ICCV 2019 Papers

In this paper, we propose a fast and practical neural architecture search (FPNAS) framework for automatic network design. FPNAS aims to discover extremely efficient networks with less than 300M FLOPs. Different from previous NAS methods, our approach searches for the whole network architecture to guarantee block diversity instead of stacking a set of similar blocks repeatedly. We model the search process as a bi-level optimization problem and propose an approximation solution. On CIFAR-10, our approach is capable of design networks with comparable performance to state-of-the-arts while using orders of magnitude less computational resource with only 20 GPU hours. Experimental results on ImageNet and ADE20K datasets further demonstrate transferability of the searched networks.

We propose a new model for fast and accurate video object segmentation. It consists of two convolutional neural networks, a Dynamic Targeting Network (DTN) and a Mask Refinement Network (MRN). DTN locates the object by dynamically focusing on regions of interest surrounding the target object. The target region is predicted by DTN via two sub-streams, Box Propagation (BP) and Box Re-identification (BR). The BP stream is faster but less effective at objects with large deformation or occlusion. The BR stream performs better in difficult scenarios at a higher computation cost. We propose a Decision Module (DM) to adaptively determine which sub-stream to use for each frame. Finally, MRN is exploited to predict segmentation within the target region. Experimental results on two public datasets demonstrate that the proposed model significantly outperforms existing methods without online training in both accuracy and efficiency, and is comparable to online training-based methods in accuracy with an order of magnitude faster speed.

We present a unified, efficient and effective framework for point-cloud based 3D object detection. Our two-stage approach utilizes both voxel representation and raw point cloud data to exploit respective advantages. The first stage network, with voxel representation as input, only consists of light convolutional operations, producing a small number of high-quality initial predictions. Coordinate and indexed convolutional feature of each point in initial prediction are effectively fused with the attention mechanism, preserving both accurate localization and context information. The second stage works on interior points with their fused feature for further refining the prediction. Our method is evaluated on KITTI dataset, in terms of both 3D and Bird's Eye View (BEV) detection, and achieves state-of-the-arts with a 15FPS detection rate.

Object detection in videos has drawn increasing attention since it is more practical in real scenarios. Most of the deep learning methods use CNNs to process each decoded frame in a video stream individually. However, the free of charge yet valuable motion information already embedded in the video compression format is usually overlooked. In this paper, we propose a fast object detection method by taking advantage of this with a novel Motion aided Memory Network (MMNet). The MMNet has two major advantages: 1) It significantly accelerates the procedure of feature extraction for compressed videos. It only need to run a complete recognition network for I-frames, i.e. a few reference frames in a video, and it produces the features for the following P frames (predictive frames) with a light weight memory network, which runs fast; 2) Unlike existing methods that establish an additional network to model motion of frames, we take full advantage of both motion vectors and residual errors that are freely available in video streams. To our best knowledge, the MMNet is the first work that investigates a deep convolutional detector on compressed videos. Our method is evaluated on the large-scale ImageNet VID dataset, and the results show that it is 3x times faster than single image detector R-FCN and 10x times faster than high-performance detector MANet at a minor accuracy loss.

Tremendous advances in image restoration tasks such as denoising and super-resolution have been achieved using neural networks. Such approaches generally employ very deep architectures, large number of parameters, large receptive fields and high nonlinear modeling capacity. In order to obtain efficient and fast image restoration networks one should improve upon the above mentioned requirements. In this paper we propose a novel activation function, the multi-bin trainable linear unit (MTLU), for increasing the nonlinear modeling capacity together with lighter and shallower networks. We validate the proposed fast image restoration networks for image denoising (FDnet) and super-resolution (FSRnet) on standard benchmarks. We achieve large improvements in both memory and runtime over current state-of-the-art for comparable or better PSNR accuracies.

State-of-the-art researches model the data of images and videos as low-dimensional manifolds and generate superpixels/supervoxels in a content-sensitive way, which is achieved by computing geodesic centroidal Voronoi tessellation (GCVT) on manifolds. However, computing exact GCVTs is slow due to computationally expensive geodesic distances. In this paper, we propose a much faster queue-based graph distance (called q-distance). Our key idea is that for manifold regions in which q-distances are different from geodesic distances, GCVT is prone to placing more generators in them, and therefore after few iterations, the q-distance-induced tessellation is an exact GCVT. This idea works well in practice and we also prove it theoretically under moderate assumption. Our method is simple and easy to implement. It runs 6-8 times faster than state-of-the-art GCVT computation, and has an optimal approximation ratio O(1) and a linear time complexity O(N) for N-pixel images or N-voxel videos. A thorough evaluation of 31 superpixel methods on five image datasets and 8 supervoxel methods on four video datasets shows that our method consistently achieves the best over-segmentation accuracy. We also demonstrate the advantage of our method on one image and two video applications.

Given an outfit, what small changes would most improve its fashionability? This question presents an intriguing new computer vision challenge. We introduce Fashion++, an approach that proposes minimal adjustments to a full-body clothing outfit that will have maximal impact on its fashionability. Our model consists of a deep image generation neural network that learns to synthesize clothing conditioned on learned per-garment encodings. The latent encodings are explicitly factorized according to shape and texture, thereby allowing direct edits for both fit/presentation and color/patterns/material, respectively. We show how to bootstrap Web photos to automatically train a fashionability model, and develop an activation maximization-style approach to transform the input image into its more fashionable self. The edits suggested range from swapping in a new garment to tweaking its color, how it is worn (e.g., rolling up sleeves), or its fit (e.g., making pants baggier). Experiments demonstrate that Fashion++ provides successful edits, both according to automated metrics and human opinion.

Recently, large-margin softmax loss methods, such as angular softmax loss (SphereFace), large margin cosine loss (CosFace), and additive angular margin loss (ArcFace), have demonstrated impressive performance on deep face recognition. These methods incorporate a fixed additive margin to all the classes, ignoring the class imbalance problem. However, imbalanced problem widely exists in various real-world face datasets, in which samples from some classes are in a higher number than others. We argue that the number of a class would influence its demand for the additive margin. In this paper, we introduce a new margin-aware reinforcement learning based loss function, namely fair loss, in which each class will learn an appropriate adaptive margin by Deep Q-learning. Specifically, we train an agent to learn a margin adaptive strategy for each class, and make the additive margins for different classes more reasonable. Our method has better performance than present large-margin loss functions on three benchmarks, Labeled Face in the Wild (LFW), Youtube Faces (YTF) and MegaFace, which demonstrates that our method could learn better face representation on imbalanced face datasets.

The rapid progress in synthetic image generation and manipulation has now come to a point where it raises significant concerns for the implications towards society. At best, this leads to a loss of trust in digital content, but could potentially cause further harm by spreading false information or fake news. This paper examines the realism of state-of-the-art image manipulations, and how difficult it is to detect them, either automatically or by humans. To standardize the evaluation of detection methods, we propose an automated benchmark for facial manipulation detection. In particular, the benchmark is based on Deep-Fakes, Face2Face, FaceSwap and NeuralTextures as prominent representatives for facial manipulations at random compression level and size. The benchmark is publicly available and contains a hidden test set as well as a database of over 1.8 million manipulated images. This dataset is over an order of magnitude larger than comparable, publicly available, forgery datasets. Based on this data, we performed a thorough analysis of data-driven forgery detectors. We show that the use of additional domain-specific knowledge improves forgery detection to unprecedented accuracy, even in the presence of strong compression, and clearly outperforms human observers.

Character customization system is an important component in Role-Playing Games (RPGs), where players are allowed to edit the facial appearance of their in-game characters with their own preferences rather than using default templates. This paper proposes a method for automatically creating in-game characters of players according to an input face photo. We formulate the above "artistic creation" process under a facial similarity measurement and parameter searching paradigm by solving an optimization problem over a large set of physically meaningful facial parameters. To effectively minimize the distance between the created face and the real one, two loss functions, i.e. a "discriminative loss" and a "facial content loss", are specifically designed. As the rendering process of a game engine is not differentiable, a generative network is further introduced as an "imitator" to imitate the physical behavior of the game engine so that the proposed method can be implemented under a neural style transfer framework and the parameters can be optimized by gradient descent. Experimental results demonstrate that our method achieves a high degree of generation similarity between the input face photo and the created in-game character in terms of both global appearance and local details. Our method has been deployed in a new game last year and has now been used by players over 1 million times.

In recent decades, 3D morphable model (3DMM) has been commonly used in image-based photorealistic 3D face reconstruction. However, face images are often corrupted by serious occlusion by non-face objects including eyeglasses, masks, and hands. Such objects block the correct capture of landmarks and shading information. Therefore, the reconstructed 3D face model is hardly reusable. In this paper, a novel method is proposed to restore de-occluded face images based on inverse use of 3DMM and generative adversarial network. We utilize the 3DMM prior to the proposed adversarial network and combine a global and local adversarial convolutional neural network to learn face de-occlusion model. The 3DMM serves not only as geometric prior but also proposes the face region for the local discriminator. Experiment results confirm the effectiveness and robustness of the proposed algorithm in removing challenging types of occlusions with various head poses and illumination. Furthermore, the proposed method reconstructs the correct 3D face model with de-occluded textures.

Deep neural networks achieve good performance in many computer vision problems such as face alignment. However, when the testing image is challenging due to low resolution, occlusion or adversarial attacks, the accuracy of a deep neural network suffers greatly. Therefore, it is important to quantify the uncertainty in its predictions. A probabilistic neural network with Gaussian distribution over the target is typically used to quantify uncertainty for regression problems. However, in real-world problems especially computer vision tasks, the Gaussian assumption is too strong. To model more general distributions, such as multi-modal or asymmetric distributions, we propose to develop a kernel density deep neural network. Specifically, for face alignment, we adapt state-of-the-art hourglass neural network into a probabilistic neural network framework with landmark probability map as its output. The model is trained by maximizing the conditional log likelihood. To exploit the output probability map, we extend the model to multi-stage so that the logits map from the previous stage can feed into the next stage to progressively improve the landmark detection accuracy. Extensive experiments on benchmark datasets against state-of-the-art unconstrained deep learning method demonstrate that the proposed kernel density network achieves comparable or superior performance in terms of prediction accuracy. It further provides aleatoric uncertainty estimation in predictions.

Beyond current image-based virtual try-on systems that have attracted increasing attention, we move a step forward to developing a video virtual try-on system that precisely transfers clothes onto the person and generates visually realistic videos conditioned on arbitrary poses. Besides the challenges in image-based virtual try-on (e.g., clothes fidelity, image synthesis), video virtual try-on further requires spatiotemporal consistency. Directly adopting existing image-based approaches often fails to generate coherent video with natural and realistic textures. In this work, we propose Flow-navigated Warping Generative Adversarial Network (FW-GAN), a novel framework that learns to synthesize the video of virtual try-on based on a person image, the desired clothes image, and a series of target poses. FW-GAN aims to synthesize the coherent and natural video while manipulating the pose and clothes. It consists of: (i) a flow-guided fusion module that warps the past frames to assist synthesis, which is also adopted in the discriminator to help enhance the coherence and quality of the synthesized video; (ii) a warping net that is designed to warp clothes image for the refinement of clothes textures; (iii) a parsing constraint loss that alleviates the problem caused by the misalignment of segmentation maps from images with different poses and various clothes. Experiments on our newly collected dataset show that FW-GAN can synthesize high-quality video of virtual try-on and significantly outperforms other methods both qualitatively and quantitatively.

We present Face Swapping GAN (FSGAN) for face swapping and reenactment. Unlike previous work, FSGAN is subject agnostic and can be applied to pairs of faces without requiring training on those faces. To this end, we describe a number of technical contributions. We derive a novel recurrent neural network (RNN)-based approach for face reenactment which adjusts for both pose and expression variations and can be applied to a single image or a video sequence. For video sequences, we introduce continuous interpolation of the face views based on reenactment, Delaunay Triangulation, and barycentric coordinates. Occluded face regions are handled by a face completion network. Finally, we use a face blending network for seamless blending of the two faces while preserving target skin color and lighting conditions. This network uses a novel Poisson blending loss which combines Poisson optimization with perceptual loss. We compare our approach to existing state-of-the-art systems and show our results to be both qualitatively and quantitatively superior.

Though Deep Neural Networks (DNN) show excellent performance across various computer vision tasks, several works show their vulnerability to adversarial samples, i.e., image samples with imperceptible noise engineered to manipulate the network's prediction. Adversarial sample generation methods range from simple to complex optimization techniques. Majority of these methods generate adversaries through optimization objectives that are tied to the pre-softmax or softmax output of the network. In this work we, (i) show the drawbacks of such attacks, (ii) propose two new evaluation metrics: Old Label New Rank (OLNR) and New Label Old Rank (NLOR) in order to quantify the extent of damage made by an attack, and (iii) propose a new attack FDA: Feature Disruptive attack, to address the drawbacks of existing attacks. FDA works by generating image perturbation that disrupts features at each layer of the network and causes deep-features to be highly corrupt. This allows FDA adversaries to severely reduce the performance of deep networks. We experimentally validate that FDA generates stronger adversaries than other state-of-the-art methods for Image classification, even in the presence of various defense measures. More importantly, we show that FDA disrupts feature-representation based tasks even without access to the task-specific network or methodology.

We propose a fully convolutional one-stage object detector (FCOS) to solve object detection in a per-pixel prediction fashion, analogue to semantic segmentation. Almost all state-of-the-art object detectors such as RetinaNet, SSD, YOLOv3, and Faster R-CNN rely on pre-defined anchor boxes. In contrast, our proposed detector FCOS is anchor box free, as well as proposal free. By eliminating the pre-defined set of anchor boxes, FCOS completely avoids the complicated computation related to anchor boxes such as calculating overlapping during training. More importantly, we also avoid all hyper-parameters related to anchor boxes, which are often very sensitive to the final detection performance. With the only post-processing non-maximum suppression (NMS), FCOS with ResNeXt-64x4d-101 achieves 44.7% in AP with single-model and single-scale testing, surpassing previous one-stage detectors with the advantage of being much simpler. For the first time, we demonstrate a much simpler and flexible detection framework achieving improved detection accuracy. We hope that the proposed FCOS framework can serve as a simple and strong alternative for many other instance-level tasks. Code is available at: https://tinyurl.com/FCOSv1

Current methods for body shape estimation either lack detail or require many images. They are usually architecturally complex and computationally expensive. We propose FACSIMILE (FAX), a method that estimates a detailed body from a single photo, lowering the bar for creating virtual representations of humans. Our approach is easy to implement and fast to execute, making it easily deployable. FAX uses an image-translation network which recovers geometry at the original resolution of the image. Counterintuitively, the main loss which drives FAX is on per-pixel surface normals instead of per-pixel depth, making it possible to estimate detailed body geometry without any depth supervision. We evaluate our approach both qualitatively and quantitatively, and compare with a state-of-the-art method.

Recently, facial landmark detection algorithms have achieved remarkable performance on static images. However, these algorithms are neither accurate nor stable in motion-blurred videos. The missing of structure information makes it difficult for state-of-the-art facial landmark detection algorithms to yield good results. In this paper, we propose a framework named FAB that takes advantage of structure consistency in the temporal dimension for facial landmark detection in motion-blurred videos. A structure predictor is proposed to predict the missing face structural information temporally, which serves as a geometry prior. This allows our framework to work as a virtuous circle. On one hand, the geometry prior helps our structure-aware deblurring network generates high quality deblurred images which lead to better landmark detection results. On the other hand, better landmark detection results help structure predictor generate better geometry prior for the next frame. Moreover, it is a flexible video-based framework that can incorporate any static image-based methods to provide a performance boost on video datasets. Extensive experiments on Blurred-300VW, the proposed Real-world Motion Blur (RWMB) datasets and 300VW demonstrate the superior performance to the state-of-the-art methods. Datasets and model will be publicly available at https://github.com/KeqiangSun/FAB https://github.com/KeqiangSun/FAB .

Image captioning is a research hotspot where encoder-decoder models combining convolutional neural network (CNN) and long short-term memory (LSTM) achieve promising results. Despite significant progress, these models generate sentences differently from human cognitive styles. Existing models often generate a complete sentence from the first word to the end, without considering the influence of the following words on the whole sentence generation. In this paper, we explore the utilization of a human-like cognitive style, i.e., building overall cognition for the image to be described and the sentence to be constructed, for enhancing computer image understanding. This paper first proposes a Mutual-aid network structure with Bidirectional LSTMs (MaBi-LSTMs) for acquiring overall contextual information. In the training process, the forward and backward LSTMs encode the succeeding and preceding words into their respective hidden states by simultaneously constructing the whole sentence in a complementary manner. In the captioning process, the LSTM implicitly utilizes the subsequent semantic information contained in its hidden states. In fact, MaBi-LSTMs can generate two sentences in forward and backward directions. To bridge the gap between cross-domain models and generate a sentence with higher quality, we further develop a cross-modal attention mechanism to retouch the two sentences by fusing their salient parts as well as the salient areas of the image. Experimental results on the Microsoft COCO dataset show that the proposed model improves the performance of encoder-decoder models and achieves state-of-the-art results.