🎯 What it does: A high-resolution semantic correspondence framework LPMFlow is designed based on layout-aware representation learning, progressive feature super-resolution, and multi-scale matching flow fusion for pixel-level semantic correspondence.
🎯 What it does: This paper proposes a point cloud pre-training framework called PointDif based on diffusion models, which learns the hierarchical geometric priors of point clouds by progressively denoising noisy point clouds at different noise levels.
🎯 What it does: A general framework called PseCo based on point-segmentation-counting is proposed for few-shot and zero-shot object counting/detection.
Point2RBox: Combine Knowledge from Synthetic Visual Patterns for End-to-end Oriented Object Detection with Single Point Supervision
Yi Yu (Harbin Institute of Technology), Junchi Yan (Shanghai Jiao Tong University)
CodeObject DetectionImage
🎯 What it does: Proposes an end-to-end single-point supervised slanted object detection method Point2RBox, utilizing synthetic visual patterns and transformation self-supervision to achieve RBox detection.
🎯 What it does: A sparse bird's-eye view (BEV) segmentation model called PointBeV is proposed, which utilizes sparse point sets instead of full grid maps for BEV feature extraction and prediction.
PointOBB: Learning Oriented Object Detection via Single Point Supervision
Junwei Luo (Wuhan University), Yansheng Li (Wuhan University)
CodeObject DetectionImage
🎯 What it does: A direction target detection framework under single-point supervision, PointOBB, is proposed, which transforms single point annotations into oriented bounding boxes.
🎯 What it does: This paper proposes a polar coordinate-based visibility data reconstruction method called PolarRec, which can interpolate sparse radio interferometric visibility data to achieve full frequency coverage and generate high-quality astronomical images.
🎯 What it does: This paper proposes a lightweight feature extraction backbone network named Poly Kernel Inception Network (PKINet) to enhance the performance of object detection in remote sensing images.
POPDG: Popular 3D Dance Generation with PopDanceSet
Zhenye Luo (Beijing Normal University), Li Yao (Beijing Normal University)
CodeGenerationTransformerDiffusion modelVideo
🎯 What it does: This paper proposes a new popular dance dataset called PopDanceSet and constructs the POPDG model based on iDDPM, utilizing a spatial enhancement algorithm and alignment module to generate 3D dances that are highly synchronized and diverse with music.
PoseIRM: Enhance 3D Human Pose Estimation on Unseen Camera Settings via Invariant Risk Minimization
Yanlu Cai (Fudan University), Cheng Jin (Fudan University)
CodePose EstimationTransformerImage
🎯 What it does: This paper proposes a multi-view 3D human pose estimation framework called PoseIRM based on Invariant Risk Minimization (IRM). By synthesizing 2D pose data from various camera setups and using IRM constraints during training, the model can maintain good performance even under unseen camera configurations.
🎯 What it does: This paper proposes a new positive/unlabeled learning framework called LaGAM, which combines hierarchical contrastive learning and meta-learning to achieve label disambiguation and high-quality representation learning.
PracticalDG: Perturbation Distillation on Vision-Language Models for Hybrid Domain Generalization
Zining Chen (Beijing University of Posts and Telecommunications), Hongying Meng (Brunel University)
CodeDomain AdaptationKnowledge DistillationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposed SCI-PD, which transfers knowledge from VLM to lightweight visual models through perturbation distillation from three perspectives: scoring, categories, and instances, achieving open set domain generalization.
Pre-trained Model Guided Fine-Tuning for Zero-Shot Adversarial Robustness
Sibo Wang (Institute of Computing Technology Chinese Academy of Sciences), Shiguang Shan (Institute of Computing Technology Chinese Academy of Sciences)
🎯 What it does: Conduct adversarial fine-tuning on the CLIP pre-trained model and add an auxiliary branch based on the original model features to enhance adversarial robustness under zero-shot conditions.
🎯 What it does: A few-shot class incremental learning framework called PriViLege is proposed, based on a pre-trained visual and language Transformer, which significantly improves the model's incremental learning performance through pre-trained knowledge tuning, entropy-based divergence loss, and semantic knowledge distillation.
🎯 What it does: A formula-driven unsupervised pre-training dataset based on 3D Mandelbulb fractal variations is proposed, and pre-training is conducted on CNN and ViT models to validate their effectiveness in classification and anomaly detection tasks.
🎯 What it does: The Predicated Diffusion framework is proposed, which converts text prompts into predicate logic propositions and uses attention maps as fuzzy predicates to provide differentiable guidance to the diffusion model, thereby generating more faithful images.
PREGO: Online Mistake Detection in PRocedural EGOcentric Videos
Alessandro Flaborea (Sapienza University of Rome), Fabio Galasso (Sapienza University of Rome)
CodeClassificationRecognitionAnomaly DetectionTransformerLarge Language ModelVideo
🎯 What it does: This paper proposes an online one-shot classification model called PREGO, which can instantly detect unseen program errors in first-person videos.
🎯 What it does: A framework is proposed that combines feature disentanglement, dual-layer fairness loss, and loss surface smoothing to achieve fairness generalization in deep fake detection models.
Privacy-Preserving Face Recognition Using Trainable Feature Subtraction
Yuxi Mi (Fudan University), Shuigeng Zhou (Fudan University)
CodeRecognitionSafty and PrivacyConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: A privacy-preserving facial recognition framework named MinusFace is proposed, which obtains the residual by performing feature subtraction between the original facial image and its generated reconstructed image. This residual is then processed through high-dimensional mapping and random channel permutation to generate a final recognizable and privacy-preserving representation.
🎯 What it does: A 3D facial action dataset based on the large-scale VoxCeleb2 videos has been constructed, metrics for measuring the quality of probabilistic models have been designed, and a two-stage residual vector quantization (RVQ) autoregressive model has been proposed to achieve speech-driven diverse 3D facial animation generation.
Probing Synergistic High-Order Interaction in Infrared and Visible Image Fusion
Naishan Zheng (University of Science and Technology of China), Feng Zhao (University of Science and Technology of China)
CodeImage TranslationRestorationTransformerImage
🎯 What it does: A fusion paradigm SHIP based on high-order spatial fine-grained interaction and channel statistical interaction is proposed for the fusion of infrared and visible images, achieving efficient computation of high-order interactions within this framework.
Probing the 3D Awareness of Visual Foundation Models
Mohamed El Banani (University of Michigan), Varun Jampani (Google Research)
CodeDepth EstimationImage
🎯 What it does: This paper evaluates the three-dimensional perception capabilities of frozen features from visual foundation models by probing them for single-view depth, normal vector prediction, and multi-view consistency.
Producing and Leveraging Online Map Uncertainty in Trajectory Prediction
Xunjiang Gu (University of Toronto), Boris Ivanovic (NVIDIA Research)
CodeAutonomous DrivingGraph Neural NetworkTransformerSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper proposes modeling the uncertainty of the position and category of map elements in online HD map estimation, and directly injecting this uncertainty information into the trajectory prediction model to enhance prediction performance.
🎯 What it does: This paper proposes an evolutionary partitioning and divide-and-conquer (PDAC) framework, which first decomposes the severe undersampling problem under high compression rates into a series of progressively decreasing moderate undersampling subproblems, and at each iteration stage, only recovers the missing information corresponding to the subspace;
Prompt Highlighter: Interactive Control for Multi-Modal LLMs
Yuechen Zhang, Jiaya Jia
CodeGenerationData SynthesisTransformerLarge Language ModelPrompt EngineeringVision Language ModelTextMultimodality
🎯 What it does: A token-level highlighting interactive reasoning method called Prompt Highlighter is proposed, which enables controllable generation of multimodal LLMs without training any models.
Jinyoung Park (Korea University), Hyunwoo J. Kim (Korea University)
CodeMeta LearningTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: ProMetaR framework is proposed to automatically adjust regularization and soft prompts through meta-learning to enhance the generalization ability of VLM tasks.
🎯 What it does: Proposes the Prompt-Free Diffusion framework, which directly drives the text-to-image diffusion model using reference images and optional structural conditions without the need for text prompts.
🎯 What it does: This paper proposes a weakly supervised 3D object detection method called Prompt3D, which utilizes random text prompts to generate diverse 3D shapes and construct synthetic scenes, achieving domain adaptation through prototype proposal feature alignment.
PromptAD: Learning Prompts with only Normal Samples for Few-Shot Anomaly Detection
Xiaofan Li (East China Normal University), Lizhuang Ma (East China Normal University)
CodeAnomaly DetectionPrompt EngineeringVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes PromptAD, which utilizes single-class normal samples to automatically learn prompts to enhance few-shot anomaly detection.
Prompting Vision Foundation Models for Pathology Image Analysis
Chong Yin (Hong Kong Baptist University), Pong C. Yuen (Hong Kong Baptist University)
CodeClassificationRecognitionAnomaly DetectionTransformerPrompt EngineeringImageBiomedical Data
🎯 What it does: A quantitative attribute-based prompting method (QAP) is proposed for liver pathological image analysis, transforming spatial (K-function) and morphological (histogram) attributes into visual prompts to enhance model performance.
🎯 What it does: This paper proposes a two-stage Prompt-to-Simulate method (ProS) based on CLIP, which learns domain and semantic Prompt units and utilizes a context-aware Prompt simulator to generate dynamic Prompts for universal cross-domain retrieval (UCDR).
ProTeCt: Prompt Tuning for Taxonomic Open Set Classification
Tz-Ying Wu (University of California), Nuno Vasconcelos (University of California)
CodeClassificationTransformerPrompt EngineeringVision Language ModelImage
🎯 What it does: A taxonomic open set (TOS) classification method called ProTeCt is proposed for visual-language foundational models, which maintains prediction consistency across different levels of label sets.
Psychometry: An Omnifit Model for Image Reconstruction from Human Brain Activity
Ruijie Quan (Zhejiang University), Yi Yang (Zhejiang University)
CodeGenerationRetrievalTransformerMixture of ExpertsDiffusion modelContrastive LearningImageMagnetic Resonance Imaging
🎯 What it does: A unified omnifit model called Psychometry was constructed to reconstruct natural images from fMRI data of multiple subjects using a single network.
🎯 What it does: A post-training quantization framework PTQ4SAM is proposed to specifically address the bimodal activation distribution and diversified softmax distribution issues in the Segment Anything Model (SAM).
🎯 What it does: A Point-Trajectory Transformer (PTT) is proposed, which utilizes only the current frame point cloud and multi-frame predicted box trajectories for temporal 3D object detection, significantly reducing memory usage.
🎯 What it does: Proposes Puff-Net, an efficient style transfer framework that combines pure content and pure style feature extractors with a transformer that only contains an encoder;
Querying as Prompt: Parameter-Efficient Learning for Multimodal Language Model
Tian Liang (Zhejiang University), Qiang Zhu (Beijing Information Science and Technology University)
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringVideoTextMultimodality
🎯 What it does: A parameter-efficient multimodal language model learning strategy called QaP is proposed, which injects multimodal information into a frozen pre-trained language model through queries as prompts.
🎯 What it does: A R-Cyclic Diffuser has been developed, which achieves 3D clothed human reconstruction from a single view by combining the Zero-1-to-3 latent diffusion method with a pixel-aligned implicit model.
🎯 What it does: A radar-camera multimodal 3D object detection framework RCBEVDet is proposed, which integrates radar and multi-view camera information in the bird's-eye view (BEV) space to enhance detection accuracy and robustness.
RCooper: A Real-world Large-scale Dataset for Roadside Cooperative Perception
Ruiyang Hao (Tsinghua University), Zaiqing Nie (Tsinghua University)
CodeObject DetectionObject TrackingAutonomous DrivingSimultaneous Localization and MappingImageMultimodalityPoint Cloud
🎯 What it does: RCooper has been released, which is the first large-scale, real-world roadside collaborative perception dataset that supports detection and tracking tasks.
🎯 What it does: This paper presents REACTO, which reconstructs 3D models of general movable parts from single-segment videos, taking into account shape, texture, and motion.
🎯 What it does: A retrieval-based image-driven robot motion planning framework called READ (Retrieval-Enhanced Asymmetric Diffusion) is proposed, which utilizes retrieval to obtain initial motion and refines it through asymmetric diffusion.
Real-Time Exposure Correction via Collaborative Transformations and Adaptive Sampling
Ziwen Li (Huazhong University of Science and Technology), Nong Sang (Huazhong University of Science and Technology)
CodeRestorationConvolutional Neural NetworkImage
🎯 What it does: A collaborative transformation framework CoTF is proposed to achieve real-time exposure correction by combining global 3D LUT with local pixel-level transformations.
Real-Time Neural BRDF with Spherically Distributed Primitives
Yishun Dou (Shanghai Jiao Tong University), Junxiang Ke (Huawei)
CodeMesh
🎯 What it does: This paper presents NeuBRDF, a lightweight neural reflection model that decomposes 4D BRDF into two low-dimensional hemispherical feature grids, and achieves real-time rendering evaluation at millisecond speeds through quantifiable neural reflection primitives (codebook) and polar coordinate grids (HEALPix).
🎯 What it does: This paper constructs a large-scale multi-sensor mobile image denoising dataset MIDD and proposes an efficient segmentation network SplitterNet, which supports processing 8MP images on mobile GPUs/NPUs in less than 1 second.
🎯 What it does: A pluggable proposal learning framework (TSP-Net) is proposed, which aligns confidence and proposal quality in point-level weakly supervised temporal action localization through center score learning and alignment boundary adaptation.
🎯 What it does: A feature reconstruction-based unsupervised defect detection framework, RealNet, is proposed, which achieves high-precision defect detection and localization using only normal images.
RecDiffusion: Rectangling for Image Stitching with Diffusion Models
Tianhao Zhou (University of Electronic Science and Technology of China), Shuaicheng Liu (University of Electronic Science and Technology of China)
CodeGenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes RecDiffusion, which utilizes motion diffusion models and content diffusion models for the rectification of stitched images.
🎯 What it does: This paper proposes to average pool the semantic regions generated by SAM (and SLIC) with strong self-supervised features like DINOv2 to obtain a sparse yet semantically rich 'region representation', and uses simple linear/MLP/Transformer decoders to accomplish tasks such as semantic segmentation, object retrieval, multi-view segmentation, and video action classification.
🎯 What it does: A mutual prompting learning framework is proposed, combining density regressors with head segmenters, utilizing point prompts and context prompts to correct label variance and improve crowd counting accuracy.
🎯 What it does: A relaxed contrastive learning framework FedRCL is proposed to address the gradient bias and feature representation collapse issues caused by data heterogeneity in federated learning.
🎯 What it does: This paper proposes a training framework called RepAn that combines reparameterization (Rep) with simulated annealing. In each iteration, it first compresses a multi-branch network into a single branch, then expands new parallel branches for learning, thereby achieving the effects of incremental learning and ensemble learning.
Jiawei Liu (Chinese Academy of Sciences), Liangqiong Qu (University of Hong Kong)
CodeRestorationGenerationDiffusion modelImage
🎯 What it does: A Residual Denoising Diffusion Model (RDDM) is proposed, which splits the traditional single denoising diffusion process into two subprocesses: residual diffusion and noise diffusion, achieving unification and interpretability of generation and recovery tasks.
🎯 What it does: The RECAP framework is proposed, which adopts an iterative pruning-finetune-update three-stage cycle, loading and training only sub-networks on the GPU, thereby achieving memory-efficient fine-tuning of large-scale Transformers.
🎯 What it does: This study investigates exemplar-free continual learning scenarios and proposes an Adversarial Drift Compensation (ADC) method that utilizes adversarial samples to estimate and compensate for the feature drift of old classes, aiming to enhance class-incremental learning performance under small initial tasks.
🎯 What it does: A dual optimization framework (Dual-Optimization) and complex exponential encoding (ACM) are proposed to address the boundary discontinuity problem in inclined object detection.
🎯 What it does: A multi-contrast MRI super-resolution method called DiffMSR, which combines a latent diffusion model with a Prior-Guide Large Window Transformer, is designed to generate high-quality, distortion-free super-resolved images in just 4 inference steps.
🎯 What it does: A systematic evaluation of the existing image generation evaluation metric FID is conducted, pointing out its assumption failures, low sample efficiency, and inconsistency with human subjective assessments, and a new metric CMMD based on CLIP embeddings and MMD is proposed.
Rethinking Interactive Image Segmentation with Low Latency High Quality and Diverse Prompts
Qin Liu (University of North Carolina at Chapel Hill), Marc Niethammer (University of North Carolina at Chapel Hill)
CodeSegmentationTransformerSupervised Fine-TuningImageVideoBiomedical Data
🎯 What it does: A novel interactive image segmentation framework called SegNext is proposed, which achieves low latency, high quality, and supports various interactive prompts through dense prompt representation and fusion.
Rethinking Multi-view Representation Learning via Distilled Disentangling
Guanzhou Ke (Beijing Jiaotong University), Shengfeng He (Singapore Management University)
CodeRepresentation LearningAuto EncoderImage
🎯 What it does: A multi-view representation learning framework called MRDD is proposed, which first learns view-consistent representations through Masked Cross-View Prediction, and then uses Distilled Disentangling to eliminate redundancy between consistency and view-specificity, resulting in high-quality view-consistent and specific representations.
Rethinking Prior Information Generation with CLIP for Few-Shot Segmentation
Jin Wang (China University of Petroleum), Weifeng Liu (China University of Petroleum)
CodeSegmentationContrastive LearningImage
🎯 What it does: This paper proposes a method for generating CLIP-based prior information without additional training, combining visual-text alignment and visual-visual matching to achieve fine localization and global guidance for few-shot semantic segmentation.
🎯 What it does: A new VQ tokenizer SeQ-GAN is proposed, which employs a two-stage training to balance semantic compression and detail preservation, and constructs a visualization process to evaluate the impact of different VQ tokenizers on generative Transformers.
🎯 What it does: Proposes a method for adaptively adjusting the Classifier-Free Guidance (CFG) scale for different semantic regions in text-to-image diffusion models;
🎯 What it does: A local pixel relationship feature based on a generative model upsampling layer (Neighboring Pixel Relationships, NPR) is proposed for generalizing deep fake detection;
🎯 What it does: A multi-scale self-supervised pre-training framework called SatMAE++ is proposed, which utilizes masked autoencoders and convolutional upsampling blocks to reconstruct satellite images at multiple scales, thereby enhancing the feature representation of multispectral and optical remote sensing images.
🎯 What it does: This paper proposes a mixed-precision quantization framework for one-shot training and search, which achieves joint optimization of multiple bit widths by sharing weights during the training phase, and obtains the best bit width configuration through greedy search during the search phase without the need for additional retraining.
Jooyeon Kim (Korea University), Hyunwoo J. Kim (Korea University)
CodeObject DetectionTransformerLarge Language ModelVision Language ModelImageRetrieval-Augmented Generation
🎯 What it does: This paper proposes a framework called RALF that enhances open vocabulary object detection models by retrieving negative class vocabulary and utilizing 'spoken concepts' generated by large language models.
🎯 What it does: This study investigates the effects of adversarial training under long-tail distribution, finding that Balanced Softmax Loss is key to RoBal, and proposes AT-BSL, which combines various data augmentation methods to alleviate robust overfitting and significantly enhance robustness.
🎯 What it does: To address the 'counterfactual' problem in visual-language tasks, the authors propose a counterfactual expression generation method based on fine-grained attributes (CSG) and an end-to-end model (C-REC) that simultaneously performs object localization and counterfactual determination.
Revisiting Global Translation Estimation with Feature Tracks
Peilin Tao (Institute of Automation, Chinese Academy of Sciences), Shuhan Shen (Institute of Automation, Chinese Academy of Sciences)
CodePose EstimationAutonomous DrivingSimultaneous Localization and MappingImage
🎯 What it does: A hybrid explicit global translation estimation framework HETA is proposed, which integrates relative translation and feature trajectories to simultaneously estimate camera position and 3D points.
🎯 What it does: A hierarchical frequency domain fusion network (HFIN) is proposed for full-resolution multispectral fusion (pansharpening) of remote sensing images.
Revisiting the Domain Shift and Sample Uncertainty in Multi-source Active Domain Transfer
Wenqiao Zhang (Zhejiang University), Siliang Tang
CodeClassificationDomain AdaptationImage
🎯 What it does: This paper proposes a Multi-source Active Domain Adaptation (MADA) framework that can transfer multi-source domain knowledge to the target domain and improve classification performance with only a limited number of labeled target samples.
🎯 What it does: This paper theoretically analyzes and empirically proves that element-wise multiplication (star operation) can implicitly map low-dimensional features to high-dimensional nonlinear spaces, and based on this, designs a minimalist and efficient network called StarNet for ImageNet image classification.
🎯 What it does: This paper proposes a neural volume rendering method for 3D reconstruction using multi-view reflectance and normal maps, suitable for multi-view photometric stereo (MVPS).
🎯 What it does: A training framework based on adversarial frequency domain mixing (AFM) is proposed to enhance the robustness of image denoising networks against unknown real noise distributions.
Robust Self-calibration of Focal Lengths from the Fundamental Matrix
Viktor Kocur (Comenius University), Zuzana Kukelova (Czech Technical University in Prague)
CodePose EstimationOptimizationSimultaneous Localization and MappingImage
🎯 What it does: An efficient iterative self-calibration method based on the Kruppa equation is proposed, with a simple check for the generation of virtual focal length models added to RANSAC;
🎯 What it does: The study fine-tunes a pre-trained stereo matching network on real-world data while maintaining its generalization ability on unseen domains, and proposes a dynamic filtering and fusion framework (DKT) based on EMA teachers to achieve this goal.
🎯 What it does: This paper proposes a framework for directly estimating the intermediate distortion flow from rolling shutter (RS) images to global shutter (GS) images, and recovers GS images through reverse warping.
S2MVTC: a Simple yet Efficient Scalable Multi-View Tensor Clustering
Zhen Long (University of Electronic Science and Technology of China), Ce Zhu (University of Electronic Science and Technology of China)
CodeOptimizationImageVideo
🎯 What it does: This paper proposes a scalable multi-view clustering algorithm based on tensors, S2MVTC, which directly learns the intra/inter-view correlations of embedded features and achieves efficient clustering through anchor graphs and tensor low-frequency approximation.
SaCo Loss: Sample-wise Affinity Consistency for Vision-Language Pre-training
Sitong Wu (Chinese University of Hong Kong), Jiaya Jia (Chinese University of Hong Kong)
CodeRetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A loss function called SaCo Loss, based on sample-level similarity consistency, is designed to enhance the similarity consistency between different modalities in visual-language pre-training models, and can be used in both zero-shot pre-training and continual pre-training.
🎯 What it does: A hierarchical saliency filtering and query refinement DETR framework is proposed, reducing encoding and selection redundancy, significantly improving small object detection performance.
🎯 What it does: Proposes the SAM-6D framework to achieve zero-shot 6D object pose estimation: first, use SAM to generate candidate segmentations, then filter target instances using semantic, appearance, and geometric matching scores; finally, predict poses using a two-stage point matching network based on background tokens (coarse-fine).
SC-Tune: Unleashing Self-Consistent Referential Comprehension in Large Vision Language Models
Tongtian Yue (University of Chinese Academy of Sciences), Jing Liu (University of Chinese Academy of Sciences)
CodeRecognitionObject DetectionGenerationTransformerReinforcement LearningVision Language ModelImageMultimodality
🎯 What it does: In large-scale visual language models (LVLM), a self-consistency tuning (SC-Tune) framework is proposed, which enhances the model's self-consistency in object-level reference expression generation (REG) and localization (REC) through cyclic descriptor-locator dual-modal training, improving generalization ability on unseen data.
Scalable 3D Registration via Truncated Entry-wise Absolute Residuals
Tianyu Huang (Chinese University of Hong Kong), Yun-Hui Liu (University of Pennsylvania)
CodeAutonomous DrivingOptimizationPoint Cloud
🎯 What it does: A robust 3D registration method based on Truncated Element-wise Absolute Residual (TEAR) is proposed, utilizing decomposition and branch-and-bound to achieve global optimality and scalability to tens of millions of point pairs.
🎯 What it does: This study investigates the scaling laws of generating synthetic images from text-to-image models and compares their effectiveness in supervised classification and CLIP training.
Scene Adaptive Sparse Transformer for Event-based Object Detection
Yansong Peng (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
CodeObject DetectionTransformerImage
🎯 What it does: A sparse Transformer architecture for event cameras (SAST) is proposed, which significantly reduces computational load while maintaining detection performance through window-tag co-sparsification.
SchurVINS: Schur Complement-Based Lightweight Visual Inertial Navigation System
Yunfei Fan (ByteDance), Guidong Wang (ByteDance)
CodeOptimizationComputational EfficiencySimultaneous Localization and MappingVideo
🎯 What it does: A lightweight visual-inertial navigation system based on Schur complement (SchurVINS) is proposed, which utilizes EKF for joint estimation of attitude and landmarks within a sliding window, significantly reducing computational complexity.
🎯 What it does: Recover 3D scene representation from a single compressed image and generate multi-view consistent high frame rate images based on that representation.
🎯 What it does: This paper proposes MixCon3D, which constructs a unified 3D object-level representation by fusing multi-view RGB and point cloud features, and achieves open-source 3D understanding through tri-modal (image, text, point cloud) contrastive learning.