ICCV 2021 Papers

Deepfakes ("deep learning" + "fake") are synthetically-generated videos from AI algorithms. While they could be entertaining, they could also be misused for falsifying speeches and spreading misinformation. The process to create deepfakes involves both visual and auditory manipulations. Exploration on detecting visual deepfakes has produced a number of detection methods as well as datasets, while audio deepfakes (e.g. synthetic speech from text-to-speech or voice conversion systems) and the relationship between the visual and auditory modalities have been relatively neglected. In this work, we propose a novel visual / auditory deepfake joint detection task and show that exploiting the intrinsic synchronization between the visual and auditory modalities could benefit deepfake detection. Experiments demonstrate that the proposed joint detection framework outperforms independently trained models, and at the same time, yields superior generalization capability on unseen types of deepfakes.

Joint Energy-based Model (JEM) is a recently proposed hybrid model that retains strong discriminative power of modern CNN classifiers, while generating samples rivaling the quality of GAN-based approaches. In this paper, we propose a variety of new training procedures and architecture features to improve JEM's accuracy, training stability, and speed altogether. 1) We propose a proximal SGLD to generate samples in the proximity of samples from previous step, which improves the stability. 2) We further treat the approximate maximum likelihood learning of EBM as a multi-step differential game, and extend the YOPO framework to cut out redundant calculations during backpropagation, which accelerates the training substantially. 3) Rather than initializing SGLD chain from random noise, we introduce a new informative initialization that samples from a distribution estimated from training data. 4) This informative initialization allows us to enable batch normalization in JEM, which further releases the power of modern CNN architectures for hybrid modeling.

Few-shot learning amounts to learning representations and acquiring knowledge such that novel tasks may be solved with both supervision and data being limited. Improved performance is possible by transductive inference, where the entire test set is available concurrently, and semi-supervised learning, where more unlabeled data is available. Focusing on these two settings, we introduce a new algorithm that leverages the manifold structure of the labeled and unlabeled data distribution to predict pseudo-labels, while balancing over classes and using the loss value distribution of a limited-capacity classifier to select the cleanest labels, iteratively improving the quality of pseudo-labels. Our solution surpasses or matches the state of the art results on four benchmark datasets, namely miniImageNet, tieredImageNet, CUB and CIFAR-FS, while being robust over feature space pre-processing and the quantity of available data. The publicly available source code can be found in https://github.com/MichalisLazarou/iLPC

Recent progress in 3D object detection from single images leverages monocular depth estimation as a way to produce 3D pointclouds, turning cameras into pseudo-lidar sensors. These two-stage detectors improve with the accuracy of the intermediate depth estimation network, which can itself be improved without manual labels via large-scale self-supervised learning. However, they tend to suffer from overfitting more than end-to-end methods, are more complex, and the gap with similar lidar-based detectors remains significant. In this work, we propose an end-to-end, single stage, monocular 3D object detector, DD3D, that can benefit from depth pre-training like pseudo-lidar methods, but without their limitations. Our architecture is designed for effective information transfer between depth estimation and 3D detection, allowing us to scale with the amount of unlabeled pre-training data. Our method achieves state-of-theart results on two challenging benchmarks, with 16.34% and 9.28% AP for Cars and Pedestrians (respectively) on the KITTI-3D benchmark, and 41.5% mAP on NuScenes.

Recently, backdoor attacks pose a new security threat to the training process of deep neural networks (DNNs). Attackers intend to inject hidden backdoors into DNNs, such that the attacked model performs well on benign samples, whereas its prediction will be maliciously changed if hidden backdoors are activated by the attacker-defined trigger. Existing backdoor attacks usually adopt the setting that triggers are sample-agnostic, i.e., different poisoned samples contain the same trigger, resulting in that the attacks could be easily mitigated by current backdoor defenses. In this work, we explore a novel attack paradigm, where backdoor triggers are sample-specific. In our attack, we only need to modify certain training samples with invisible perturbation, while not need to manipulate other training components (e.g., training loss, and model structure) as required in many existing attacks. Specifically, inspired by the recent advance in DNN-based image steganography, we generate sample-specific invisible additive noises as backdoor triggers by encoding an attacker-specified string into benign images through an encoder-decoder network. The mapping from the string to the target label will be generated when DNNs are trained on the poisoned dataset. Extensive experiments on benchmark datasets verify the effectiveness of our method in attacking models with or without defenses.

Rolling shutter (RS) images can be viewed as the result of the row-wise combination of global shutter (GS) images captured by a virtual moving GS camera over the period of camera readout time. The RS effect brings tremendous difficulties for the downstream applications. In this paper, we propose to invert the above RS imaging mechanism, i.e., recovering a high framerate GS video from consecutive RS images to achieve RS temporal super-resolution (RSSR). This extremely challenging problem, e.g., recovering 1440 GS images from two 720-height RS images, is far from being solved end-to-end. To address this challenge, we exploit the geometric constraint in the RS camera model, thus achieving geometry-aware inversion. Specifically, we make three contributions in resolving the above difficulties: (i) formulating the bidirectional RS undistortion flows under the constant velocity motion model, (ii) building the connection between the RS undistortion flow and optical flow via a scaling operation, and (iii) developing a mutual conversion scheme between varying RS undistortion flows that correspond to different scanlines. Building upon these formulations, we propose the first RS temporal super-resolution network in a cascaded structure to extract high framerate global shutter video. Our method explores the underlying spatio-temporal geometric relationships within a deep learning framework, where no extra supervision besides the middle-scanline ground truth GS image is needed. Essentially, our method can be very efficient for explicit propagation to generate GS images under any scanline. Experimental results on both synthetic and real data show that our method can produce high-quality GS image sequences with rich details, outperforming state-of-the-art methods.

With the strength of deep generative models, 3D pose transfer regains intensive research interests in recent years. Existing methods mainly rely on a variety of constraints to achieve the pose transfer over 3D meshes, e.g., the need for manually encoding for shape and pose disentanglement. In this paper, we present an unsupervised approach to conduct the pose transfer between any arbitrate given 3D meshes. Specifically, a novel Intrinsic-Extrinsic Preserved Generative Adversarial Network (IEP-GAN) is presented for both intrinsic (i.e., shape) and extrinsic (i.e., pose) information preservation. Extrinsically, we propose a co-occurrence discriminator to capture the structural/pose invariance from distinct Laplacians of the mesh. Meanwhile, intrinsically, a local intrinsic-preserved loss is introduced to preserve the geodesic priors while avoiding heavy computations. At last, we show the possibility of using IEP-GAN to manipulate 3D human meshes in various ways, including pose transfer, identity swapping and pose interpolation with latent code vector arithmetic. The extensive experiments on various 3D datasets of humans, animals and hands qualitatively and quantitatively demonstrate the generality of our approach. Our proposed model produces better results and is substantially more efficient compared to recent state-of-the-art methods. Code is available: https://github.com/mikecheninoulu/Unsupervised_IEPGAN

We propose IntraTomo, a powerful framework that combines the benefits of learning-based and model-based approaches for solving highly ill-posed inverse problems in the Computed Tomography (CT) context. IntraTomo is composed of two core modules: a novel sinogram prediction module, and a geometry refinement module, which are applied iteratively. In the first module, the unknown density field is represented as a continuous and differentiable function, parameterized by a deep neural network. This network is learned, in a self-supervised fashion, from the incomplete or/and degraded input sinogram. After getting estimated through the sinogram prediction module, the density field is consistently refined in the second module using local and non-local geometrical priors. With these two core modules, we show that IntraTomo significantly outperforms existing approaches on several ill-posed inverse problems, such as limited angle tomography with a range of 45 degrees, sparse view tomographic reconstruction with as few as eight views, or super-resolution tomography with eight times increased resolution. The experiments on simulated and real data show that our approach can achieve results of unprecedented quality.

This paper aims to explain adversarial attacks in terms of how adversarial perturbations contribute to the attacking task. We estimate attributions of different image regions to the decrease of the attacking cost based on the Shapley value. We define and quantify interactions among adversarial perturbation pixels, and decompose the entire perturbation map into relatively independent perturbation components. The decomposition of the perturbation map shows that adversarially-trained DNNs have more perturbation components in the foreground than normally-trained DNNs. Moreover, compared to the normally-trained DNN, the adversarially-trained DNN have more components which mainly decrease the score of the true category. Above analyses provide new insights into the understanding of adversarial attacks.

Communication between embodied AI agents has received increasing attention in recent years. Despite its use, it is still unclear whether the learned communication is interpretable and grounded in perception. To study the grounding of emergent forms of communication, we first introduce the collaborative multi-object navigation task 'CoMON.' In this task, an 'oracle agent' has detailed environment information in the form of a map. It communicates with a 'navigator agent' that perceives the environment visually and is tasked to find a sequence of goals. To succeed at the task, effective communication is essential. CoMON hence serves as a basis to study different communication mechanisms between heterogeneous agents, that is, agents with different capabilities and roles. We study two common communication mechanisms and analyze their communication patterns through an egocentric and spatial lens. We show that the emergent communication can be grounded to the agent observations and the spatial structure of the 3D environment.

We study the problem of concept induction in visual reasoning, i.e., identifying concepts and their hierarchical relationships from question-answer pairs associated with images; and achieve an interpretable model via working on the induced symbolic concept space. To this end, we first design a new framework named object-centric compositional attention model (OCCAM) to perform the visual reasoning task with object-level visual features. Then, we come up with a method to induce concepts of objects and relations using clues from the attention patterns between objects' visual features and question words. Finally, we achieve a higher level of interpretability by imposing OCCAM on the objects represented in the induced symbolic concept space. Experiments on the CLEVR and GQA datasets demonstrate: 1) our OCCAM achieves a new state of the art without human-annotated functional programs; 2) our induced concepts are both accurate and sufficient as OCCAM achieves an on-par performance on objects represented either in visual features or in the induced symbolic concept space.

Humans usually explain their reasoning (e.g. classification) by dissecting the image and pointing out the evidence from these parts to the concepts in their minds. Inspired by this cognitive process, several part-level interpretable neural network architectures have been proposed to explain the predictions. However, they suffer from the complex data structure and confusing the effect of the individual part to output category. In this work, an interpretable image recognition deep network is designed by introducing a plug-in transparent embedding space (TesNet) to bridge the high-level input patches (e.g. CNN feature maps) and the output categories. This plug-in embedding space is spanned by transparent basis concepts which are constructed on the Grassmann manifold. These basis concepts are enforced to be category-aware and within-category concepts are orthogonal to each other, which makes sure the embedding space is disentangled. Meanwhile, each basis concept can be traced back to the particular image patches, thus they are transparent and friendly to explain the reasoning process. By comparing with state-of-the-art interpretable methods, TesNet is much more beneficial to classification tasks, esp. providing better interpretability on predictions and improve the final accuracy.

Sparse voxel-based 3D convolutional neural networks (CNNs) are widely used for various 3D vision tasks. Sparse voxel-based 3D CNNs create sparse non-empty voxels from input point clouds and perform standard convolution operations on them only. We propose a simple and effective padding scheme --- interpolation-aware padding to pad a few empty voxels adjacent to the non-empty voxels and involving them in the CNN computation so that all neighboring voxels exist when computing point-wise features via the trilinear interpolation. For fine-grained 3D vision tasks where point-wise features are essential, like semantic segmentation and 3D detection, our network achieves higher prediction accuracy than the existing networks using the nearest neighbor interpolation or normalized trilinear interpolation with the zero-padding or the octree-padding scheme. Through extensive comparisons on various 3D segmentation and detection tasks, we demonstrate the superiority of 3D sparse CNNs with our sparse padding scheme in conjunction with feature interpolation.

We propose a novel framework for video inpainting by adopting an internal learning strategy. Unlike previous methods that use optical flow for cross-frame context propagation to inpaint unknown regions, we show that this can be achieved implicitly by fitting a convolutional neural network to known regions. Moreover, to handle challenging sequences with ambiguous backgrounds or long-term occlusion, we design two regularization terms to preserve high-frequency details and long-term temporal consistency. Extensive experiments on the DAVIS dataset demonstrate that the proposed method achieves state-of-the-art inpainting quality quantitatively and qualitatively. We further extend the proposed method to another challenging task: learning to remove an object from a video giving a single object mask in only one frame in a 4K video.

Egocentric video recognition is a challenging task that requires to identify both the actor's motion and the active object that the actor interacts with. Recognizing the active object is particularly hard due to the cluttered background with distracting objects, the frequent field of view changes, severe occlusion, etc. To improve the active object classification, most existing methods use object detectors or human gaze information, which are computationally expensive or require labor-intensive annotations. To avoid these additional costs, we propose an end-to-end Interactive Prototype Learning (IPL) framework to learn better active object representations by leveraging the motion cues from the actor. First, we introduce a set of verb prototypes to disentangle active object features from distracting object features. Each prototype corresponds to a primary motion pattern of an egocentric action, offering a distinctive supervision signal for active object feature learning. Second, we design two interactive operations to enable the extraction of active object features, i.e., noun-to-verb assignment and verb-to-noun selection. These operations are parameter-efficient and can learn judicious location-aware features on top of 3D CNN backbones. We demonstrate that the IPL framework can generalize to different backbones and outperform the state-of-the-art on three large-scale egocentric video datasets, i.e., EPIC-KITCHENS-55, EPIC-KITCHENS-100 and EGTEA.

In this work, we devote to address the challenging problem of scene parsing. Previous methods, though capture context to exploit global clues, handle scene parsing as a pixel-independent task. However, it is well known that pixels in an image are highly correlated with each other, especially those from the same semantic region, while treating pixels independently fails to take advantage of such correlations. In this work, we treat each respective region in an image as a whole, and capture the structure topology as well as the affinity among different regions. To this end, we first divide the entire feature maps to different regions and extract respective global features from them. Next, we construct a directed graph whose nodes are regional features, and the edge connecting every two nodes is the affinity between the regional features they represent. After that, we transfer the affinity-aware nodes in the directed graph back to corresponding regions of the image, which helps to model the region dependencies and mitigate unrealistic results. In addition, to further boost the correlation among pixels, we propose a region-level loss that evaluates all pixels in a region as a whole and motivates the network to learn the exclusive regional feature per class. With the proposed approach, we achieves new state-of-the-art segmentation results on PASCAL-Context, ADE20K, and COCO-Stuff consistently.

We study the problem of multi-label zero-shot recognition in which labels are in the form of human-object interactions (combinations of actions on objects), each image may contain multiple interactions and some interactions do not have training images. We propose a novel compositional learning framework that decouples interaction labels into separate action and object scores that incorporate the spatial compatibility between the two components. We combine these scores to efficiently recognize seen and unseen interactions. However, learning action-object spatial relations, in principle, requires bounding-box annotations, which are costly to gather. Moreover, it is not clear how to generalize spatial relations to unseen interactions. We address these challenges by developing a cross-attention mechanism that localizes objects from action locations and vice versa by predicting displacements between them, referred to as relational directions. During training, we estimate the relational directions as ones maximizing the scores of ground-truth interactions that guide predictions toward compatible action-object regions. By extensive experiments, we show the effectiveness of our framework, where we improve the state of the art by 2.6% mAP score and 5.8% recall score on HICO and Visual Genome datasets, respectively. Code is available at https://github.com/hbdat/iccv21_relational_direction.

In this paper, we propose a novel deep learning framework to reconstruct 3D hand poses and shapes of two interacting hands from a single color image. Previous methods designed for single hand cannot be easily applied for the two hand scenario because of the heavy inter-hand occlusion and larger solution space. In order to address the occlusion and similar appearance between hands that may confuse the network, we design a hand pose-aware attention module to extract features associated to each individual hand respectively. We then leverage the two hand context presented in interaction and propose a context-aware cascaded refinement that improves the hand pose and shape accuracy of each hand conditioned on the context between interacting hands. Extensive experiments on the main benchmark datasets demonstrate that our method predicts accurate 3D hand pose and shape from single color image, and achieves the state-of-the-art performance. Code is available in project webpage https://baowenz.github.io/Intershape/.

Adversarial data examples have drawn significant attention from the machine learning and security communities. A line of work on tackling adversarial examples is certified robustness via randomized smoothing that can provide a theoretical robustness guarantee. However, such a mechanism usually uses floating-point arithmetic for calculations in inference and requires large memory footprints and daunting computational costs. These defensive models cannot run efficiently on edge devices nor be deployed on integer-only logical units such as Turing Tensor Cores or integer-only ARM processors. To overcome these challenges, we propose an integer randomized smoothing approach with quantization to convert any classifier into a new smoothed classifier, which uses integer-only arithmetic for certified robustness against adversarial perturbations. We prove a tight robustness guarantee under L2-norm for the proposed approach. We show our approach can obtain a comparable accuracy and 4x 5x speedup over floating-point arithmetic certified robust methods on general-purpose CPUs and mobile devices on two distinct datasets (CIFAR-10 and Caltech-101).