ECCV 2018 Papers

We propose a dynamic filtering strategy with large sampling field for ConvNets (LS-DFN), where the position-specific kernels learn from not only the identical position but also multiple sampled neighbour regions. During sampling, residual learning is introduced to ease training and an attention mechanism is applied to fuse features from different samples. Such multiple samples enlarge the kernels’ receptive fields significantly without requiring more parameters. While LS-DFN inherits the advantages of DFN, namely avoiding feature map blurring by position-wise kernels while keeping translation invariance, it also efficiently alleviates the overfitting issue caused by much more parameters than normal CNNs. Our model is efficient and can be trained end-to-end via standard back-propagation. We demonstrate the merits of our LS-DFN on both sparse and dense prediction tasks involving object detection, semantic segmentation and flow estimation. Our results show LS-DFN enjoys stronger recognition abilities in object detection and semantic segmentation tasks on VOC benchmark and sharper responses in flow estimation on FlyingChairs dataset compared to strong baselines.

This paper proposes a novel Dynamic Conditional Convolutional Network (DCCN) to handle conditional few-shot learning, i.e, only a few training samples are available for each condition. DCCN consists of dual subnets: DyConvNet contains a dynamic convolutional layer with a bank of basis filters; CondiNet predicts a set of adaptive weights from conditional inputs to linearly combine the basis filters. In this manner, a specific convolutional kernel can be dynamically obtained for each conditional input. The filter bank is shared between all conditions thus only a low-dimension weight vector needs to be learned. This significantly facilitates the parameter learning across different conditions when training data are limited. We evaluate DCCN on four tasks which can be formulated as conditional model learning, including specific object counting, multi-modal image classification, phrase grounding and identity based face generation. Extensive experiments demonstrate the superiority of the proposed model in the conditional few-shot learning setting.

In this paper, we propose a dual-agent deep reinforcement learning (DADRL) method for deformable face tracking, which generates bounding boxes and detects facial landmarks interactively from face videos. Most existing deformable face tracking methods learn models for these two tasks individually, and perform these two procedures subsequently during the testing phase, which ignore the intrinsic connections of these two tasks. Motivated by the fact that the performance of facial landmark detection depends heavily on the accuracy of the generated bounding boxes, we exploit the interactions of these two tasks in probabilistic manner by following a Bayesian model and propose a unified framework for simultaneous bounding box tracking and landmark detection. By formulating it as a Markov decision process, we define two agents to exploit the relationships and pass messages via an adaptive sequence of actions under a deep reinforcement learning framework to iteratively adjust the positions of the bounding boxes and facial landmarks. Our proposed DADRL achieves performance improvements over the state-of-the-art deformable face tracking methods on the most challenging category of the 300-VW dataset.

Double JPEG detection is essential for detecting various image manipulations. This paper proposes a novel deep convolutional neural network for double JPEG detection using statistical histogram features from each block with a vectorized quantization table. In contrast to previous methods, the proposed approach handles mixed JPEG quality factors and is suitable for real-world situations. We collected real-world JPEG images from the image forensic service and generated a new double JPEG dataset with 1120 quantization tables to train the network. The proposed approach was verified experimentally to produce a state-of-the-art performance, successfully detecting various image manipulations.

We introduce a layer-wise unsupervised domain adaptation approach for the task of semantic segmentation. Instead of merely matching the output distributions of the source and target domains, our approach aligns the distributions of activations of intermediate layers. This scheme exhibits two key advantages. First, matching across intermediate layers introduces more constraints for training the network in the target domain, making the optimization problem better conditioned. Second, the matched activations at each layer provide similar inputs to the next layer for both training and adaptation, and thus alleviate covariate shift. We use a Generative Adversarial Network (or GAN) to align activation distributions. Experimental results show that our approach achieves state-of-the-art results on a variety of popular domain adaptation tasks, including (1) from GTA to Cityscapes for semantic segmentation, (2) from SYNTHIA to Cityscapes for semantic segmentation, and (3) adaptations on USPS and MNIST for image classification.

Drastic variations in illumination across surveillance cameras make the person re-identification problem extremely challenging. Current large scale re-identification datasets have a significant number of training subjects, but lack diversity in lighting conditions. As a result, a trained model requires fine-tuning to become effective under an unseen illumination condition. To alleviate this problem, we introduce a new synthetic dataset that contains hundreds of illumination conditions. Specifically, we use 100 virtual humans illuminated with multiple HDR environment maps which accurately model realistic indoor and outdoor lighting. To achieve better accuracy in unseen illumination conditions we propose a novel domain adaptation technique that takes advantage of our synthetic data and performs fine-tuning in a completely unsupervised way. Our approach yields significantly higher accuracy than semi-supervised and unsupervised state-of-the-art methods, and is very competitive with supervised techniques.

Hazy images are common in real scenarios and many dehazing methods have been developed to automatically remove the haze from images. Typically, the goal of image dehazing is to produce clearer images from which human vision can better identify the object and structural details present in the images. When the ground-truth haze-free image is available for a hazy image, quantitative evaluation of image dehazing is usually based on objective metrics, such as Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). However, in many applications, large-scale images are collected not for visual examination by human. Instead, they are used for many high-level vision tasks, such as automatic classification, recognition and categorization. One fundamental problem here is whether various dehazing methods can produce clearer images that can help improve the performance of the high-level tasks. In this paper, we empirically study this problem in the important task of image classification by using both synthetic and real hazy image datasets. From the experimental results, we find that the existing image-dehazing methods cannot improve much the image-classification performance and sometimes even reduce the image classification performance.

In this paper, we propose a new Dividing and Aggregating Network (DA-Net) for multi-view action recognition. In our DA-Net, we learn view-independent representations shared by all views at lower layers, while we learn one view-specific representation for each view at higher layers. We then train view-specific action classifiers based on the view-specific representation for each view and a view classifier based on the shared representation at lower layers. The view classifier is used to predict how likely each video belongs to each view. Finally, the predicted view probabilities from multiple views are used as the weights when fusing the prediction scores of view-specific action classifiers. We also propose a new approach based on the conditional random field (CRF) formulation to pass message among view-specific representations from different branches to help each other. Comprehensive experiments on two benchmark datasets clearly demonstrate the effectiveness of our proposed DA-Net for multi-view action recognition.

Visualizing features in deep neural networks (DNNs) can help understanding their computations. Many previous studies aimed to visualize the selectivity of individual units by finding meaningful images that maximize their activation. However, comparably little attention has been paid to visualizing to what image transformations units in DNNs are invariant. Here we propose a method to discover invariances in the responses of hidden layer units of deep neural networks. Our approach is based on simultaneously searching for a batch of images that strongly activate a unit while at the same time being as distinct from each other as possible. We find that even early convolutional layers in VGG-19 exhibit various forms of response invariance: near-perfect phase invariance in some units and invariance to local diffeomorphic transformations in others. At the same time, we uncover representational differences with ResNet-50 in its corresponding layers. We conclude that invariance transformations are a major computational component learned by DNNs and we provide a systematic method to study them.

Paragraph generation from images, which has gained popularity recently, is an important task for video summarization, editing, and support of the disabled. Traditional image captioning methods fall short on this front, since they aren't designed to generate long informative descriptions. Moreover, the vanilla approach of simply concatenating multiple short sentences, possibly synthesized from a classical image captioning system, doesn't embrace the intricacies of paragraphs: coherent sentences, globally consistent structure, and diversity. To address those challenges, we propose to augment paragraph generation techniques with ``coherence vectors,'' ``global topic vectors,'' and modeling of the inherent ambiguity of associating paragraphs with images, via a variational auto-encoder formulation. We demonstrate the effectiveness of the developed approach on two datasets, outperforming existing state-of-the-art techniques on both.

Image-to-image translation aims to learn the mapping between two visual domains. There are two main challenges for many applications: 1) the lack of aligned training pairs and 2) multiple possible outputs from a single input image. In this work, we present an approach based on disentangled representation for producing diverse outputs without paired training images. To achieve diversity, we propose to embed images onto two spaces: a domain-invariant content space capturing shared information across domains and a domain-specific attribute space. Using the disentangled features as inputs greatly reduces mode collapse. To handle unpaired training data, we introduce a novel cross-cycle consistency loss. Qualitative results show that our model can generate diverse and realistic images on a wide range of tasks. We validate the effectiveness of our approach through extensive evaluation.

Recent advances in Deep Learning and probabilistic modeling have let to strong improvements in generative models for images. On the one hand, GANs have contributed a highly effective adversarial learning procedure, but still suffer from stability issues. On the other hand, CVAE models provide a sound way of conditional modeling but suffer from mode-mixing issues. Therefore, recent work has turned back to simple and stable regression models that are effective at generation but give up on the sampling mechanism and the latent code representation. We propose a novel and efficient stochastic regression approach with latent drop-out codes that combines the merits of both lines of research. In addition, a new training objective enforces coverage of the training distribution leading to improvements over the state of the art in terms of accuracy as well as diversity.

Recently, Siamese networks have drawn great attention in visual tracking community because of their balanced accuracy and speed. However, features used in most Siamese tracking approaches can only discriminate foreground from the non-semantic backgrounds. The semantic backgrounds are always considered as distractors, which hinders the robustness of Siamese trackers. In this paper, we focus on learning distractor-aware Siamese networks for accurate and long-term tracking. To this end, features used in traditional Siamese trackers are analyzed at first. We observe that the imbalanced distribution of training data makes the learned features less discriminative. During the off-line training phase, an effective sampling strategy is introduced to control this distribution and make the model focus on the semantic distractors. During inference, a novel distractor-aware module is designed to perform incremental learning, which can effectively transfer the general embedding to the current video domain. In addition, we extend the proposed approach for long-term tracking by introducing a simple yet effective local-to-global search region strategy. Extensive experiments on benchmarks show that our approach significantly outperforms the state-of-the-arts, yielding 9.6% relative gain in VOT2016 dataset and 35.9% relative gain in UAV20L dataset. The proposed tracker can perform at 160 FPS on short-term benchmarks and 110 FPS on long-term benchmarks.

There is a high demand of 3D data for 360° panoramic images and videos, pushed by the growing availability on the market of specialized hardware for both capturing (e.g., omnidirectional cameras) as well as visualizing in 3D (e.g., head mounted displays) panoramic images and videos. At the same time, 3D sensors able to capture 3D panoramic data are expensive and/or hardly available. To fill this gap, we propose a learning approach for panoramic depth map estimation from a single image. Thanks to a specifically developed distortion-aware deformable convolution filter, our method can be trained by means of conventional perspective images, then used to regress depth for panoramic images, thus bypassing the effort needed to create annotated panoramic training dataset. We also demonstrate our approach for emerging tasks such as panoramic monocular SLAM, panoramic semantic segmentation and panoramic style transfer.

We introduce effective training algorithms for Generative Adversarial Networks (GAN) to alleviate mode collapse and gradient vanishing. In our system, we constrain the generator by an Autoencoder (AE). We propose a formulation to consider the reconstructed samples from AE as "real'' samples for the discriminator. This couples the convergence of the AE with that of the discriminator, effectively slowing down the convergence of discriminator and reducing gradient vanishing. Importantly, we propose two novel distance constraints to improve the generator. First, we propose a latent-data distance constraint to enforce compatibility between the latent sample distances and the corresponding data sample distances. We use this constraint to explicitly prevent the generator from mode collapse. Second, we propose a discriminator-score distance constraint to align the distribution of the generated samples with that of the real samples through the discriminator score. We use this constraint to guide the generator to synthesize samples that resemble the real ones. Our proposed GAN using these distance constraints, namely Dist-GAN, can achieve better results than state-of-the-art methods across benchmark datasets: synthetic, MNIST, MNIST-1K, CelebA, CIFAR-10 and STL-10 datasets. Our code is published here (https://github.com/tntrung/gan) for research.

Generative models that learn disentangled representations for different factors of variation in an image can be very useful for targeted data augmentation. By sampling from the disentangled latent subspace of interest, we can efficiently generate new data necessary for a particular task. Learning disentangled representations is a challenging problem, especially when certain factors of variation are difficult to label. In this paper, we introduce a novel architecture that disentangles the latent space into two complementary subspaces by using only weak supervision in form of pairwise similarity labels. Inspired by the recent success of cycle-consistent adversarial architectures, we use cycle-consistency in a variational auto-encoder framework. Our non-adversarial approach is in contrast with the recent works that combine adversarial training with auto-encoders to disentangle representations. We show compelling results of disentangled latent subspaces on three datasets and compare with recent works that leverage adversarial training.

Image-to-image translation has been made much progress with embracing Generative Adversarial Networks (GANs). However, it's still very challenging for translation tasks that require high quality, especially at high-resolution and photorealism. In this paper, we present Discriminative Region Proposal Adversarial Networks (DRPAN) for high-quality image-to-image translation. We decompose the procedure of image-to-image translation task into three iterated steps, first is to generate an image with global structure but some local artifacts (via GAN), second is using our DRPnet to propose the most fake region from the generated image, and third is to implement "image inpainting" on the most fake region for more realistic result through a reviser, so that the system (DRPAN) can be gradually optimized to synthesize images with more attention on the most artifact local part. Experiments on a variety of image-to-image translation tasks and datasets validate that our method outperforms state-of-the-arts for producing high-quality translation results in terms of both human perceptual studies and automatic quantitative measures.

Neglecting the effects of rolling-shutter cameras for visual odometry (VO) severely degrades accuracy and robustness. In this paper, we propose a novel direct monocular VO method that incorporates a rolling-shutter model. Our approach extends direct sparse odometry which performs direct bundle adjustment of a set of recent keyframe poses and the depths of a sparse set of image points. We estimate the velocity at each keyframe and impose a constant-velocity prior for the optimization. In this way, we obtain a near real-time, accurate direct VO method. Our approach achieves improved results on challenging rolling-shutter sequences over state-of-the-art global-shutter VO.

Despite the recent progress in video understanding and the continuous rate of improvement in temporal action localization throughout the years, it is still unclear how far (or close?) we are to solving the problem. To this end, we introduce a new diagnostic tool to analyze the performance of temporal action detectors in videos and compare different methods beyond a single scalar metric. We exemplify the use of our tool by analyzing the performance of the top rewarded entries in the latest ActivityNet action localization challenge. Our analysis shows that the most impactful areas to work on are: strategies to better handle temporal context around the instances, improving the robustness w.r.t. the instance absolute and relative size, and strategies to reduce the localization errors. Moreover, our experimental analysis finds the lack of agreement among annotator is not a major roadblock to attain progress in the field. Our diagnostic tool is publicly available to keep fueling the minds of other researchers with additional insights about their algorithms.