WACV 2024 Papers

Most deep neural networks are trained under fixed network architectures and require retraining when the architecture changes. If expanding the network's size is needed, it is necessary to retrain from scratch, which is expensive. To avoid this, one can grow from a small network by adding random weights over time to gradually achieve the target network size. However, this naive approach falls short in practice as it brings too much noise to the growing process. Prior work tackled this issue by leveraging the already learned weights and training data for generating new weights through conducting a computationally expensive analysis step. In this paper, we introduce MixtureGrowth, a new approach to growing networks that circumvents the initialization overhead in prior work. Before growing, each layer in our model is generated with a linear combination of parameter templates. Newly grown layer weights are generated by using a new linear combination of existing templates for a layer. On one hand, these templates are already trained for the task, providing a strong initialization. On the other, the new coefficients provide flexibility for the added layer weights to learn something new. We show that our approach boosts top-1 accuracy over the state-of-the-art by 2-2.5% on CIFAR-100 and ImageNet datasets, while achieving comparable performance with fewer FLOPs to a larger network trained from scratch. Code is available at https://github.com/chaudatascience/mixturegrowth

Federated learning reduces the risk of information leakage, but remains vulnerable to attack. We show that well-mixed gradients provide numerical resistance to gradient inversion in neural networks. For example, we can enhance mixing gradients in a batch by choosing an appropriate loss function and drawing identical labels, and we support this with an approximate solution of batch inversion for linear layers. These simple architecture choices show no degradation to classification performance as opposed to noise perturbation defense. To accurately assess data recovery, we propose to use a variation distance metric for information leakage in images, derived from total variation. In contrast to Mean Squared Error or Structural Similarity Index metrics, it provides a continuous metric for information recovery. Finally, our empirical results of information recovery from various inversion attacks and training performance supports our defense strategies. These simple architecture choices found to be also useful for practical size of convolutional neural networks but depends on their size. We hope this work will trigger further defense studies using gradient mixing, towards achieving a trustful federation policy.

While deep learning has significantly improved ReID model accuracy under the independent and identical distribution (IID) assumption, it has also become clear that such models degrade notably when applied to an unseen novel domain due to unpredictable/unknown domain shift. Contemporary domain generalization (DG) ReID models struggle in learning domain-invariant representation through solely training on an instance classification objective. We consider that a deep learning model is heavily influenced therefore biased towards domain-specific characteristics, e.g., background clutter, scale and viewpoint variations, limiting the generalizability of the learned model, and hypothesize that the pedestrians are domain invariant owning they share the same structural characteristics. To enable ReID model to be less domain-specific from these pure pedestrians and domain-specific factors, we introduce a method that guides model learning of the primary ReID instance classification objective by a concurrent auxiliary learning objective on weakly labeled pedestrian saliency detection. To solve the problem of conflicting optimization criteria in the model parameter space between the two learning objectives, we introduce a Primary-Auxiliary Objectives Association (PAOA) mechanism to calibrate the loss gradients of the auxiliary task towards the primary learning task gradients. Benefited from the harmonious multitask learning design, our model can be extended with the recent test-time diagram to form the PAOA+, which performs on-the-fly optimization against the auxiliary objective in order to maximize the model's generative capacity in the test target domain. Experiments demonstrate the superiority of the proposed PAOA model.

Using multiple spatial modalities has been proven helpful in improving semantic segmentation performance. However, there are several real-world challenges that have yet to be addressed: (a) improving label efficiency and (b) enhancing robustness in realistic scenarios where modalities are missing at the test time. To address these challenges, we first propose a simple yet efficient multi-modal fusion mechanism Linear Fusion, that performs better than the state-of-the-art multi-modal models even with limited supervision. Second, we propose M3L: Multi-modal Teacher for Masked Modality Learning, a semi-supervised framework that not only improves the multi-modal performance but also makes the model robust to the realistic missing modality scenario using unlabeled data. We create the first benchmark for semi-supervised multi-modal semantic segmentation and also report the robustness to missing modalities. Our proposal shows an absolute improvement of up to 5% on robust mIoU above the most competitive baselines. Our project page is at https://harshm121.github.io/projects/m3l.html

We introduce a novel weakly-supervised video actor-action segmentation (VAAS) framework, where only video-level tags are available. Previous VAAS methods follow a synthesize-and-refine scheme, i.e., they first synthesize the pseudo-segmentation and recursively refine the segmentation. However, this process requires significant time costs and heavily relies on the quality of the initial segmentation. Unlike existing works, our method hierarchically mines contrastive relations to supplement each other for learning a visually-plausible segmentation model. Specifically, three contrastive relations are abstracted from the pixel-level and frame-level, i.e., low-level edge-aware, class-activation map aware, and semantic tag-aware relations. Then, the discovered contrastive relations are unified into a universal objective for training the segmentation model, regardless of their heterogeneity. Moreover, we incorporate motion cues and unlabeled samples to increase the discriminative power and robustness of the segmentation model. Extensive experiments indicate that our proposed method produces reasonable segmentation.

Federated Learning (FL) is an emerging machine learning framework that enables multiple clients (coordinated by a server) to collaboratively train a global model by aggregating the locally trained models without sharing any client's training data. It has been observed in recent works that learning in a federated manner may lead the aggregated global model to converge to a 'sharp minimum' thereby adversely affecting the generalizability of this FL-trained model. Therefore, in this work, we aim to improve the generalization performance of models trained in a federated setup by introducing a 'flatness' constrained FL optimization problem. This flatness constraint is imposed on the top eigenvalue of the Hessian computed from the training loss. %of the global model. As each client trains a model on its local data, we further re-formulate this complex problem utilizing the client loss functions and propose a new computationally efficient regularization technique, dubbed 'MAN' which Minimizes Activation's Norm of each layer on client-side models. We also theoretically show that minimizing the activation norm reduces the top eigenvalue of the layer-wise Hessian of the client's loss, which in turn decreases the overall Hessian's top eigenvalue, ensuring convergence to a flat minimum. We apply our proposed flatness-constrained optimization to the existing FL techniques and obtain significant improvements, thereby establishing new state-of-the-art.

Vision Transformers (ViTs) have become one of the dominant architectures in computer vision, and pre-trained ViT models are commonly adapted to new tasks via fine-tuning. Recent works proposed several parameter-efficient transfer learning methods, such as adapters, to avoid the prohibitive training and storage cost of fine-tuning. In this work, we observe that adapters perform poorly when the dimension of adapters is small, and we propose MiMi, a training framework that addresses this issue. We start with large adapters which can reach high performance, and iteratively reduce the size of every adapter. We introduce a scoring function that compares neuron importance across layers and consequently allows automatic estimation of the hidden dimension of every adapter. Our method outperforms existing methods in finding the best trade-off between accuracy and trained parameters across the three dataset benchmarks DomainNet, VTAB, and Multi-task, for a total of 29 datasets. We will release our code publicly upon acceptance.

One of the daunting challenges in federated learning (FL) is the heterogeneity across clients that hinders the successful federation of a global model. When the heterogeneity becomes worse, personalized federated learning (PFL) pursues to detour the hardship of capturing the commonality across clients by allowing the personalization of models built upon the federation. In the scope of PFL for visual models, on the contrary, the recent effort for aggregating an effective global representation rather than chasing further personalization draws great attention. Along the same lines, we aim to train a large-margin global representation with a strong generalization across clients by adopting the meta-learning framework and margin-based loss, which are widely accepted to be effective in handling multiple visual tasks. Our method called MetVers achieves state-of-the-art accuracies for the PFL benchmarks with the CIFAR-10, CIFAR-100, and CINIC-10 datasets while showing robustness against data reconstruction attacks. Noteworthy, the versatile representation of MetaVers exhibits a strong generalization when tested on new clients with novel classes.

Beyond the Transformer, it is important to explore how to exploit the capacity of the MetaFormer, an architecture that is fundamental to the performance improvements of the Transformer. Previous studies have exploited it only for the backbone network. Unlike previous studies, we explore the capacity of the Metaformer architecture more extensively in the semantic segmentation task. We propose a powerful semantic segmentation network, MetaSeg, which leverages the Metaformer architecture from the backbone to the decoder. Our MetaSeg shows that the MetaFormer architecture plays a significant role in capturing the useful contexts for the decoder as well as for the backbone. In addition, recent segmentation methods have shown that using a CNN-based backbone for extracting the spatial information and a decoder for extracting the global information is more effective than using a transformer-based backbone with a CNN-based decoder. This motivates us to adopt the CNN-based backbone using the MetaFormer block and design our MetaFormer-based decoder, which consists of a novel self-attention module to capture the global contexts. To consider both the global contexts extraction and the computational efficiency of the self-attention for semantic segmentation, we propose a Channel Reduction Attention (CRA) module that reduces the channel dimension of the query and key into the one dimension. In this way, our proposed MetaSeg outperforms the previous state-of-the-art methods with more efficient computational costs on popular semantic segmentation and a medical image segmentation benchmark, including ADE20K, Cityscapes, COCO-stuff, and Synapse.

Recent image degradation estimation methods have enabled single-image super-resolution (SR) approaches to better upsample real-world images. Among these methods, explicit kernel estimation approaches have demonstrated unprecedented performance at handling unknown degradations. Nonetheless, a number of limitations constrain their efficacy when used by downstream SR models. Specifically, this family of methods yields i) excessive inference time due to long per-image adaptation times and ii)inferior image fidelity due to kernel mismatch. In this work, we introduce a learning-to-learn approach that meta-learns from the information contained in a distribution of images, thereby enabling significantly faster adaptation to new images with substantially improved performance in both kernel estimation and image fidelity. Specifically, we meta-train a kernel-generating GAN, named MetaKernelGAN, on a range of tasks, such that when a new image is presented, the generator starts from an informed kernel estimate and the discriminator starts with a strong capability to distinguish between patch distributions. Compared with state-of-the-art methods, our experiments show that MetaKernelGAN better estimates the magnitude and covariance of the kernel, leading to state-of-the-art blind SR results within a similar computational regime when combined with a non-blind SR model. Through supervised learning of an unsupervised learner, our method maintains the generalizability of the unsupervised learner, improves the optimization stability of kernel estimation, and hence image adaptation, and leads to a faster inference with a speedup between 14.24 to 102.1x over existing methods.

Zero-shot learning (ZSL) is a promising approach to generalizing a model to categories unseen during training by leveraging class attributes, but challenges remain. Recently, methods using generative models to combat bias towards classes seen during training have pushed state of the art, but these generative models can be slow or computationally expensive to train. Also, these generative models assume that the attribute vector of each unseen class is available a priori at training, which is not always practical. Additionally, while many previous ZSL methods assume a one-time adaptation to unseen classes, in reality, the world is always changing, necessitating a constant adjustment of deployed models. Models unprepared to handle a sequential stream of data are likely to experience catastrophic forgetting. We propose a Meta-learned Attribute self-Interaction Network (MAIN) for continual ZSL. By pairing attribute self-interaction trained using meta-learning with inverse regularization of the attribute encoder, we are able to outperform state-of-the-art results without leveraging the unseen class attributes while also being able to train our models substantially faster (>100x) than expensive generative-based approaches. We demonstrate this with experiments on five standard ZSL datasets (CUB, aPY, AWA1, AWA2, and SUN) in the generalized zero-shot learning and continual (fixed/dynamic) zero-shot learning settings. Extensive ablations and analyses demonstrate the efficacy of various components proposed.

Member inference (MI) attacks aim to determine if a specific data sample was used to train a machine learning model. Thus, MI is a major privacy threat to models trained on private sensitive data, such as medical records. In MI attacks one may consider the black-box settings, where the model's parameters and activations are hidden from the adversary, or the white-box case where they are available to the attacker. In this work, we focus on the latter and present a novel MI attack for it that employs influence functions, or more specifically the samples' self-influence scores, to perform MI prediction. The proposed method is evaluated on CIFAR-10, CIFAR-100, and Tiny ImageNet datasets using various architectures such as AlexNet, ResNet, and DenseNet. Our new attack method achieves new state-of-the-art (SOTA) results for MI even with limited adversarial knowledge, and is effective against MI defense methods such as data augmentation and differential privacy. Our code is available at https: //github.com/giladcohen/sif_mi_attack.

Recent works have shown that the computational efficiency of 3D medical image (e.g. CT and MRI) segmentation can be impressively improved by dynamic inference based on slice-wise complexity. As a pioneering work, a dynamic architecture network for medical volumetric segmentation (i.e. Med-DANet) has achieved a favorable accuracy and efficiency trade-off by dynamically selecting a suitable 2D candidate model from the pre-defined model bank for different slices. However, the issues of incomplete data analysis, high training costs, and the two-stage pipeline in Med-DANet require further improvement. To this end, this paper further explores a unified formulation of the dynamic inference framework from the perspective of both the data itself and the model structure. For each slice of the input volume, our proposed method dynamically selects an important foreground region for segmentation based on the policy generated by our Decision Network and Crop Position Network. Besides, we propose to insert a stage-wise quantization selector to the employed segmentation model (e.g. U-Net) for dynamic architecture adapting. Extensive experiments on BraTS 2019 and 2020 show that our method achieves comparable or better performance than previous state-of-the-art methods with much less model complexity. Compared with previous methods Med-DANet and TransBTS with dynamic and static architecture respectively, our framework improves the model efficiency by up to nearly 4.1 and 17.3 times with comparable segmentation results on BraTS 2019. Code will be available at https://github.com/Rubics-Xuan/Med-DANet.

Federated learning (FL) aims at keeping client data local to preserve privacy. Instead of gathering the data itself, the server only collects aggregated gradient updates from clients. Following the popularity of FL, there has been considerable amount of work, revealing the vulnerability of FL approaches by reconstructing the input data from gradient updates. Yet, most existing works assume an FL setting with unrealistically small batch size, and have poor image quality when the batch size is large. Other works modify the neural network architectures or parameters to the point of being suspicious, and thus, can be detected by clients. Moreover, most of them can only reconstruct one sample input from a large batch. To address these limitations, we propose a novel and completely analytical approach, referred to as the maximum knowledge orthogonality reconstruction (MKOR), to reconstruct clients' input data. Our proposed method reconstructs a mathematically proven high quality image from large batches. MKOR only requires the server to send secretly modified parameters to clients and can efficiently and inconspicuously reconstruct the input images from clients' gradient updates. We evaluate MKOR's performance on the MNIST, CIFAR-100, and ImageNet dataset and compare it with the state-of-the-art works. The results show that MKOR outperforms the existing approaches, and draws attention to a pressing need for further research on the privacy protection of FL so that comprehensive defense approaches can be developed.

While image data starts to enjoy the simple-but-effective self-supervised learning scheme built upon masking and self-reconstruction objective thanks to the introduction of tokenization procedure and vision transformer backbone, convolutional neural networks as another important and widely-adopted architecture for image data, though having contrastive-learning techniques to drive the self-supervised learning, still face the difficulty of leveraging such straightforward and general masking operation to benefit their learning process significantly. In this work, we aim to alleviate the burden of including masking operation into the contrastive-learning framework for convolutional neural networks as an extra augmentation method. In addition to the additive but unwanted edges (between masked and unmasked regions) as well as other adverse effects caused by the masking operations for ConvNets, which have been discussed by prior works, we particularly identify the potential problem where for one view in a contrastive sample-pair the randomly-sampled masking regions could be overly concentrated on important/salient objects thus resulting in misleading contrastiveness to the other view. To this end, we propose to explicitly take the saliency constraint into consideration in which the masked regions are more evenly distributed among the foreground and background for realizing the masking-based augmentation. Moreover, we introduce hard negative samples by masking larger regions of salient patches in an input image. Extensive experiments conducted on various datasets, contrastive learning mechanisms, and downstream tasks well verify the efficacy as well as the superior performance of our proposed method with respect to several state-of-the-art baselines.

Event cameras asynchronously capture brightness changes with low latency, high temporal resolution, and high dynamic range. However, annotation of event data is a costly and laborious process, which limits the use of deep learning methods for classification and other semantic tasks with the event modality. To reduce the dependency on labeled event data, we introduce Masked Event Modeling (MEM), a self-supervised framework for events. Our method pretrains a neural network on unlabeled events, which can originate from any event camera recording. Subsequently, the pretrained model is finetuned on a downstream task, leading to a consistent improvement of the task accuracy. For example, our method reaches state-of-the-art classification accuracy across three datasets, N-ImageNet, N-Cars, and N-Caltech101, increasing the top-1 accuracy of previous work by significant margins. When tested on real-world event data, MEM is even superior to supervised RGB-based pretraining. The models pretrained with MEM are also label-efficient and generalize well to the dense task of semantic image segmentation.

This study introduces an efficacious approach, Masked Collaborative Contrast (MCC), to highlight semantic regions in weakly supervised semantic segmentation. MCC adroitly draws inspiration from masked image modeling and contrastive learning to devise a novel framework that induces keys to contract toward semantic regions. Unlike prevalent techniques that directly eradicate patch regions in the input image when generating masks, we scrutinize the neighborhood relations of patch tokens by exploring masks considering keys on the affinity matrix. Moreover, we generate positive and negative samples in contrastive learning by utilizing the masked local output and contrasting it with the global output. Elaborate experiments on commonly employed datasets evidences that the proposed MCC mechanism effectively aligns global and local perspectives within the image, attaining impressive performance. The source code is available at https://github.com/fwu11/MCC.

In recent years, transformer-based models have dominated panoptic segmentation, thanks to their strong modeling capabilities and their unified representation for both semantic and instance classes as global binary masks. In this paper, we revisit pure convolution model and propose a novel panoptic architecture named MaskConver. MaskConver proposes to fully unify things and stuff representation by predicting their centers. To that extent, it creates a lightweight class embedding module that can break the ties when multiple centers co-exist in the same location. Furthermore, our study shows that the decoder design is critical in ensuring that the model has sufficient context for accurate detection and segmentation. We introduce a powerful ConvNeXt-UNet decoder that closes the performance gap between convolution- and transformer-based models. With ResNet50 backbone, our MaskConver achieves 53.6% PQ on the COCO panoptic val set, out-performing the modern convolution-based model, Panoptic FCN, by 9.3% as well as transformer-based models such as Mask2Former (+1.7% PQ) and kMaX-DeepLab (+0.6% PQ). Additionally, MaskConver with a MobileNet backbone reaches 37.2% PQ, improving over Panoptic-DeepLab by +6.4% under the same FLOPs/latency constraints. A further optimized version of MaskConver achieves 29.7% PQ, while running in real-time on mobile devices. The code and model weights will be publicly available.

Martian landslide segmentation is a challenging task compared to the same task on Earth. One of the reasons is that vegetation is typically lost or significantly less compared to its surroundings in the regions of landslide on Earth. In contrast, Mars is a desert planet, and there is no vegetation to aid landslide detection and segmentation. Recent work has demonstrated the strength of vision transformer (ViT) based deep learning models for various computer vision tasks. Inspired by the multi-head attention mechanism in ViT, which can model the global long-range spatial correlation between local regions in the input image, we hypothesize self-attention mechanism can effectively capture pertinent contextual information for the Martian landslide segmentation task. Furthermore, considering parameter efficiency or model size is another important factor for deep learning algorithms, we construct a new feature representation block, namely Progressively Expanded Neuron Attention (PEN-Attention), to extract more relevant features with significantly fewer trainable parameters. Overall, we refer to our deep learning architecture as the Martian landslide segmentation network (MarsLS-Net). In addition to the new architecture, we introduce a new multi-modal Martian landslide segmentation dataset for the first time, which will be made publicly available at https://github.com/MAIN-Lab/Multimodal-Martian-Landslides-Dataset