CVPR 2024 Papers

Continual learning requires the model to learn multiple tasks sequentially. In continual learning the model should possess the ability to maintain its performance on old tasks (stability) and the ability to adapt to new tasks continuously (plasticity). Recently parameter-efficient fine-tuning (PEFT) which involves freezing a pre-trained model and injecting a small number of learnable parameters to adapt to downstream tasks has gained increasing popularity in continual learning. Although existing continual learning methods based on PEFT have demonstrated superior performance compared to those not based on PEFT most of them do not consider how to eliminate the interference of the new task on the old tasks which inhibits the model from making a good trade-off between stability and plasticity. In this work we propose a new PEFT method called interference-free low-rank adaptation (InfLoRA) for continual learning. InfLoRA injects a small number of parameters to reparameterize the pre-trained weights and shows that fine-tuning these injected parameters is equivalent to fine-tuning the pre-trained weights within a subspace. Furthermore InfLoRA designs this subspace to eliminate the interference of the new task on the old tasks making a good trade-off between stability and plasticity. Experimental results show that InfLoRA outperforms existing state-of-the-art continual learning methods on multiple datasets.

We present a novel semantic segmentation approach for incremental nuclei segmentation from histopathological images which is a very challenging task as we have to incrementally optimize existing models to make them perform well in both old and new classes without using training samples of old classes. Yet it is an indispensable component of computer-aided diagnosis systems. The proposed approach has two key techniques. First we propose a new future-class awareness mechanism by separating some potential regions for future classes from background based on their similarities to both old and new classes in the representation space. With this mechanism we can not only reserve more parameter space for future updates but also enhance the representation capability of learned features. We further propose an innovative compatibility-inspired distillation scheme to make our model take full advantage of the knowledge learned by the old model. We conducted extensive experiments on two famous histopathological datasets and the results demonstrate the proposed approach achieves much better performance than state-of-the-art approaches. The code is available at https://github.com/why19991/InSeg.

Existing object recognition models have been shown to lack robustness in diverse geographical scenarios due to domain shifts in design and context. Class representations need to be adapted to more accurately reflect an object concept under these shifts. In the absence of training data from target geographies we hypothesize that geographically diverse descriptive knowledge of categories can enhance robustness. For this purpose we explore the feasibility of probing a large language model for geography-based object knowledge and we examine the effects of integrating knowledge into zero-shot and learnable soft prompting with CLIP. Within this exploration we propose geography knowledge regularization to ensure that soft prompts trained on a source set of geographies generalize to an unseen target set. Accuracy gains over prompting baselines on DollarStreet while training only on Europe data are up to +2.8/1.2/1.6 on target data from Africa/Asia/Americas and +4.6 overall on the hardest classes. Competitive performance is shown vs. few-shot target training and analysis is provided to direct future study of geographical robustness.

Inspired by the long-range modeling ability of ViTs large-kernel convolutions are widely studied and adopted recently to enlarge the receptive field and improve model performance like the remarkable work ConvNeXt which employs 7x7 depthwise convolution. Although such depthwise operator only consumes a few FLOPs it largely harms the model efficiency on powerful computing devices due to the high memory access costs. For example ConvNeXt-T has similar FLOPs with ResNet-50 but only achieves 60% throughputs when trained on A100 GPUs with full precision. Although reducing the kernel size of ConvNeXt can improve speed it results in significant performance degradation which poses a challenging problem: How to speed up large-kernel-based CNN models while preserving their performance. To tackle this issue inspired by Inceptions we propose to decompose large-kernel depthwise convolution into four parallel branches along channel dimension i.e. small square kernel two orthogonal band kernels and an identity mapping. With this new Inception depthwise convolution we build a series of networks namely IncepitonNeXt which not only enjoy high throughputs but also maintain competitive performance. For instance InceptionNeXt-T achieves 1.6x higher training throughputs than ConvNeX-T as well as attains 0.2% top-1 accuracy improvement on ImageNet-1K. We anticipate InceptionNeXt can serve as an economical baseline for future architecture design to reduce carbon footprint.

We propose InNeRF360 an automatic system that accurately removes text-specified objects from 360-degree Neural Radiance Fields (NeRF). The challenge is to effectively remove objects while inpainting perceptually consistent content for the missing regions which is particularly demanding for existing NeRF models due to their implicit volumetric representation. Moreover unbounded scenes are more prone to floater artifacts in the inpainted region than frontal-facing scenes as the change of object appearance and background across views is more sensitive to inaccurate segmentations and inconsistent inpainting. With a trained NeRF and a text description our method efficiently removes specified objects and inpaints visually consistent content without artifacts. We apply depth-space warping to enforce consistency across multiview text-encoded segmentations and then refine the inpainted NeRF model using perceptual priors and 3D diffusion-based geometric priors to ensure visual plausibility. Through extensive experiments in segmentation and inpainting on 360-degree and frontal-facing NeRFs we show that InNeRF360 is effective and enhances NeRF's editability. Project page: https://ivrl.github.io/InNeRF360.

Dual-camera compressive hyperspectral imaging (DCCHI) offers the capability to reconstruct 3D hyperspectral image (HSI) by fusing compressive and panchromatic (PAN) image which has shown great potential for snapshot hyperspectral imaging in practice. In this paper we introduce a novel DCCHI reconstruction network intra-inter similarity exploiting Transformer (In2SET). Our key insight is to make full use of the PAN image to assist the reconstruction. To this end we propose to use the intra-similarity within the PAN image as a proxy for approximating the intra-similarity in the original HSI thereby offering an enhanced content prior for more accurate HSI reconstruction. Furthermore we propose to use the inter-similarity to align the features between HSI and PAN images thereby maintaining semantic consistency between the two modalities during the reconstruction process. By integrating In2SET into a PAN-guided deep unrolling (PGDU) framework our method substantially enhances the spatial-spectral fidelity and detail of the reconstructed images providing a more comprehensive and accurate depiction of the scene. Experiments conducted on both real and simulated datasets demonstrate that our approach consistently outperforms existing state-of-the-art methods in terms of reconstruction quality and computational complexity. The code is available at https://github.com/2JONAS/In2SET.

To alleviate the utility degradation of deep learning image classification with differential privacy (DP) employing extra public data or pre-trained models has been widely explored. Recently the use of in-distribution public data has been investigated where tiny subsets of datasets are released publicly. In this paper we investigate a framework that leverages recent diffusion models to amplify the information of public data. Subsequently we identify data diversity and generalization gap between public and private data as critical factors addressing the limited public data. While assuming 4% of training data as public our method achieves 85.48% on CIFAR-10 with a privacy budget of ?=2 without employing extra public data for training.

3D-aware GANs offer new capabilities for view synthesis while preserving the editing functionalities of their 2D counterparts. GAN inversion is a crucial step that seeks the latent code to reconstruct input images or videos subsequently enabling diverse editing tasks through manipulation of this latent code. However a model pre-trained on a particular dataset (e.g. FFHQ) often has difficulty reconstructing images with out-of-distribution (OOD) objects such as faces with heavy make-up or occluding objects. We address this issue by explicitly modeling OOD objects from the input in 3D-aware GANs. Our core idea is to represent the image using two individual neural radiance fields: one for the in-distribution content and the other for the out-of-distribution object. The final reconstruction is achieved by optimizing the composition of these two radiance fields with carefully designed regularization. We demonstrate that our explicit decomposition alleviates the inherent trade-off between reconstruction fidelity and editability. We evaluate reconstruction accuracy and editability of our method on challenging real face images and videos and showcase favorable results against other baselines.

Neural field is an emerging paradigm in data representation that trains a neural network to approximate the given signal. A key obstacle that prevents its widespread adoption is the encoding speed---generating neural fields requires an overfitting of a neural network which can take a significant number of SGD steps to reach the desired fidelity level. In this paper we delve into the impacts of data transformations on the speed of neural field training specifically focusing on how permuting pixel locations affect the convergence speed of SGD. Counterintuitively we find that randomly permuting the pixel locations can considerably accelerate the training. To explain this phenomenon we examine the neural field training through the lens of PSNR curves loss landscapes and error patterns. Our analyses suggest that the random pixel permutations remove the easy-to-fit patterns which facilitate easy optimization in the early stage but hinder capturing fine details of the signal.

The success of large language models has inspired the computer vision community to explore image segmentation foundation model that is able to zero/few-shot generalize through prompt engineering. Segment-Anything (SAM) among others is the state-of-the-art image segmentation foundation model demonstrating strong zero/few-shot generalization. Despite the success recent studies reveal the weakness of SAM under strong distribution shift. In particular SAM performs awkwardly on corrupted natural images camouflaged images medical images etc. Motivated by the observations we aim to develop a self-training based strategy to adapt SAM to target distribution. Given the unique challenges of large source dataset high computation cost and incorrect pseudo label we propose a weakly supervised self-training architecture with anchor regularization and low-rank finetuning to improve the robustness and computation efficiency of adaptation. We validate the effectiveness on 5 types of downstream segmentation tasks including natural clean/corrupted images medical images camouflaged images and robotic images. Our proposed method is task-agnostic in nature and outperforms pre-trained SAM and state-of-the-art domain adaptation methods on almost all downstream tasks with the same testing prompt inputs.

Visual scenes are naturally organized in a hierarchy where a coarse semantic is recursively comprised of several fine details. Exploring such a visual hierarchy is crucial to recognize the complex relations of visual elements leading to a comprehensive scene understanding. In this paper we propose a Visual Hierarchy Mapper (Hi-Mapper) a novel approach for enhancing the structured understanding of the pre-trained Deep Neural Networks (DNNs). Hi-Mapper investigates the hierarchical organization of the visual scene by 1) pre-defining a hierarchy tree through the encapsulation of probability densities; and 2) learning the hierarchical relations in hyperbolic space with a novel hierarchical contrastive loss. The pre-defined hierarchy tree recursively interacts with the visual features of the pre-trained DNNs through hierarchy decomposition and encoding procedures thereby effectively identifying the visual hierarchy and enhancing the recognition of an entire scene. Extensive experiments demonstrate that Hi-Mapper significantly enhances the representation capability of DNNs leading to an improved performance on various tasks including image classification and dense prediction tasks.

Learning representations to capture the very fundamental understanding of the world is a key challenge in machine learning. The hierarchical structure of explanatory factors hidden in data is such a general representation and could be potentially achieved with a hierarchical VAE. However training a hierarchical VAE always suffers from the "posterior collapse" where the data information is hard to propagate to the higher-level latent variables hence resulting in a bad hierarchical representation. To address this issue we first analyze the shortcomings of existing methods for mitigating the "posterior collapse" from an information theory perspective then highlight the necessity of regularization for explicitly propagating data information to higher-level latent variables while maintaining the dependency between different levels. This naturally leads to formulating the inference of the hierarchical latent representation as a sequential decision process which could benefit from applying reinforcement learning (RL). Aligning RL's objective with the regularization we first introduce a "skip-generative path" to acquire a reward for evaluating the information content of an inferred latent representation and then the developed Q-value function based on it could have a consistent optimization direction of the regularization. Finally policy gradient one of the typical RL methods is employed to train a hierarchical VAE without introducing a gradient estimator. Experimental results firmly support our analysis and demonstrate that our proposed method effectively mitigates the "posterior collapse" issue learns an informative hierarchy acquires explainable latent representations and significantly outperforms other hierarchical VAE-based methods in downstream tasks.

Various transfer attack methods have been proposed to evaluate the robustness of deep neural networks (DNNs). Although manifesting remarkable performance in generating untargeted adversarial perturbations existing proposals still fail to achieve high targeted transferability. In this work we discover that the adversarial perturbations' overfitting towards source models of mediocre generalization capability can hurt their targeted transferability. To address this issue we focus on enhancing the source model's generalization capability to improve its ability to conduct transferable targeted adversarial attacks. In pursuit of this goal we propose a novel model self-enhancement method that incorporates two major components: Sharpness-Aware Self-Distillation (SASD) and Weight Scaling (WS). Specifically SASD distills a fine-tuned auxiliary model which mirrors the source model's structure into the source model while flattening the source model's loss landscape. WS obtains an approximate ensemble of numerous pruned models to perform model augmentation which can be conveniently synergized with SASD to elevate the source model's generalization capability and thus improve the resultant targeted perturbations' transferability. Extensive experiments corroborate the effectiveness of the proposed method. Notably under the black-box setting our approach can outperform the state-of-the-art baselines by a significant margin of 12.2% on average in terms of the obtained targeted transferability. Code is available at https://github.com/g4alllf/SASD.

Diffusion models emerging as powerful deep generative tools excel in various applications. They operate through a two-steps process: introducing noise into training samples and then employing a model to convert random noise into new samples (e.g. images). However their remarkable generative performance is hindered by slow training and sampling. This is due to the necessity of tracking extensive forward and reverse diffusion trajectories and employing a large model with numerous parameters across multiple timesteps (i.e. noise levels). To tackle these challenges we present a multi-stage framework inspired by our empirical findings. These observations indicate the advantages of employing distinct parameters tailored to each timestep while retaining universal parameters shared across all time steps. Our approach involves segmenting the time interval into multiple stages where we employ custom multi-decoder U-net architecture that blends time-dependent models with a universally shared encoder. Our framework enables the efficient distribution of computational resources and mitigates inter-stage interference which substantially improves training efficiency. Extensive numerical experiments affirm the effectiveness of our framework showcasing significant training and sampling efficiency enhancements on three state-of-the-art diffusion models including large-scale latent diffusion models. Furthermore our ablation studies illustrate the impact of two important components in our framework: (i) a novel timestep clustering algorithm for stage division and (ii) an innovative multi-decoder U-net architecture seamlessly integrating universal and customized hyperparameters.

In subject-driven text-to-image synthesis the synthesis process tends to be heavily influenced by the reference images provided by users often overlooking crucial attributes detailed in the text prompt. In this work we propose Subject-Agnostic Guidance (SAG) a simple yet effective solution to remedy the problem. We show that through constructing a subject-agnostic condition and applying our proposed dual classifier-free guidance one could obtain outputs consistent with both the given subject and input text prompts. We validate the efficacy of our approach through both optimization-based and encoder-based methods. Additionally we demonstrate its applicability in second-order customization methods where an encoder-based model is fine-tuned with DreamBooth. Our approach is conceptually simple and requires only minimal code modifications but leads to substantial quality improvements as evidenced by our evaluations and user studies.

How to effectively utilize the spectral and spatial characteristics of Hyperspectral Image (HSI) is always a key problem in spectral snapshot reconstruction. Recently the spectra-wise transformer has shown great potential in capturing inter-spectra similarities of HSI but the classic design of the transformer i.e. multi-head division in the spectral (channel) dimension hinders the modeling of global spectral information and results in mean effect. In addition previous methods adopt the normal spatial priors without taking imaging processes into account and fail to address the unique spatial degradation in snapshot spectral reconstruction. In this paper we analyze the influence of multi-head division and propose a novel Spectral-Spatial Rectification (SSR) method to enhance the utilization of spectral information and improve spatial degradation. Specifically SSR includes two core parts: Window-based Spectra-wise Self-Attention (WSSA) and spAtial Rectification Block (ARB). WSSA is proposed to capture global spectral information and account for local differences whereas ARB aims to mitigate the spatial degradation using a spatial alignment strategy. The experimental results on simulation and real scenes demonstrate the effectiveness of the proposed modules and we also provide models at multiple scales to demonstrate the superiority of our approach.

In this work we tackle the problem of domain generalization for object detection specifically focusing on the scenario where only a single source domain is available. We propose an effective approach that involves two key steps: diversifying the source domain and aligning detections based on class prediction confidence and localization. Firstly we demonstrate that by carefully selecting a set of augmentations a base detector can outperform existing methods for single domain generalization by a good margin. This highlights the importance of domain diversification in improving the performance of object detectors. Secondly we introduce a method to align detections from multiple views considering both classification and localization outputs. This alignment procedure leads to better generalized and well-calibrated object detector models which are crucial for accurate decision-making in safety-critical applications. Our approach is detector-agnostic and can be seamlessly applied to both single-stage and two-stage detectors. To validate the effectiveness of our proposed methods we conduct extensive experiments and ablations on challenging domain-shift scenarios. The results consistently demonstrate the superiority of our approach compared to existing methods.

Recent self-supervised models produce visual features that are not only effective at encoding image-level but also pixel-level semantics. They have been reported to obtain impressive results for dense visual semantic correspondence estimation even outperforming fully-supervised methods. Nevertheless these models still fail in the presence of challenging image characteristics such as symmetries and repeated parts. To address these limitations we propose a new semantic correspondence estimation method that supplements state-of-the-art self-supervised features with 3D understanding via a weak geometric spherical prior. Compared to more involved 3D pipelines our model provides a simple and effective way of injecting informative geometric priors into the learned representation while requiring only weak viewpoint information. We also propose a new evaluation metric that better accounts for repeated part and symmetry-induced mistakes. We show that our method succeeds in distinguishing between symmetric views and repeated parts across many object categories in the challenging SPair-71k dataset and also in generalizing to previously unseen classes in the AwA dataset.