π― What it does: This paper proposes a Channel-wise Lightweight Reprogramming (CLR) method, which achieves task reshaping in continual learning by adding a minimal number of channel-level learnable convolutional kernels on a frozen shared CNN backbone, thereby avoiding catastrophic forgetting.
π― What it does: A clustering-based supervised learning framework is proposed, utilizing online clustering within each category to discover potential subcategory patterns, which are used as auxiliary constraints to enhance point cloud representation learning.
π― What it does: For nighttime semantic segmentation, an unsupervised cross-modal domain adaptation framework CMDA is proposed, utilizing information from images and event sensors to achieve the transfer of source domain daytime images to target domain nighttime images.
π― What it does: Using 2D pose sequences and image features from videos, we first estimate the 3D skeleton of the mid-frame, and then regress the 3D human mesh through an image-guided pose-mesh collaborative evolution network.
π― What it does: A coarse-to-fine amodal segmentation method called C2F-Seg is proposed, which first generates a rough complete mask in the vector quantization latent space using a transformer, and then refines it to obtain a fine mask using convolution.
Collaborative Tracking Learning for Frame-Rate-Insensitive Multi-Object Tracking
Yiheng Liu (ByteDance Inc), Yi Fu (ByteDance Inc)
CodeObject TrackingTransformerVideo
π― What it does: This paper proposes a Collaborative Tracking Learning (ColTrack) framework that enhances multi-object tracking performance in low frame rate videos by using multiple historical queries to jointly track the same target.
π― What it does: This paper proposes the CoDis method, which handles deep learning tasks with noisy labels in a robust manner by using dual network collaboration and selecting samples with significant differences in predicted probabilities during training.
π― What it does: A single-step synthetic feature compressor (3SFC) is proposed to achieve communication efficiency in federated learning with an extremely low compression ratio.
π― What it does: A bias-free scene graph generation method called CFA is designed, which increases the feature diversity of tail class relationship triples by utilizing intrinsic feature replacement and extrinsic feature mixing.
π― What it does: Proposes the Concept-Tuning approach, which fine-tunes pre-trained models at the level of concepts (patches) to address rare features and spurious correlations, significantly reducing negative transfer.
Confidence-aware Pseudo-label Learning for Weakly Supervised Visual Grounding
Yang Liu (Peking University), Yuxin Peng (Peking University)
CodeObject DetectionRetrievalTransformerPrompt EngineeringVision Language ModelImageText
π― What it does: A confidence-aware pseudo-label learning framework for weakly supervised visual localization (CPL) is proposed, which constructs reliable image-text correspondences through the automatic generation of diverse pseudo queries, unified modality query propagation, and cross-modal confidence verification.
π― What it does: The ConSlide framework is proposed to achieve Whole Slide Image (WSI) analysis in a continuous learning environment, addressing challenges such as large image sizes, hierarchical structure utilization, and catastrophic forgetting.
π― What it does: A dual depth prediction and checkerboard selection strategy is proposed to enhance the accuracy of 3D reconstruction by constructing hanging-shaped depth cells in multi-view stereo.
Content-Aware Local GAN for Photo-Realistic Super-Resolution
JoonKyu Park (Seoul National University), Kyoung Mu Lee (Seoul National University)
CodeGenerationSuper ResolutionMixture of ExpertsGenerative Adversarial NetworkImage
π― What it does: Proposes the Content-Aware Local GAN (CAL-GAN) framework, which uses multiple expert discriminators to classify local features of different contents, thereby enhancing the realism of single-image super-resolution.
π― What it does: A continuous zero-shot learning framework (ICGZSL) is proposed that completely relies on unseen category semantic information, achieving realistic modeling of unseen visual spaces through generative models and random walk loss.
Continuously Masked Transformer for Image Inpainting
Keunsoo Ko (Catholic University of Korea), Chang-Su Kim (Korea University)
CodeRestorationTransformerImage
π― What it does: This paper proposes a Continuous Mask-aware Transformer (CMT) for image inpainting, utilizing continuous masks and overlapping tokens for multi-layer masked self-attention and mask updates. After generating an initial inpainting result, a refinement network is employed to enhance details.
π― What it does: Achieve adaptive cross-condition semantic segmentation models through contrastive learning without accessing labeled data from the source domain;
π― What it does: A new Open World DeepFake Attribution (OW-DFA) benchmark is proposed, along with a Contrastive Pseudo Learning (CPL) framework designed to simultaneously identify known and unknown types of forgery in an open environment where labeled and unlabeled samples coexist.
π― What it does: A single-stage multi-person human mesh recovery framework (CoordFormer) is proposed, which can directly detect, track, and recover 3D human meshes of multiple people from videos simultaneously.
COPILOT: Human-Environment Collision Prediction and Localization from Egocentric Videos
Boxiao Pan (Stanford University), Leonidas J. Guibas (Stanford University)
CodeAutonomous DrivingTransformerVideo
π― What it does: We propose a system called COPILOT that predicts and locates human-environment collisions from multi-view first-person videos, providing collision probabilities, involved joints, and collision heatmaps.
π― What it does: This paper proposes and implements the CORE framework, which enhances the perception performance of multi-agent systems through collaborative reconstruction learning.
CroCo v2: Improved Cross-view Completion Pre-training for Stereo Matching and Optical Flow
Philippe Weinzaepfel (NAVER LABS Europe), Jerome Revaud (NAVER LABS Europe)
CodeDepth EstimationTransformerOptical FlowImage
π― What it does: This paper studies an improved cross-view completion (CroCo v2) pre-training framework aimed at enhancing the performance of dense geometric tasks such as stereo matching and optical flow.
π― What it does: This paper proposes an online one-stage Cross-Contrast Feature Perturbation framework (CCFP) that simulates domain transfer by applying learnable perturbations to features in the latent space, and introduces a semantic consistency constraint during training to enhance the model's generalization performance on unseen target domains.
π― What it does: This paper proposes a large-scale cross-domain e-commerce dataset ROPE, which covers three domains: product pages, short videos, and live broadcasts. It also designs a unified cross-domain product representation framework COPE, using multi-modal contrastive learning and classification loss to learn cross-domain shared features. The framework is then evaluated on cross-domain retrieval and one-shot few-sample classification tasks.
Cross-Modal Translation and Alignment for Survival Analysis
Fengtao Zhou (Hong Kong University of Science and Technology), Hao Chen (Hong Kong University of Science and Technology)
CodeClassificationData-Centric LearningTransformerMultimodalityBiomedical Data
π― What it does: A cross-modal translation and alignment framework (CMTA) is proposed, which jointly predicts the survival time of cancer patients using pathological images and genomic data.
π― What it does: A consistent training method (CTVIS) is proposed, utilizing long video training, a memory bank, momentum average embeddings, and noise injection to align the training process of online video instance segmentation with the inference process, thereby enhancing the distinguishability of instance embeddings and addressing challenges such as occlusion and re-identification.
π― What it does: A new knowledge distillation method is proposed - Cumulative Spatial Knowledge Distillation (CSKD), which directly uses the dense spatial predictions of CNN as supervision for ViT, avoiding the difficulties of aligning intermediate features;
Hemanth Saratchandran (Australian Institute of Machine Learning, University of Adelaide), Simon Lucey (Australian Institute of Machine Learning, University of Adelaide)
π― What it does: This study investigates how to use second-order optimization methods (L-BFGS) to accelerate the training of Coordinate Networks (Coordinate-MLP).
CVSformer: Cross-View Synthesis Transformer for Semantic Scene Completion
Haotian Dong (Tianjin University), Di Lin (Tianjin University)
CodeSegmentationTransformerImage
π― What it does: A cross-view synthesis Transformer (CVSformer) framework is proposed, which generates multi-view features from a single RGB-D image by rotating 3D convolutional kernels and utilizes a cross-view Transformer to fuse these features for semantic scene completion.
π― What it does: A framework D3G is proposed for video sentence localization using glance annotation, aiming to reduce annotation costs while maintaining performance close to fully supervised methods.
DALL-Eval: Probing the Reasoning Skills and Social Biases of Text-to-Image Generation Models
Jaemin Cho (University of North Carolina Chapel Hill), Mohit Bansal (University of North Carolina Chapel Hill)
CodeObject DetectionGenerationTransformerVision Language ModelImage
π― What it does: This paper systematically evaluates the visual reasoning capabilities and social biases of text-to-image generation models by designing a new assessment framework.
Dancing in the Dark: A Benchmark towards General Low-light Video Enhancement
Huiyuan Fu (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)
CodeRestorationVideoBenchmark
π― What it does: A high-quality dynamic low-light video dataset, DID, is proposed, and an adaptive illumination iterative enhancement network, LAN, is designed based on the Retinex theory for enhancing low-light videos.
π― What it does: A method for unpaired nighttime image data augmentation is proposed, which converts daytime images to nighttime images by training a lightweight GAN (HED N GAN) and jointly trains an edge detector to ensure that the edges of the generated nighttime images are consistent with the original images, used for metric learning in image retrieval.
π― What it does: This paper proposes a data-free knowledge distillation framework for fine-grained visual classification (FGVC) called DFKD-FGVC, which efficiently transfers fine-grained knowledge by combining three key technologies: spatial attention generator, mixed high-order attention distillation (MHAD), and semantic feature contrastive learning (SFCL).
π― What it does: Proposes the Dataset Quantization (DQ) framework, which compresses large-scale datasets into smaller subsets, achieving no significant performance loss across various network architectures while also considering storage and computational efficiency.
π― What it does: A dense visual prediction framework DDP based on conditional diffusion models is proposed, which unifies the handling of tasks such as semantic segmentation, BEV segmentation, and depth estimation.
DDS2M: Self-Supervised Denoising Diffusion Spatio-Spectral Model for Hyperspectral Image Restoration
Yuchun Miao (Wuhan University), Dacheng Tao (University of Sydney)
CodeRestorationDiffusion modelImage
π― What it does: A self-supervised diffusion-based spatiotemporal spectral model DDS2M is proposed for recovering clear images from degraded hyperspectral images.
π― What it does: A multi-contrast MRI super-resolution and reconstruction network based on a variational model has been constructed, which can decompose the reference image into common and unique components, and only pass the common information to the target image;
π― What it does: A decoupled iterative refinement framework is proposed to achieve pixel-level 3D reconstruction of interacting hands from a single RGB image.
Deep Directly-Trained Spiking Neural Networks for Object Detection
Qiaoyi Su (University of Chinese Academy of Sciences), Guoqi Li (University of Chinese Academy of Sciences)
CodeObject DetectionSpiking Neural NetworkImage
π― What it does: A deep pulse neural network framework EMS-YOLO based on direct training is proposed, utilizing the full pulse residual network EMS-ResNet to achieve low-energy target detection.
Deep Feature Deblurring Diffusion for Detecting Out-of-Distribution Objects
Aming Wu (Xidian University), Cheng Deng (Xidian University)
CodeObject DetectionDiffusion modelImage
π― What it does: A deep feature diffusion method based on Gaussian blur for forward diffusion and reverse deblurring (DFDD) is proposed for unsupervised discrete distribution object detection.
Deep Fusion Transformer Network with Weighted Vector-Wise Keypoints Voting for Robust 6D Object Pose Estimation
Jun Zhou (Hong Kong Polytechnic University), Jing Qin (Hong Kong Polytechnic University)
CodePose EstimationTransformerPoint Cloud
π― What it does: This paper proposes the Deep Fusion Transformer (DFTr) network, which fuses RGB and depth features through a cross-modal Transformer and designs a weighted vector-wise voting algorithm to achieve non-iterative localization of 3D keypoints, thereby enabling robust 6D object pose estimation.
π― What it does: A global guidance-based feature transformation (GIFT) and relational distillation image harmonization network called GiftNet is proposed, along with the contribution of a new ccHarmony dataset.
π― What it does: A learnable enhancement network SycoNet is proposed for automatically generating diverse synthetic images and dynamically augmenting training data, thereby improving the effect of image harmonization.
π― What it does: A cross-view contrastive learning (CVCL) framework is proposed, which first learns view-related features using deep autoencoders, and then aligns the soft clustering assignments of different views through cluster-level contrastive loss, thereby achieving multi-view clustering.
DeePoint: Visual Pointing Recognition and Direction Estimation
Shu Nakamura (Kyoto University), Ko Nishino (RIKEN)
CodeRecognitionPose EstimationTransformerVideo
π― What it does: A model named DeePoint is proposed, which can automatically recognize pointing behavior in fixed-view RGB videos and estimate its 3D direction, and a large-scale pointing dataset called DP Dataset is constructed.
π― What it does: Under low-view angle conditions, high-precision 3D reconstruction of human heads is achieved using neural implicit functions through geometric decomposition and two-stage coarse-to-fine training.
π― What it does: Utilizing audio input, dense 2D and 3D predictions of indoor environments are achieved through cross-modal distillation with a visual teacher model, including depth, semantic segmentation, and 3D reconstruction.
π― What it does: This paper proposes a density-invariant feature extraction method based on Group-wise Contrastive Learning (GCL) to address the density mismatch problem in long-range LiDAR point cloud registration.
DETA: Denoised Task Adaptation for Few-Shot Learning
Ji Zhang (University of Electronic Science and Technology of China), Jingkuan Song (Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China)
Shilong Liu (Tsinghua University), Lei Zhang (International Digital Economy Academy)
CodeObject DetectionSegmentationTransformerImage
π― What it does: To address the matching instability issue that arises between multiple decoder layers in the DETR series detectors, two improvement schemes are proposed: position-supervised loss and position-modulated cost. Additionally, dense memory fusion is introduced to accelerate convergence, ultimately resulting in the construction of the Stable-DINO model.
DETRs with Collaborative Hybrid Assignments Training
Zhuofan Zong (SenseTime Research), Yu Liu (SenseTime Research)
CodeObject DetectionTransformerImage
π― What it does: A new collaborative hybrid allocation training scheme called C o-DETR is proposed to improve the efficiency and effectiveness of DETR-based detectors.
π― What it does: The DetZero framework is proposed, utilizing complete long sequences of LiDAR point clouds for offline 3D object detection, and generating complete trajectories through multi-frame detection and offline tracking, followed by an attention-based attribute-level refinement module (geometry, position, confidence) to further enhance detection accuracy.
π― What it does: Designed and implemented 3D Deformable Attention (DFA3D) to elevate multi-view 2D image features into a unified 3D space, refining features layer by layer through a hierarchical Transformer;
DIFFGUARD: Semantic Mismatch-Guided Out-of-Distribution Detection Using Pre-Trained Diffusion Models
Ruiyuan Gao (Chinese University of Hong Kong), Qiang Xu (Chinese University of Hong Kong)
CodeAnomaly DetectionDiffusion modelImage
π― What it does: A semantic mismatch detection framework called DIFFGUARD is proposed, which utilizes a pre-trained diffusion model to identify OOD samples in image classifiers.
π― What it does: DiffIR proposes a two-stage diffusion model framework for efficient image restoration, first extracting a compact recovery prior representation through the CPEN network, then denoising only that representation during the diffusion process, and finally reconstructing the image using DIRformer;
π― What it does: This paper proposes a multi-hypothesis 3D human pose estimation method called DiffPose, which is based on a conditional diffusion model. It generates 3D poses that conform to the distribution by sampling multiple joint candidates from 2D heatmaps and embedding them into a Transformer.
π― What it does: Treating temporary action detection as a denoising diffusion process from noisy proposals to action boundaries, the DiffTAD model is proposed.
Diffusion-Based 3D Human Pose Estimation with Multi-Hypothesis Aggregation
Wenkang Shan (National Engineering Research Center of Visual Technology Peking University), Wen Gao (National Engineering Research Center of Visual Technology Peking University)
CodePose EstimationDiffusion modelImage
π― What it does: A 3D human pose estimation framework based on diffusion models (D3DP) is proposed, which generates multiple 3D pose hypotheses conditioned on 2D keypoints and uses a joint-level multi-hypothesis aggregation method based on projection error (JPMA) to obtain a final single high-quality 3D pose.
π― What it does: A text-video retrieval framework called DiffusionRet based on diffusion models is proposed, treating the retrieval task as gradually generating a joint distribution of text and video from noise.
π― What it does: A video-to-speech (V2S) synthesis system called DiffV2S has been developed, which completes speech reconstruction under no audio conditions through vision-guided speaker embedding.
π― What it does: A method called DIRE, based on the reconstruction error of diffusion models, is proposed for detecting images generated by diffusion models.
Discovering Spatio-Temporal Rationales for Video Question Answering
Yicong Li (National University of Singapore), Tat-Seng Chua (National University of Singapore)
CodeRecognitionRetrievalExplainability and InterpretabilityTransformerVideo
π― What it does: A differentiable spatiotemporal rationalization (STR) module and the TranSTR network are proposed to automatically select key frames and key objects in long videos to support question answering.
π― What it does: Proposes the DiST framework, freezing CLIP ViT as a spatial encoder, combined with a lightweight temporal encoder and fusion branch to achieve efficient video transfer learning.
π― What it does: By using cross-modal knowledge distillation, the features of the LiDAR-based detector are transferred to the multi-camera BEV detector, thereby enhancing its 3D detection performance.
π― What it does: A differentiable particle rendering framework is designed to generate realistic, three-dimensional consistent, and temporally coherent weather effects. This framework is used to conduct adversarial attacks on optical flow estimation while evaluating and enhancing the network's robustness under weather conditions.
Distribution-Aware Prompt Tuning for Vision-Language Models
Eulrang Cho (Korea University), Hyunwoo J Kim
CodeDomain AdaptationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: A distribution-aware prompt tuning method DAPT is proposed to enhance the performance of pre-trained vision-language models in few-shot learning and domain transfer tasks.
π― What it does: This paper proposes a distribution-consistent modal recovery framework called DiCMoR, which first maps observable modalities to a Gaussian distribution space, then transfers the distribution to the missing modality through a cross-modal flow model to generate corresponding features, and finally inputs the recovered modality along with the existing modalities into a multimodal Transformer for downstream classification.
Diverse Data Augmentation with Diffusions for Effective Test-time Prompt Tuning
Chun-Mei Feng (Institute of High Performance Computing Agency for Science Technology and Research), Wangmeng Zuo (Institute of High Performance Computing Agency for Science Technology and Research)
CodeData SynthesisDomain AdaptationOptimizationTransformerPrompt EngineeringVision Language ModelDiffusion modelImage
π― What it does: This paper proposes DiffTPT, a method for adaptive prompt tuning on a single image during testing, which generates diverse augmented images using a stable diffusion model and employs cosine similarity filtering to ensure semantic consistency.
π― What it does: An end-to-end 3D point cloud instance segmentation network PBNet is proposed, which utilizes point density binarization and local scene reconstruction to achieve finer instance segmentation.
Divide and Conquer: a Two-Step Method for High Quality Face De-identification with Model Explainability
Yunqian Wen (Shanghai Jiao Tong University), Li Song (Shanghai Jiao Tong University)
CodeRestorationGenerationPose EstimationSafty and PrivacyConvolutional Neural NetworkNeural Radiance FieldGenerative Adversarial NetworkImage
π― What it does: A two-step facial de-identification method called IDeudemon is proposed, which first perturbs the identity code using 3D NeRF and then restores details using GAN.
π― What it does: Proposed the DOLCE model, which combines conditional diffusion and forward measurement consistency to achieve limited-angle CT reconstruction;
π― What it does: This paper proposes the Domain Adaptation Few-Shot Open Set Learning (DA-FSOS) task and designs an end-to-end model DAFOS-NET, which can simultaneously recognize known categories and reject unknown samples under the condition of fully supervised source domain and extremely few supervised target domain.
π― What it does: A 3DLabelProp method based on the accumulation of LiDAR point cloud sequences and geometric label propagation is proposed, and the first domain generalization benchmark for 3D semantic segmentation is established.
Liang Chen (Tencent AI Lab), Lingqiao Liu (University of Adelaide)
CodeClassificationDomain AdaptationExplainability and InterpretabilityConvolutional Neural NetworkImageBenchmark
π― What it does: This paper introduces a 'rationale' matrix in domain generalization, enforcing consistent decision contributions among similar samples, thereby enhancing the model's robustness to unknown domains.
π― What it does: The DomainDrop framework is proposed, which suppresses channels in the source domain that are susceptible to domain shift through channel dropout guided by a domain discriminator during training, thereby enhancing the model's generalization ability to unseen target domains.
Doppelgangers: Learning to Disambiguate Images of Similar Structures
Ruojin Cai (Cornell University), Noah Snavely (Cornell University)
CodeClassificationRecognitionConvolutional Neural NetworkSimultaneous Localization and MappingImage
π― What it does: This paper proposes a visual ambiguity discrimination method based on a binary classification network, which automatically determines whether two similar images correspond to the same 3D surface and integrates it into the SfM process to avoid erroneous reconstructions.
π― What it does: This study investigates the optimization process of knowledge distillation and finds that introducing distillation loss leads to a trade-off between task loss and distillation loss. It proposes the Distillation-Oriented Trainer (DOT), which sets different momenta for distillation loss and task loss, allowing the distillation loss to dominate the optimization, thereby reducing both losses simultaneously and achieving flatter, better generalizing minima.
Downscaled Representation Matters: Improving Image Rescaling with Collaborative Downscaled Images
Bingna Xu (South China University of Technology), Jian Chen (South China University of Technology)
CodeRestorationSuper ResolutionOptimizationImage
π― What it does: Proposes a Hierarchical Collaborative Downsampling (HCD) method that directly optimizes gradients on low-resolution images to enhance image reconstruction quality.
π― What it does: A fast diffusion probability model (DPM-OT) is proposed, which combines semi-discrete optimal transport. It directly maps the Gaussian prior to the intermediate latent space in one step, and then generates samples using a small number of reverse diffusion steps.
π― What it does: A framework DQS3D is proposed for semi-supervised 3D detection using dense matching and online compensation for quantization errors.
DRAW: Defending Camera-shooted RAW Against Image Manipulation
Xiaoxiao Hu (Fudan University), Xinpeng Zhang (Fudan University)
CodeSafty and PrivacyConvolutional Neural NetworkImage
π― What it does: Embed an invisible protection signal in RAW images, so that the rendered RGB images carry this signal, allowing for accurate localization of tampered areas in subsequent modifications.
π― What it does: This paper proposes the DREAM (Dataset distillation by REpresentative Matching) method, which utilizes representative samples for gradient matching to achieve efficient dataset distillation.
DREAMWALKER: Mental Planning for Continuous Vision-Language Navigation
Hanqing Wang (Beijing Institute of Technology), Wenguan Wang (Zhejiang University)
CodeOptimizationExplainability and InterpretabilityRobotic IntelligenceRecurrent Neural NetworkGraph Neural NetworkReinforcement LearningVision Language ModelWorld ModelMultimodality
π― What it does: This paper studies a vision-language navigation (VLN-CE) agent called DREAMWALKER, which can perform 'mental planning' in an internally abstract discrete environment and then map the best plan to low-level actions in a real continuous environment.
π― What it does: This paper proposes a deep learning framework called DReg-NeRF, which requires no manual labeling or initialization, to align multiple NeRF blocks trained in different coordinate systems to a common global coordinate system.
π― What it does: Proposes the DriveAdapter framework, which directly uses a frozen RL teacher model for planning in end-to-end autonomous driving, with the student model only responsible for perception. It aligns features through a learnable Adapter and incorporates masked feature alignment and action guidance loss to alleviate distribution differences and teacher errors.
π― What it does: A Dual-Axis Aggregation Transformer (DAT) is proposed for single image super-resolution, capable of aggregating spatial and channel features simultaneously, achieving stronger representation ability.
Dual Learning with Dynamic Knowledge Distillation for Partially Relevant Video Retrieval
Jianfeng Dong (Zhejiang Gongshang University), Baolong Liu (Zhejiang Gongshang University)
CodeRetrievalKnowledge DistillationConvolutional Neural NetworkTransformerVision Language ModelContrastive LearningVideoText
π― What it does: A dual-branch student network (inheritance branch and exploration branch) based on the large-scale visual language pre-training model CLIP is proposed for dynamic knowledge distillation to address the problem of Partial Relevant Video Retrieval (PRVR).
π― What it does: A gaze estimation network called DV-Gaze is proposed, which can directly predict the gaze direction from a pair of camera images under a dual-camera perspective.
DVIS: Decoupled Video Instance Segmentation Framework
Tao Zhang (Wuhan University), Pengfei Wan (Kuaishou Technology)
CodeObject TrackingSegmentationTransformerVideo
π― What it does: The DVIS framework is proposed, which breaks down the video instance segmentation task into three main sub-tasks: segmentation, tracking, and refinement. A lightweight referring tracker and temporal refiner are designed for each sub-task.
π― What it does: A unified rendering and simulation pipeline has been constructed, enabling the coupling of NeRF scenes and polygon meshes within the same physical space, and achieving real-time rendering and dynamic simulation on the GPU.
π― What it does: This paper proposes a Dynamic Residual Classifier (DRC) to address the dynamically worsening sample imbalance problem in Class Incremental Learning (CIL) as tasks increase, and integrates it with three mainstream CIL processes (MDT, MEC, MAF);
Dynamic Snake Convolution Based on Topological Geometric Constraints for Tubular Structure Segmentation
Yaolei Qi (Southeast University), Guanyu Yang (Southeast University)
CodeSegmentationConvolutional Neural NetworkBiomedical Data
π― What it does: This paper proposes a segmentation network specifically designed for tubular structures (such as blood vessels and roads) called DSCNet. It combines dynamic snake-like convolution, multi-view feature fusion, and topology continuity constraint loss based on persistent homology, significantly improving the segmentation accuracy and connectivity of tubular structures.
E3Sym: Leveraging E(3) Invariance for Unsupervised 3D Planar Reflective Symmetry Detection
Ren-Wu Li (Institute of Computing Technology, Chinese Academy of Sciences), Lin Gao (Institute of Computing Technology, Chinese Academy of Sciences)
CodeRecognitionSegmentationPoint Cloud
π― What it does: This paper proposes an unsupervised, end-to-end 3D global planar symmetry detection method called E3Sym, which constructs correspondences using E(3) invariant features and then obtains symmetric planes through differentiable clustering.