Few-shot Class-incremental learning (FSCIL) is a challenging task in machine learning that aims to recognize new classes from a limited number of instances while preserving the ability to classify previously learned classes without retraining the entire model. This presents challenges in updating the model with new classes using limited training data, particularly in balancing acquiring new knowledge while retaining the old. We propose a novel method named Multiple Mxing Self-Distillation (M2SD) during the training phase to address these issues. Specifically, we propose a dual-branch structure that facilitates the expansion of the entire feature space to accommodate new classes. Furthermore, we introduce a feature enhancement component that can pass additional enhanced information back to the base network by self-distillation, resulting in improved classification performance upon adding new classes. After training, we discard both structures, leaving only the primary network to classify new class instances. Extensive experiments demonstrate that our approach achieves superior performance over previous state-of-the-art methods.
论文检索
输入标题、作者或关键词，从 8,216 篇学术成果中精准定位
The goal of this paper is to alleviate the training cost for few-shot semantic segmentation (FSS) models. Despite that FSS in nature improves model generalization to new concepts using only a handful of test exemplars, it relies on strong supervision from a considerable amount of labeled training data for base classes. However, collecting pixel-level annotations is notoriously expensive and time-consuming, and small-scale training datasets convey low information density that limits test-time generalization. To resolve the issue, we take a pioneering step towards label-efficient training of FSS models from fully unlabeled training data, or additionally a few labeled samples to enhance the performance. This motivates an approach based on a novel unsupervised meta-training paradigm. In particular, the approach first distills pre-trained unsupervised pixel embedding into compact semantic clusters from which a massive number of pseudo meta-tasks is constructed. To mitigate the noise in the pseudo meta-tasks, we further advocate a robust Transformer-based FSS model with a novel prototype-based cross-attention design. Extensive experiments have been conducted on two standard benchmarks, i.e., PASCAL-5i and COCO-20i, and the results show that our method produces impressive performance without any annotations, and is comparable to fully supervised competitors even using only 20% of the annotations. Our code is available at: https://github.com/SSSKYue/UMTFSS.
Clustering-based methods are emerging as a ubiquitous technology in unsupervised object Re-Identification (ReID), which alternate between pseudo-label generation and representation learning. Recent advances in this field mainly fall into two groups: pseudo-label correction and robust representation learning. Differently, in this work, we improve unsupervised object ReID from feature calibration, a completely different but complementary insight from the current approaches. Specifically, we propose to insert a conceptually simple yet empirically powerful Feature Calibration Module (FCM) before pseudo-label generation. In practice, FCM calibrates the features using a nonparametric graph attention network, enforcing similar instances to move together in the feature space while allowing dissimilar instances to separate. As a result, we can generate more reliable pseudo-labels using the calibrated features and further improve subsequent representation learning. FCM is simple, effective, parameter-free, training-free, plug-and-play, and can be considered as a catalyst, increasing the ’chemical reaction’ between pseudo-label generation and representation learning. Moreover, it maintains the efficiency of testing time with negligible impact on training time. In this paper, we insert FCM into a simple baseline. Experiments across different scenarios and benchmarks show that FCM consistently improves the baseline (e.g., 8.2% mAP gain on MSMT17), and achieves the new state-of-the-art results. Code is available at: https://github.com/lhf12278/FCM-ReID.
Audio-driven talking head synthesis is a promising topic with wide applications in digital human, film making and virtual reality. Recent NeRF-based approaches have shown superiority in quality and fidelity compared to previous studies. However, when it comes to few-shot talking head generation, a practical scenario where only few seconds of talking video is available for one identity, two limitations emerge: 1) they either have no base model, which serves as a facial prior for fast convergence, or ignore the importance of audio when building the prior; 2) most of them overlook the degree of correlation between different face regions and audio, e.g., mouth is audio related, while ear is audio independent. In this paper, we present Audio Enhanced Neural Radiance Field (AE-NeRF) to tackle the above issues, which can generate realistic portraits of a new speaker with few-shot dataset. Specifically, we introduce an Audio Aware Aggregation module into the feature fusion stage of the reference scheme, where the weight is determined by the similarity of audio between reference and target image. Then, an Audio-Aligned Face Generation strategy is proposed to model the audio related and audio independent regions respectively, with a dual-NeRF framework. Extensive experiments have shown AE-NeRF surpasses the state-of-the-art on image fidelity, audio-lip synchronization, and generalization ability, even in limited training set or training iterations.
The task of Composed Image Retrieval (CoIR) involves queries that combine image and text modalities, allowing users to express their intent more effectively. However, current CoIR datasets are orders of magnitude smaller compared to other vision and language (V&L) datasets. Additionally, some of these datasets have noticeable issues, such as queries containing redundant modalities. To address these shortcomings, we introduce the Large Scale Composed Image Retrieval (LaSCo) dataset, a new CoIR dataset which is ten times larger than existing ones. Pre-training on our LaSCo, shows a noteworthy improvement in performance, even in zero-shot. Furthermore, we propose a new approach for analyzing CoIR datasets and methods, which detects modality redundancy or necessity, in queries. We also introduce a new CoIR baseline, the Cross-Attention driven Shift Encoder (CASE). This baseline allows for early fusion of modalities using a cross-attention module and employs an additional auxiliary task during training. Our experiments demonstrate that this new baseline outperforms the current state-of-the-art methods on established benchmarks like FashionIQ and CIRR.
Masked Image Modeling (MIM) methods, like Masked Autoencoders (MAE), efficiently learn a rich representation of the input. However, for adapting to downstream tasks, they require a sufficient amount of labeled data since their rich features code not only objects but also less relevant image background. In contrast, Instance Discrimination (ID) methods focus on objects. In this work, we study how to combine the efficiency and scalability of MIM with the ability of ID to perform downstream classification in the absence of large amounts of labeled data. To this end, we introduce Masked Autoencoder Contrastive Tuning (MAE-CT), a sequential approach that utilizes the implicit clustering of the Nearest Neighbor Contrastive Learning (NNCLR) objective to induce abstraction in the topmost layers of a pre-trained MAE. MAE-CT tunes the rich features such that they form semantic clusters of objects without using any labels. Notably, MAE-CT does not rely on hand-crafted augmentations and frequently achieves its best performances while using only minimal augmentations (crop & flip). Further, MAE-CT is compute efficient as it requires at most 10% overhead compared to MAE re-training. Applied to large and huge Vision Transformer (ViT) models, MAE-CT excels over previous self-supervised methods trained on ImageNet in linear probing, k-NN and low-shot classification accuracy as well as in unsupervised clustering accuracy. With ViT-H/16 MAE-CT achieves a new state-of-the-art in linear probing of 82.2%. Project page: github.com/ml-jku/MAE-CT.
Existing learning-based methods for object pose estimation in RGB images are mostly model-specific or category based. They lack the capability to generalize to new object categories at test time, hence severely hindering their practicability and scalability. Notably, recent attempts have been made to solve this issue, but they still require accurate 3D data of the object surface at both train and test time. In this paper, we introduce a novel approach that can estimate in a single forward pass the pose of objects never seen during training, given minimum input. In contrast to existing state-of-the-art approaches, which rely on task-specific modules, our proposed model is entirely based on a transformer architecture, which can benefit from recently proposed 3D-geometry general pretraining. We conduct extensive experiments and report state-of-the-art one-shot performance on the challenging LINEMOD benchmark. Finally, extensive ablations allow us to determine good practices with this relatively new type of architecture in the field.
Image matching and object detection are two fundamental and challenging tasks, while many related applications consider them two individual tasks (i.e. task-individual). In this paper, a collaborative framework called MatchDet (i.e. task-collaborative) is proposed for image matching and object detection to obtain mutual improvements. To achieve the collaborative learning of the two tasks, we propose three novel modules, including a Weighted Spatial Attention Module (WSAM) for Detector, and Weighted Attention Module (WAM) and Box Filter for Matcher. Specifically, the WSAM highlights the foreground regions of target image to benefit the subsequent detector, the WAM enhances the connection between the foreground regions of pair images to ensure high-quality matches, and Box Filter mitigates the impact of false matches. We evaluate the approaches on a new benchmark with two datasets called Warp-COCO and miniScanNet. Experimental results show our approaches are effective and achieve competitive improvements.
We present a real-time method for robust estimation of multiple instances of geometric models from noisy data. Geometric models such as vanishing points, planar homographies or fundamental matrices are essential for 3D scene analysis. Previous approaches discover distinct model instances in an iterative manner, thus limiting their potential for speedup via parallel computation. In contrast, our method detects all model instances independently and in parallel. A neural network segments the input data into clusters representing potential model instances by predicting multiple sets of sample and inlier weights. Using the predicted weights, we determine the model parameters for each potential instance separately in a RANSAC-like fashion. We train the neural network via task-specific loss functions, i.e. we do not require a ground-truth segmentation of the input data. As suitable training data for homography and fundamental matrix fitting is scarce, we additionally present two new synthetic datasets. We demonstrate state-of-the-art performance on these as well as multiple established datasets, with inference times as small as five milliseconds per image.
The lack of large-scale data seriously hinders the development of palmprint recognition. Recent approaches address this issue by generating large-scale realistic pseudo palmprints from Bézier curves. However, the significant difference between Bézier curves and real palmprints limits their effectiveness. In this paper, we divide the Bézier-Real difference into creases and texture differences, thus reducing the generation difficulty. We introduce a new palm crease energy (PCE) domain as a bridge from Bézier curves to real palmprints and propose a two-stage generation model. The first stage generates PCE images (realistic creases) from Bézier curves, and the second stage outputs realistic palmprints (realistic texture) with PCE images as input. In addition, we also design a lightweight plug-and-play line feature enhancement block to facilitate domain transfer and improve recognition performance. Extensive experimental results demonstrate that the proposed method surpasses state-of-the-art methods. Under extremely few data settings like 40 IDs (only 2.5% of the total training set), our model achieves a 29% improvement over RPG-Palm and outperforms ArcFace with 100% training set by more than 6% in terms of TAR@FAR=1e-6.
MWSIS: Multimodal Weakly Supervised Instance Segmentation with 2D Box Annotations for Autonomous Driving
PDF ↗Instance segmentation is a fundamental research in computer vision, especially in autonomous driving. However, manual mask annotation for instance segmentation is quite time-consuming and costly. To address this problem, some prior works attempt to apply weakly supervised manner by exploring 2D or 3D boxes. However, no one has ever successfully segmented 2D and 3D instances simultaneously by only using 2D box annotations, which could further reduce the annotation cost by an order of magnitude. Thus, we propose a novel framework called Multimodal Weakly Supervised Instance Segmentation (MWSIS), which incorporates various fine-grained label correction modules for both 2D and 3D modalities, along with a new multimodal cross-supervision approach. In the 2D pseudo label generation branch, the Instance-based Pseudo Mask Generation (IPG) module utilizes predictions for self-supervised correction. Similarly, in the 3D pseudo label generation branch, the Spatial-based Pseudo Label Generation (SPG) module generates pseudo labels by incorporating the spatial prior information of the point cloud. To further refine the generated pseudo labels, the Point-based Voting Label Correction (PVC) module utilizes historical predictions for correction. Additionally, a Ring Segment-based Label Correction (RSC) module is proposed to refine the predictions by leveraging the depth prior information from the point cloud. Finally, the Consistency Sparse Cross-modal Supervision (CSCS) module reduces the inconsistency of multimodal predictions by response distillation. Particularly, transferring the 3D backbone to downstream tasks not only improves the performance of the 3D detectors, but also outperforms fully supervised instance segmentation with only 5% fully supervised annotations. On the Waymo dataset, the proposed framework demonstrates significant improvements over the baseline, especially achieving 2.59% mAP and 12.75% mAP increases for 2D and 3D instance segmentation tasks, respectively. The code is available at https://github.com/jiangxb98/mwsis-plugin.
Recent significant advancements in diffusion models have revolutionized image generation, enabling the synthesis of highly realistic images with text-based guidance. These breakthroughs have paved the way for constructing datasets via generative artificial intelligence (AI), offering immense potential for various applications. However, two critical challenges hinder the widespread adoption of synthesized data: computational cost and the generation of peculiar images. While computational costs have improved through various approaches, the issue of peculiar image generation remains relatively unexplored. Existing solutions rely on heuristics, extra training, or AI-based post-processing to mitigate this problem. In this paper, we present a novel approach to address both issues simultaneously. We establish that both gradient descent and diffusion sampling are specific cases of the generalized expectation maximization algorithm. We hypothesize and empirically demonstrate that peculiar image generation is akin to the local minima problem in optimization. Inspired by optimization techniques, we apply naive momentum and positive-negative momentum to diffusion sampling. Last, we propose new metrics to evaluate the peculiarity. Experimental results show momentum effectively prevents peculiar image generation without extra computation.
Existing GAN-based approaches to caricature generation mainly focus on exaggerating a character’s global facial structure. This often leads to the failure in highlighting significant facial features such as big eyes and hook nose. To address this limitation, we propose a new approach termed as G2L-CariGAN, which uses feature maps of spatial dimensions instead of latent codes for geometric exaggeration. G2L-CariGAN first exaggerates the global facial structure of the character on a low-dimensional feature map and then exaggerates its local facial features on a high-dimensional feature map. Moreover, we develop a caricature identity loss function based on feature maps, which well retains the character's identity after exaggeration. Our experiments have demonstrated that G2L-CariGAN outperforms the state-of-arts in terms of the quality of exaggerating a character and retaining its identity.
Vision Language Models (VLMs), which extend Large Language Models (LLM) by incorporating visual understanding capability, have demonstrated significant advancements in addressing open-ended visual question-answering (VQA) tasks. However, these models cannot accurately interpret images infused with text, a common occurrence in real-world scenarios. Standard procedures for extracting information from images often involve learning a fixed set of query embeddings. These embeddings are designed to encapsulate image contexts and are later used as soft prompt inputs in LLMs. Yet, this process is limited to the token count, potentially curtailing the recognition of scenes with text-rich context. To improve upon them, the present study introduces BLIVA: an augmented version of InstructBLIP with Visual Assistant. BLIVA incorporates the query embeddings from InstructBLIP and also directly projects encoded patch embeddings into the LLM, a technique inspired by LLaVA. This approach assists the model to capture intricate details potentially missed during the query decoding process. Empirical evidence demonstrates that our model, BLIVA, significantly enhances performance in processing text-rich VQA benchmarks (up to 17.76% in OCR-VQA benchmark) and in undertaking general (not particularly text-rich) VQA benchmarks (up to 7.9% in Visual Spatial Reasoning benchmark), and achieved 17.72% overall improvement in a comprehensive multimodal LLM benchmark (MME), comparing to our baseline InstructBLIP. BLIVA demonstrates significant capability in decoding real-world images, irrespective of text presence. To demonstrate the broad industry applications enabled by BLIVA, we evaluate the model using a new dataset comprising YouTube thumbnails paired with question-answer sets across 11 diverse categories. For researchers interested in further exploration, our code and models are freely accessible at https://github.com/mlpc-ucsd/BLIVA.
How to frame (or crop) a photo often depends on the image subject and its context; e.g., a human portrait. Recent works have defined the subject-aware image cropping task as a nuanced and practical version of image cropping. We propose a weakly-supervised approach (GenCrop) to learn what makes a high-quality, subject-aware crop from professional stock images. Unlike supervised prior work, GenCrop requires no new manual annotations beyond the existing stock image collection. The key challenge in learning from this data, however, is that the images are already cropped and we do not know what regions were removed. Our insight is to combine a library of stock images with a modern, pre-trained text-to-image diffusion model. The stock image collection provides diversity, and its images serve as pseudo-labels for a good crop. The text-image diffusion model is used to out-paint (i.e., outward inpainting) realistic uncropped images. Using this procedure, we are able to automatically generate a large dataset of cropped-uncropped training pairs to train a cropping model. Despite being weakly-supervised, GenCrop is competitive with state-of-the-art supervised methods and significantly better than comparable weakly-supervised baselines on quantitative and qualitative evaluation metrics.
The aim of audio-visual segmentation (AVS) is to precisely differentiate audible objects within videos down to the pixel level. Traditional approaches often tackle this challenge by combining information from various modalities, where the contribution of each modality is implicitly or explicitly modeled. Nevertheless, the interconnections between different modalities tend to be overlooked in audio-visual modeling. In this paper, inspired by the human ability to mentally simulate the sound of an object and its visual appearance, we introduce a bidirectional generation framework. This framework establishes robust correlations between an object's visual characteristics and its associated sound, thereby enhancing the performance of AVS. To achieve this, we employ a visual-to-audio projection component that reconstructs audio features from object segmentation masks and minimizes reconstruction errors. Moreover, recognizing that many sounds are linked to object movements, we introduce an implicit volumetric motion estimation module to handle temporal dynamics that may be challenging to capture using conventional optical flow methods. To showcase the effectiveness of our approach, we conduct comprehensive experiments and analyses on the widely recognized AVSBench benchmark. As a result, we establish a new state-of-the-art performance level in the AVS benchmark, particularly excelling in the challenging MS3 subset which involves segmenting multiple sound sources. Code is released in: https://github.com/OpenNLPLab/AVS-bidirectional.
This paper proposes an end-to-end framework for generating 3D human pose datasets using Neural Radiance Fields (NeRF). Public datasets generally have limited diversity in terms of human poses and camera viewpoints, largely due to the resource-intensive nature of collecting 3D human pose data. As a result, pose estimators trained on public datasets significantly underperform when applied to unseen out-of-distribution samples. Previous works proposed augmenting public datasets by generating 2D-3D pose pairs or rendering a large amount of random data. Such approaches either overlook image rendering or result in suboptimal datasets for pre-trained models. Here we propose PoseGen, which learns to generate a dataset (human 3D poses and images) with a feedback loss from a given pre-trained pose estimator. In contrast to prior art, our generated data is optimized to improve the robustness of the pre-trained model. The objective of PoseGen is to learn a distribution of data that maximizes the prediction error of a given pre-trained model. As the learned data distribution contains OOD samples of the pre-trained model, sampling data from such a distribution for further fine-tuning a pre-trained model improves the generalizability of the model. This is the first work that proposes NeRFs for 3D human data generation. NeRFs are data-driven and do not require 3D scans of humans. Therefore, using NeRF for data generation is a new direction for convenient user-specific data generation. Our extensive experiments show that the proposed PoseGen improves two baseline models (SPIN and HybrIK) on four datasets with an average 6% relative improvement.
In recent years, object detection in deep learning has experienced rapid development. However, most existing object detection models perform well only on closed-set datasets, ignoring a large number of potential objects whose categories are not defined in the training set. These objects are often identified as background or incorrectly classified as pre-defined categories by the detectors. In this paper, we focus on the challenging problem of Novel Class Discovery and Localization (NCDL), aiming to train detectors that can detect the categories present in the training data, while also actively discover, localize, and cluster new categories. We analyze existing NCDL methods and identify the core issue: object detectors tend to be biased towards seen objects, and this leads to the neglect of unseen targets. To address this issue, we first propose an Debiased Region Mining (DRM) approach that combines class-agnostic Region Proposal Network (RPN) and class-aware RPN in a complementary manner. Additionally, we suggest to improve the representation network through semi-supervised contrastive learning by leveraging unlabeled data. Finally, we adopt a simple and efficient mini-batch K-means clustering method for novel class discovery. We conduct extensive experiments on the NCDL benchmark, and the results demonstrate that the proposed DRM approach significantly outperforms previous methods, establishing a new state-of-the-art.
Fewer Steps, Better Performance: Efficient Cross-Modal Clip Trimming for Video Moment Retrieval Using Language
PDF ↗Given an untrimmed video and a sentence query, video moment retrieval using language (VMR) aims to locate a target query-relevant moment. Since the untrimmed video is overlong, almost all existing VMR methods first sparsely down-sample each untrimmed video into multiple fixed-length video clips and then conduct multi-modal interactions with the query feature and expensive clip features for reasoning, which is infeasible for long real-world videos that span hours. Since the video is downsampled into fixed-length clips, some query-related frames may be filtered out, which will blur the specific boundary of the target moment, take the adjacent irrelevant frames as new boundaries, easily leading to cross-modal misalignment and introducing both boundary-bias and reasoning-bias. To this end, in this paper, we propose an efficient approach, SpotVMR, to trim the query-relevant clip. Besides, our proposed SpotVMR can serve as plug-and-play module, which achieves efficiency for state-of-the-art VMR methods while maintaining good retrieval performance. Especially, we first design a novel clip search model that learns to identify promising video regions to search conditioned on the language query. Then, we introduce a set of low-cost semantic indexing features to capture the context of objects and interactions that suggest where to search the query-relevant moment. Also, the distillation loss is utilized to address the optimization issues arising from end-to-end joint training of the clip selector and VMR model. Extensive experiments on three challenging datasets demonstrate its effectiveness.
Recent advances in event-based research prioritize sparsity and temporal precision. Approaches learning sparse point-based representations through graph CNNs (GCN) become more popular. Yet, these graph techniques hold lower performance than their frame-based counterpart due to two issues: (i) Biased graph structures that don't properly incorporate varied attributes (such as semantics, and spatial and temporal signals) for each vertex, resulting in inaccurate graph representations. (ii) A shortage of robust pretrained models. Here we solve the first problem by proposing a new event-based GCN (EDGCN), with a dynamic aggregation module to integrate all attributes of vertices adaptively. To address the second problem, we introduce a novel learning framework called cross-representation distillation (CRD), which leverages the dense representation of events as a cross-representation auxiliary to provide additional supervision and prior knowledge for the event graph. This frame-to-graph distillation allows us to benefit from the large-scale priors provided by CNNs while still retaining the advantages of graph-based models. Extensive experiments show our model and learning framework are effective and generalize well across multiple vision tasks.