-attention map from transition process, the guidance of target text embeddings in diffusion process is optimized. In order to demonstrate editing capability, effectiveness and efficiency of the proposed BARET, we have conducted extensive qualitative and quantitative experiments. Moreover, results derived from user study and ablation study further prove the superiority over other methods.