ICCV 2021 Papers

Nowadays advanced image editing tools and technical skills produce tampered images more realistically, which can easily evade image forensic systems and make authenticity verification of images more difficult. To tackle this challenging problem, we introduce TransForensics, a novel image forgery localization method inspired by Transformers. The two major components in our framework are dense self-attention encoders and dense correction modules. The former is to model global context and all pairwise interactions between local patches at different scales, while the latter is used for improving the transparency of the hidden layers and correcting the outputs from different branches. Compared to previous traditional and deep learning methods, TransForensics not only can capture discriminative representations and obtain high-quality mask predictions but is also not limited by tampering types and patch sequence orders. By conducting experiments on main benchmarks, we show that TransForensics outperforms the state-of-the-art methods by a large margin.

Facial expression recognition (FER) has received increasing interest in computer vision. We propose the TransFER model which can learn rich relation-aware local representations. It mainly consists of three components: Multi-Attention Dropping (MAD), ViT-FER, and Multi-head Self-Attention Dropping (MSAD). First, local patches play an important role in distinguishing various expressions, however, few existing works can locate discriminative and diverse local patches. This can cause serious problems when some patches are invisible due to pose variations or viewpoint changes. To address this issue, the MAD is proposed to randomly drop an attention map. Consequently, models are pushed to explore diverse local patches adaptively. Second, to build rich relations between different local patches, the Vision Transformers (ViT) are used in FER, called ViT-FER. Since the global scope is used to reinforce each local patch, a better representation is obtained to boost the FER performance. Thirdly, the multi-head self-attention allows ViT to jointly attend to features from different information subspaces at different positions. Given no explicit guidance, however, multiple self-attentions may extract similar relations. To address this, the MSAD is proposed to randomly drop one self-attention module. As a result, models are forced to learn rich relations among diverse local patches. Our proposed TransFER model outperforms the state-of-the-art methods on several FER benchmarks, showing its effectiveness and usefulness.

Event-based video frame interpolation is promising as event cameras capture dense motion signals that can greatly facilitate motion-aware synthesis. However, training existing frameworks for this task requires high frame-rate videos with synchronized events, posing challenges to collect real training data. In this work we show event-based frame interpolation can be trained without the need of high framerate videos. This is achieved via a novel weakly supervised framework that 1) corrects image appearance by extracting complementary information from events and 2) supplants motion dynamics modeling with attention mechanisms. For the latter we propose subpixel attention learning, which supports searching high-resolution correspondence efficiently on low-resolution feature grid. Though trained on low frame-rate videos, our framework outperforms existing models trained with full high frame-rate videos (and events) on both GoPro dataset and a new real event-based dataset. Codes, models and dataset will be made available at: https://github.com/YU-Zhiyang/WEVI.

In multi-object detection using neural networks, the fundamental problem is, "How should the network learn a variable number of bounding boxes in different input images?". Previous methods train a multi-object detection network through a procedure that directly assigns the ground truth bounding boxes to the specific locations of the network's output. However, this procedure makes the training of a multi-object detection network too heuristic and complicated. In this paper, we reformulate the multi-object detection task as a problem of density estimation of bounding boxes. Instead of assigning each ground truth to specific locations of network's output, we train a network by estimating the probability density of bounding boxes in an input image using a mixture model. For this purpose, we propose a novel network for object detection called Mixture Density Object Detector (MDOD), and the corresponding objective function for the density-estimation-based training. We applied MDOD to MS COCO dataset. Our proposed method not only deals with multi-object detection problems in a new approach, but also improves detection performances through MDOD. The code is available: https://github.com/yoojy31/MDOD.

Vehicle tracking is an essential task in the multi-object tracking (MOT) field. A distinct characteristic in vehicle tracking is that the trajectories of vehicles are fairly smooth in both the world coordinate and the image coordinate. Hence, models that capture motion consistencies are of high necessity. However, tracking with the standalone motion-based trackers is quite challenging because targets could get lost easily due to limited information, detection error and occlusion. Leveraging appearance information to assist object re-identification could resolve this challenge to some extent. However, doing so requires extra computation while appearance information is sensitive to occlusion as well. In this paper, we try to explore the significance of motion patterns for vehicle tracking without appearance information. We propose a novel approach that tackles the association issue for long-term tracking with the exclusive fully-exploited motion information. We address the tracklet embedding issue with the proposed reconstruct-to-embed strategy based on deep graph convolutional neural networks (GCN). Comprehensive experiments on the KITTI-car tracking dataset and UA-Detrac dataset show that the proposed method, though without appearance information, could achieve competitive performance with the state-of-the-art (SOTA) trackers. The source code will be available at https://github.com/GaoangW/LGMTracker.

The abundance and richness of Internet photos of landmarks and cities has led to significant progress in 3D vision over the past two decades, including automated 3D reconstructions of the world's landmarks from tourist photos. However, a major source of information available for these 3D-augmented collections---language, e.g., from image captions---has been virtually untapped. In this work, we present WikiScenes, a new, large-scale dataset of landmark photo collections that contains descriptive text in the form of captions and hierarchical category names. WikiScenes forms a new testbed for multimodal reasoning involving images, text, and 3D geometry. We demonstrate the utility of WikiScenes for learning semantic concepts over images and 3D models. Our weakly-supervised framework connects images, 3D structure and semantics---utilizing the strong constraints provided by 3D geometry---to associate semantic concepts to image pixels and points in 3D space.

Visual text recognition is undoubtedly one of the most extensively researched topics in computer vision. Great progress have been made to date, with the latest models starting to focus on the more practical "in-the-wild" setting. However, a salient problem still hinders practical deployment -- prior arts mostly struggle with recognising unseen (or rarely seen) character sequences. In this paper, we put forward a novel framework to specifically tackle this "unseen" problem. Our framework is iterative in nature, in that it utilises predicted knowledge of character sequences from a previous iteration, to augment the main network in improving the next prediction. Key to our success is a unique cross-modal variational autoencoder to act as a feedback module, which is trained with the presence of textual error distribution data. This module importantly translate a discrete predicted character space, to a continuous affine transformation parameter space used to condition the visual feature map at next iteration. Experiments on common datasets have shown competitive performance over state-of-the-arts under the conventional setting. Most importantly, under the new disjoint setup where train-test labels are mutually exclusive, ours offers the best performance thus showcasing the capability of generalising onto unseen words.

This paper proposes to handle the practical problem of learning a universal model for crowd counting across scenes and datasets. We dissect that the crux of this problem is the catastrophic sensitivity of crowd counters to scale shift, which is very common in the real world and caused by factors such as different scene layouts and image resolutions. Therefore it is difficult to train a universal model that can be applied to various scenes. To address this problem, we propose scale alignment as a prime module for establishing a novel crowd counting framework. We derive a closed-form solution to get the optimal image rescaling factors for alignment by minimizing the distances between their scale distributions. A novel neural network together with a loss function based on an efficient sliced Wasserstein distance is also proposed for scale distribution estimation. Benefiting from the proposed method, we have learned a universal model that generally works well on several datasets where can even outperform state-of-the-art models that are particularly fine-tuned for each dataset significantly. Experiments also demonstrate the much better generalizability of our model to unseen scenes.

Colorization has attracted increasing interest in recent years. Classic reference-based methods usually rely on external color images for plausible results. A large image database or online search engine is inevitably required for retrieving such exemplars. Recent deep-learning-based methods could automatically colorize images at a low cost. However, unsatisfactory artifacts and incoherent colors are always accompanied. In this work, we aim at recovering vivid colors by leveraging the rich and diverse color priors encapsulated in a pretrained Generative Adversarial Networks (GAN). Specifically, we first "retrieve" matched features (similar to exemplars) via a GAN encoder and then incorporate these features into the colorization process with feature modulations. Thanks to the powerful generative color prior and delicate designs, our method could produce vivid colors with a single forward pass. Moreover, it is highly convenient to obtain diverse results by modifying GAN latent codes. Our method also inherits the merit of interpretable controls of GANs and could attain controllable and smooth transitions by walking through GAN latent space. Extensive experiments and user studies demonstrate that our method achieves superior performance than previous works.

Recently, some works found an interesting phenomenon that adversarially robust classifiers can generate good images comparable to generative models. We investigate this phenomenon from an energy perspective and provide a novel explanation. We reformulate adversarial example generation, adversarial training, and image generation in terms of an energy function. We find that adversarial training contributes to obtaining an energy function that is flat and has low energy around the real data, which is the key for generative capability. Based on our new understanding, we further propose a better adversarial training method, Joint Energy Adversarial Training (JEAT), which can generate high-quality images and achieve new state-of-the-art robustness under a wide range of attacks. The Inception Score of the images (CIFAR-10) generated by JEAT is 8.80, much better than original robust classifiers (7.50). In particular, we achieve new state-of-the-art robustness on CIFAR-10 (from 57.20% to 62.04%) and CIFAR-100 (from 30.03% to 30.18%) without extra training data.

Rotation augmentations generally improve a model's invariance/equivariance to rotation - except in object detection. In object detection the shape is not known, therefore rotation creates a label ambiguity. We show that the de-facto method for bounding box label rotation, the Largest Box Method, creates very large labels, leading to poor performance and in many cases worse performance than using no rotation at all. We propose a new method of rotation augmentation that can be implemented in a few lines of code. First, we create a differentiable approximation of label accuracy and show that axis-aligning the bounding box around an ellipse is optimal. We then introduce Rotation Uncertainty (RU) Loss, allowing the model to adapt to the uncertainty of the labels. On five different datasets (including COCO, PascalVOC, and Transparent Object Bin Picking), this approach improves the rotational invariance of both one-stage and two-stage architectures when measured with AP, AP50, and AP75.

Recent studies have demonstrated the vulnerability of deep neural networks against adversarial examples. Inspired by the observation that adversarial examples often lie outside the natural image data manifold and the intrinsic dimension of image data is much smaller than its pixel space dimension, we propose to embed high-dimensional input images into a low-dimensional space and apply regularization on the embedding space to push the adversarial examples back to the manifold. The proposed framework is called Embedding Regularized Classifier (ER-Classifier), which improves the adversarial robustness of the classifier through embedding regularization. Besides improving classification accuracy against adversarial examples, the framework can be combined with detection methods to detect adversarial examples. Experimental results on several benchmark datasets show that, our proposed framework achieves good performance against strong adversarial attack methods.

Prohibited items detection in X-ray images often plays an important role in protecting public safety, which often deals with color-monotonous and luster-insufficient objects, resulting in unsatisfactory performance. Till now, there have been rare studies touching this topic due to the lack of specialized high-quality datasets. In this work, we first present a High-quality X-ray (HiXray) security inspection image dataset, which contains 102,928 common prohibited items of 8 categories. It is the largest dataset of high quality for prohibited items detection, gathered from the real-world airport security inspection and annotated by professional security inspectors. Besides, for accurate prohibited item detection, we further propose the Lateral Inhibition Module (LIM) inspired by the fact that humans recognize these items by ignoring irrelevant information and focusing on identifiable characteristics, especially when objects are overlapped with each other. Specifically, LIM, the elaborately designed flexible additional module, suppresses the noisy information flowing maximumly by the Bidirectional Propagation (BP) module and activates the most identifiable charismatic, boundary, from four directions by Boundary Activation (BA) module. We evaluate our method extensively on HiXray and OPIXray and the results demonstrate that it outperforms SOTA detection methods.

Automatic security inspection using computer vision technology is a challenging task in real-world scenarios due to various factors, including intra-class variance, class imbalance, and occlusion. Most of the previous methods rarely solve the cases that the prohibited items are deliberately hidden in messy objects due to the lack of large-scale datasets, restricted their applications in real-world scenarios. Towards real-world prohibited item detection, we collect a large-scale dataset, named as PIDray, which covers various cases in real-world scenarios for prohibited item detection, especially for deliberately hidden items. With an intensive amount of effort, our dataset contains 12 categories of prohibited items in 47,677 X-ray images with high-quality annotated segmentation masks and bounding boxes. To the best of our knowledge, it is the largest prohibited items detection dataset to date. Meanwhile, we design the selective dense attention network (SDANet) to construct a strong baseline, which consists of the dense attention module and the dependency refinement module. The dense attention module formed by the spatial and channel-wise dense attentions, is designed to learn the discriminative features to boost the performance. The dependency refinement module is used to exploit the dependencies of multi-scale features. Extensive experiments conducted on the collected PIDray dataset demonstrate that the proposed method performs favorably against the state-of-the-art methods, especially for detecting the deliberately hidden items.

Open-set domain adaptation (OSDA) considers that the target domain contains samples from novel categories unobserved in external source domain. Unfortunately, existing OSDA methods always ignore the demand for the information of unseen categories and simply recognize them as "unknown" set without further explanation. This motivates us to understand the unknown categories more specifically by exploring the underlying structures and recovering their interpretable semantic attributes. In this paper, we propose a novel framework to accurately identify the seen categories in target domain, and effectively recover the semantic attributes for unseen categories. Specifically, structure preserving partial alignment is developed to recognize the seen categories through domain-invariant feature learning. Attribute propagation over visual graph is designed to smoothly transit attributes from seen to unseen categories via visual-semantic mapping. Moreover, two new cross-main benchmarks are constructed to evaluate the proposed framework in the novel and practical challenge. Experimental results on open-set recognition and semantic recovery demonstrate the superiority of the proposed method over other compared baselines.

Quantization is a widely used technique to compress and accelerate deep neural networks. However, conventional quantization methods use the same bit-width for all (or most of) the layers, which often suffer significant accuracy degradation in the ultra-low precision regime and ignore the fact that emergent hardware accelerators begin to support mixed-precision computation. Consequently, we present a novel and principled framework to solve the mixed-precision quantization problem in this paper. Briefly speaking, we first formulate the mixed-precision quantization as a discrete constrained optimization problem. Then, to make the optimization tractable, we approximate the objective function with second-order Taylor expansion and propose an efficient approach to compute its Hessian matrix. Finally, based on the above simplification, we show that the original problem can be reformulated as a Multiple Choice Knapsack Problem (MCKP) and propose a greedy search algorithm to solve it efficiently. Compared with existing mixed-precision quantization works, our method is derived in a principled way and much more computationally efficient. Moreover, extensive experiments conducted on the ImageNet dataset and various kinds of network architectures also demonstrate its superiority over existing uniform and mixed-precision quantization approaches.

Deep neural networks (DNN) have shown superior performance in a variety of tasks. As they rapidly evolve, their escalating computation and memory demands make it challenging to deploy them on resource-constrained edge devices. Though extensive efficient accelerator designs, from traditional electronics to emerging photonics, have been successfully demonstrated, they are still bottlenecked by expensive memory accesses due to tremendous gaps between the bandwidth/power/latency of electrical memory and computing cores. Previous solutions fail to fully-leverage the ultra-fast computational speed of emerging DNN accelerators to break through the critical memory bound. In this work, we propose a general and unified framework to trade expensive memory transactions with ultra-fast on-chip computations, directly translating to performance improvement. We are the first to jointly explore the intrinsic correlations and bit-level redundancy within DNN kernels and propose a multi-level in situ generation mechanism with mixed-precision bases to achieve on-the-fly recovery of high-resolution parameters with minimum hardware overhead. Extensive experiments demonstrate that our proposed joint method can boost the memory efficiency by 10-20x with comparable accuracy over four state-of-the-art designs when benchmarked on ResNet-18/DenseNet-121/MobileNetV2/V3 with various tasks.

The backbone of traditional CNN classifier is generally considered as a feature extractor, followed by a linear layer which performs the classification. We propose a novel loss function, termed as CAM-loss, to constrain the embedded feature maps with the class activation maps (CAMs) which indicate the spatially discriminative regions of an image for particular categories. CAM-loss drives the backbone to express the features of target category and suppress the features of non-target categories or background, so as to obtain more discriminative feature representations. It can be simply applied in any CNN architecture with neglectable additional parameters and calculations. Experimental results show that CAM-loss is applicable to a variety of network structures and can be combined with mainstream regularization methods to improve the performance of image classification. The strong generalization ability of CAM-loss is validated in the transfer learning and few shot learning tasks. Based on CAM-loss, we also propose a novel CAAM-CAM matching knowledge distillation method. This method directly uses the CAM generated by the teacher network to supervise the CAAM generated by the student network, which effectively improves the accuracy and convergence rate of the student network.

Deep networks for Monocular Depth Estimation (MDE) have achieved promising performance recently and it is of great importance to further understand the interpretability of these networks. Existing methods attempt to provide post-hoc explanations by investigating visual cues, which may not explore the internal representations learned by deep networks. In this paper, we find that some hidden units of the network are selective to certain ranges of depth, and thus such behavior can be served as a way to interpret the internal representations. Based on our observations, we quantify the interpretability of a deep MDE network by the depth selectivity of its hidden units. Moreover, we then propose a method to train interpretable MDE deep networks without changing their original architectures, by assigning a depth range for each unit to select. Experimental results demonstrate that our method is able to enhance the interpretability of deep MDE networks by largely improving the depth selectivity of their units, while not harming or even improving the depth estimation accuracy. We further provide comprehensive analysis to show the reliability of selective units, the applicability of our method on different models and layers, and a demonstration on monocular depth completion. We further provide comprehensive analysis to show the reliability of selective units, the applicability of our method on different layers, models, and datasets, and a demonstration on analysis of model error.