CVPR 2022 Papers

Patch attacks, one of the most threatening forms of physical attack in adversarial examples, can lead networks to induce misclassification by modifying pixels arbitrarily in a continuous region. Certifiable patch defense can guarantee robustness that the classifier is not affected by patch attacks. Existing certifiable patch defenses sacrifice the clean accuracy of classifiers and only obtain a low certified accuracy on toy datasets. Furthermore, the clean and certified accuracy of these methods is still significantly lower than the accuracy of normal classification networks, which limits their application in practice. To move towards a practical certifiable patch defense, we introduce Vision Transformer (ViT) into the framework of Derandomized Smoothing (DS). Specifically, we propose a progressive smoothed image modeling task to train Vision Transformer, which can capture the more discriminable local context of an image while preserving the global semantic information. For efficient inference and deployment in the real world, we innovatively reconstruct the global self-attention structure of the original ViT into isolated band unit self-attention. On ImageNet, under 2% area patch attacks our method achieves 41.70% certified accuracy, a nearly 1-fold increase over the previous best method (26.00%). Simultaneously, our method achieves 78.58% clean accuracy, which is quite close to the normal ResNet-101 accuracy. Extensive experiments show that our method obtains state-of-the-art clean and certified accuracy with inferring efficiently on CIFAR-10 and ImageNet.

Image-level label based weakly supervised semantic segmentation has attracted much attention since image labels are very easy to obtain. Existing methods usually generate pseudo labels from class activation map (CAM) and then train a segmentation model. CAM usually highlights partial objects and produce incomplete pseudo labels. Some methods explore object contour by training a contour model with CAM seed label supervision and then propagate CAM score from discriminative regions to non-discriminative regions with contour guidance. The propagation process suffers from the noisy intra-object contours, and inadequate propagation results produce incomplete pseudo labels. This is because the coarse CAM seed label lacks sufficient precise semantic information to suppress contour noise. In this paper, we train a SANCE model which utilizes an auxiliary segmentation module to supplement high-level semantic information for contour training by backbone feature sharing and online label supervision. The auxiliary segmentation module also provides more accurate localization map than CAM for pseudo label generation. We evaluate our approach on Pascal VOC 2012 and MS COCO 2014 benchmarks and achieve state-of-the-art performance, demonstrating the effectiveness of our method.

Light field applications, especially light field rendering and depth estimation, developed rapidly in recent years. While state-of-the-art light field rendering methods handle semi-transparent and reflective objects well, depth estimation methods either ignore these cases altogether or only deliver a weak performance. We argue that this is due current methods only considering a single "true" depth, even when multiple objects at different depths contributed to the color of a single pixel. Based on the simple idea of outputting a posterior depth distribution instead of only a single estimate, we develop and explore several different deep-learning-based approaches to the problem. Additionally, we contribute the first "multimodal light field depth dataset" that contains the depths of all objects which contribute to the color of a pixel. This allows us to supervise the multimodal depth prediction and also validate all methods by measuring the KL divergence of the predicted posteriors. With our thorough analysis and novel dataset, we aim to start a new line of depth estimation research that overcomes some of the long-standing limitations of this field.

Recent years have witnessed significant progress in the area of single image dehazing, thanks to the employment of deep neural networks and diverse datasets. Most of the existing methods perform well when the training and testing are conducted on a single dataset. However, they are not able to handle different types of hazy images using a dehazing model trained on a particular dataset. One possible remedy is to perform training on multiple datasets jointly. However, we observe that this training strategy tends to compromise the model performance on individual datasets. Motivated by this observation, we propose a test-time training method which leverages a helper network to assist the dehazing model in better adapting to a domain of interest. Specifically, during the test time, the helper network evaluates the quality of the dehazing results, then directs the dehazing network to improve the quality by adjusting its parameters via self-supervision. Nevertheless, the inclusion of the helper network does not automatically ensure the desired performance improvement. For this reason, a meta-learning approach is employed to make the objectives of the dehazing and helper networks consistent with each other. We demonstrate the effectiveness of the proposed method by providing extensive supporting experiments.

Malaria, a fatal but curable disease claims hundreds of thousands of lives every year. Early and correct diagnosis is vital to avoid health complexities, however, it depends upon the availability of costly microscopes and trained experts to analyze blood-smear slides. Deep learning-based methods have the potential to not only decrease the burden of experts but also improve diagnostic accuracy on low-cost microscopes. However, this is hampered by the absence of a reasonable size dataset. One of the most challenging aspects is the reluctance of the experts to annotate the dataset at low magnification on low-cost microscopes. We present a dataset to further the research on malaria microscopy over the low-cost microscopes at low magnification. Our large-scale dataset consists of images of blood-smear slides from several malaria-infected patients, collected through microscopes at two different cost spectrums and multiple magnifications. Malarial cells are annotated for the localization and life-stage classification task on the images collected through the high-cost microscope at high magnification. We design a mechanism to transfer these annotations from the high-cost microscope at high magnification to the low-cost microscope, at multiple magnifications. Multiple object detectors and domain adaptation methods are presented as the baselines. Furthermore, a partially supervised domain adaptation method is introduced to adapt the object-detector to work on the images collected from the low-cost microscope. The dataset and benchmark models will be made publicly available.

One of the major challenges in training text-to-image generation models is the need of a large number of high-quality text-image pairs. While image samples are often easily accessible, the associated text description typically requires careful human captioning, which is particularly time- and cost-consuming. In this paper, we propose the first work to train text-to-image generation models without any text data. It intelligently leverages the well-aligned cross-modal semantic space of the powerful pre-trained CLIP model: the requirement of text-conditioning is alleviated via generating text features from image features. Extensive experiments are conducted to illustrate the effectiveness of the proposed method. We obtain state-of-the-art results in the standard text-to-image generation tasks. Importantly, the proposed language-free model outperforms most existing models trained with full text-image pairs. Furthermore, our method can be applied in fine-tuning pre-trained models, which saves both training time and cost in training text-to-image generation models. Our pre-trained model obtains competitive results in zero-shot text-to-image generation on MS-COCO dataset, yet with around only 1% of the model size compared to the recently proposed large DALL-E model.

In this work, we explore the challenging task of generating 3D shapes from text. Beyond the existing works, we propose a new approach for text-guided 3D shape generation, capable of producing high-fidelity shapes with colors that match the given text description. This work has several technical contributions. First, we decouple the shape and color predictions for learning features in both texts and shapes, and propose the word-level spatial transformer to correlate word features from text with spatial features from shape. Also, we design a cyclic loss to encourage consistency between text and shape, and introduce the shape IMLE to diversify the generated shapes. Further, we extend the framework to enable text-guided shape manipulation. Extensive experiments on the largest existing text-shape benchmark manifest the superiority of this work. The code and the models are available at https://github.com/liuzhengzhe/Towards-Implicit Text-Guided-Shape-Generation.

Scene text detection and document layout analysis have long been treated as two separate tasks in different image domains. In this paper, we bring them together and introduce the task of unified scene text detection and layout analysis. The first hierarchical scene text dataset is introduced to enable this novel research task. We also propose a novel method that is able to simultaneously detect scene text and form text clusters in a unified way. Comprehensive experiments show that our unified model achieves better performance than multiple well-designed baseline methods. Additionally, this model achieves state-of-the-art results on multiple scene text detection datasets without the need of complex post-processing. Dataset and code: https://github.com/google-research-datasets/hiertext.

Recently, Sharpness-Aware Minimization (SAM), which connects the geometry of the loss landscape and generalization, has demonstrated a significant performance boost on training large-scale models such as vision transformers. However, the update rule of SAM requires two sequential (non-parallelizable) gradient computations at each step, which can double the computational overhead. In this paper, we propose a novel algorithm LookSAM - that only periodically calculates the inner gradient ascent, to significantly reduce the additional training cost of SAM. The empirical results illustrate that LookSAM achieves similar accuracy gains to SAM while being tremendously faster - it enjoys comparable computational complexity with first-order optimizers such as SGD or Adam. To further evaluate the performance and scalability of LookSAM, we incorporate a layer-wise modification and perform experiments in the large-batch training scenario, which is more prone to converge to sharp local minima. Equipped with the proposed algorithms, we are the first to successfully scale up the batch size when training Vision Transformers (ViTs). With a 64k batch size, we are able to train ViTs from scratch in minutes while maintaining competitive performance. The code is available here: https://github.com/yong-6/LookSAM

After the 2017 TuSimple Lane Detection Challenge, its dataset and evaluation based on accuracy and F1 score have become the de facto standard to measure the performance of lane detection methods. While they have played a major role in improving the performance of lane detection methods, the validity of this evaluation method in downstream tasks has not been adequately researched. In this study, we design 2 new driving-oriented metrics for lane detection: End-to-End Lateral Deviation metric (E2E-LD) is directly formulated based on the requirements of autonomous driving, a core task downstream of lane detection; Per-frame Simulated Lateral Deviation metric (PSLD) is a lightweight surrogate metric of E2E-LD. To evaluate the validity of the metrics, we conduct a large-scale empirical study with 4 major types of lane detection approaches on the TuSimple dataset and our newly constructed dataset Comma2k19-LD. Our results show that the conventional metrics have strongly negative correlations (<=-0.55) with E2E-LD, meaning that some recent improvements purely targeting the conventional metrics may not have led to meaningful improvements in autonomous driving, but rather may actually have made it worse by overfitting to the conventional metrics. On the contrary, PSLD shows statistically significant strong positive correlations (>=0.38) with E2E-LD. As a result, the conventional metrics tend to overestimate less robust models. As autonomous driving is a security/safety-critical system, the underestimation of robustness hinders the sound development of practical lane detection models. We hope that our study will help the community achieve more downstream task-aware evaluations for lane detection.

The ability to synthesize long-term human motion sequences in real-world scenes can facilitate numerous applications. Previous approaches for scene-aware motion synthesis are constrained by pre-defined target objects or positions and thus limit the diversity of human-scene interactions for synthesized motions. In this paper, we focus on the problem of synthesizing diverse scene-aware human motions under the guidance of target action sequences. To achieve this, we first decompose the diversity of scene aware human motions into three aspects, namely interaction diversity (e.g. sitting on different objects with different poses in the given scenes), path diversity (e.g. moving to the target locations following different paths), and the motion diversity (e.g. having various body movements during moving). Based on this factorized scheme, a hierarchical framework is proposed with each sub-module responsible for modeling one aspect. We assess the effectiveness of our framework on two challenging datasets for scene-aware human motion synthesis. The experiment results show that the proposed framework remarkably outperforms the previous methods in terms of diversity and naturalness.

Discriminative representation is crucial for the association step in multi-object tracking. Recent work mainly utilizes features in single or neighboring frames for constructing metric loss and empowering networks to extract representation of targets. Although this strategy is effective, it fails to fully exploit the information contained in a whole trajectory. To this end, we propose a strategy, namely multi-view trajectory contrastive learning, in which each trajectory is represented as a center vector. By maintaining all the vectors in a dynamically updated memory bank, a trajectory-level contrastive loss is devised to explore the inter-frame information in the whole trajectories. Besides, in this strategy, each target is represented as multiple adaptively selected keypoints rather than a pre-defined anchor or center. This design allows the network to generate richer representation from multiple views of the same target, which can better characterize occluded objects. Additionally, in the inference stage, a similarity-guided feature fusion strategy is developed for further boosting the quality of the trajectory representation. Extensive experiments have been conducted on MOTChallenge to verify the effectiveness of the proposed techniques. The experimental results indicate that our method has surpassed preceding trackers and established new state-of-the-art performance.

Semi-supervised learning (SSL) has been studied for a long time to solve vision tasks in data-efficient application scenarios. SSL aims to learn a good classification model using a few labeled data together with large-scale unlabeled data. Recent advances achieve the goal by combining multiple SSL techniques, e.g., self-training and consistency regularization. From unlabeled samples, they usually adopt a confidence filter (CF) to select reliable ones with high prediction confidence. In this work, we study whether the moderately confident samples are useless and how to select the useful ones to improve model optimization. To answer these problems, we propose a novel Taylor expansion inspired filtration (TEIF) framework, which admits the samples of moderate confidence with similar feature or gradient to the respective one averaged over the labeled and highly confident unlabeled data. It can produce a stable and new information induced network update, leading to better generalization. Two novel filters are derived from this framework and can be naturally explained in two perspectives. One is gradient synchronization filter (GSF), which strengthens the optimization dynamic of fully-supervised learning; it selects the samples whose gradients are similar to class-wise majority gradients. The other is prototype proximity filter (PPF), which involves more prototypical samples in training to learn better semantic representations; it selects the samples near class-wise prototypes. They can be integrated into SSL methods with CF. We use the state-of-the-art FixMatch as the baseline. Experiments on popular SSL benchmarks show that we achieve the new state of the art.

Machine learning models deployed as a service (MLaaS) are susceptible to model stealing attacks, where an adversary attempts to steal the model within a restricted access framework. While existing attacks demonstrate near-perfect clone-model performance using softmax predictions of the classification network, most of the APIs allow access to only the top-1 labels. In this work, we show that it is indeed possible to steal Machine Learning models by accessing only top-1 predictions (Hard Label setting) as well, without access to model gradients (Black-Box setting) or even the training dataset (Data-Free setting) within a low query budget. We propose a novel GAN-based framework that trains the student and generator in tandem to steal the model effectively while overcoming the challenge of the hard label setting by utilizing gradients of the clone network as a proxy to the victim's gradients. We propose to overcome the large query costs associated with a typical Data-Free setting by utilizing publicly available (potentially unrelated) datasets as a weak image prior. We additionally show that even in the absence of such data, it is possible to achieve state-of-the-art results within a low query budget using synthetically crafted samples. We are the first to demonstrate the scalability of Model Stealing in a restricted access setting on a 100 class dataset as well.

Deep learning based single image super-resolution models have been widely studied and superb results are achieved in upscaling low-resolution images with fixed scale factor and downscaling degradation kernel. To improve real world applicability of such models, there are growing interests to develop models optimized for arbitrary upscaling factors. Our proposed method is the first to treat arbitrary rescaling, both upscaling and downscaling, as one unified process. Using joint optimization of both directions, the proposed model is able to learn upscaling and downscaling simultaneously and achieve bidirectional arbitrary image rescaling. It improves the performance of current arbitrary upscaling models by a large margin while at the same time learns to maintain visual perception quality in downscaled images. The proposed model is further shown to be robust in cycle idempotence test, free of severe degradations in reconstruction accuracy when the downscaling-to-upscaling cycle is applied repetitively. This robustness is beneficial for image rescaling in the wild when this cycle could be applied to one image for multiple times. It also performs well on tests with arbitrary large scales and asymmetric scales, even when the model is not trained with such tasks. Extensive experiments are conducted to demonstrate the superior performance of our model.

Deep neural networks are very successful on many vision tasks, but hard to interpret due to their black box nature. To overcome this, various post-hoc attribution methods have been proposed to identify image regions most influential to the models' decisions. Evaluating such methods is challenging since no ground truth attributions exist. We thus propose three novel evaluation schemes to more reliably measure the faithfulness of those methods, to make comparisons between them more fair, and to make visual inspection more systematic. To address faithfulness, we propose a novel evaluation setting (DiFull) in which we carefully control which parts of the input can influence the output in order to distinguish possible from impossible attributions. To address fairness, we note that different methods are applied at different layers, which skews any comparison, and so evaluate all methods on the same layers (ML-Att) and discuss how this impacts their performance on quantitative metrics. For more systematic visualizations, we propose a scheme (AggAtt) to qualitatively evaluate the methods on complete datasets. We use these evaluation schemes to study strengths and shortcomings of some widely used attribution methods. Finally, we propose a post-processing smoothing step that significantly improves the performance of some attribution methods, and discuss its applicability.

Plasticity-stability dilemma is a main problem for incremental learning, where plasticity is referring to the ability to learn new knowledge, and stability retains the knowledge of previous tasks. Many methods tackle this problem by storing previous samples, while in some applications, training data from previous tasks cannot be legally stored. In this work, we propose to employ mode connectivity in loss landscapes to achieve better plasticity-stability trade-off without any previous samples. We give an analysis of why and how to connect two independently optimized optima of networks, null-space projection for previous tasks and simple SGD for the current task, can attain a meaningful balance between preserving already learned knowledge and granting sufficient flexibility for learning a new task. This analysis of mode connectivity also provides us a new perspective and technology to control the trade-off between plasticity and stability. We evaluate the proposed method on several benchmark datasets. The results indicate our simple method can achieve notable improvement, and perform well on both the past and current tasks. On 10-split-CIFAR-100 task, our method achieves 79.79% accuracy, which is 6.02% higher. Our method also achieves 6.33% higher accuracy on TinyImageNet. Code is available at https://github.com/lingl1024/Connector.

Accurate facial landmarks are essential prerequisites for many tasks related to human faces. In this paper, an accurate facial landmark detector is proposed based on cascaded transformers. We formulate facial landmark detection as a coordinate regression task such that the model can be trained end-to-end. With self-attention in transformers, our model can inherently exploit the structured relationships between landmarks, which would benefit landmark detection under challenging conditions such as large pose and occlusion. During cascaded refinement, our model is able to extract the most relevant image features around the target landmark for coordinate prediction, based on deformable attention mechanism, thus bringing more accurate alignment. In addition, we propose a novel decoder that refines image features and landmark positions simultaneously. With few parameter increasing, the detection performance improves further. Our model achieves new state-of- the-art performance on several standard facial landmark detection benchmarks, and shows good generalization ability in cross-dataset evaluation.