ICCV 2019 Papers

We introduce a novel approach for keypoint detection task that combines handcrafted and learned CNN filters within a shallow multi-scale architecture. Handcrafted filters provide anchor structures for learned filters, which localize, score and rank repeatable features. Scale-space representation is used within the network to extract keypoints at different levels. We design a loss function to detect robust features that exist across a range of scales and to maximize the repeatability score. Our Key.Net model is trained on data synthetically created from ImageNet and evaluated on HPatches benchmark. Results show that our approach outperforms state-of-the-art detectors in terms of repeatability, matching performance and complexity.

Deep convolutional neural networks (CNNs), trained on corresponding pairs of high- and low-resolution images, achieve state-of-the-art performance in single-image super-resolution and surpass previous signal-processing based approaches. However, their performance is limited when applied to real photographs. The reason lies in their training data: low-resolution (LR) images are obtained by bicubic interpolation of the corresponding high-resolution (HR) images. The applied convolution kernel significantly differs from real-world camera-blur. Consequently, while current CNNs well super-resolve bicubic-downsampled LR images, they often fail on camera-captured LR images. To improve generalization and robustness of deep super-resolution CNNs on real photographs, we present a kernel modeling super-resolution network (KMSR) that incorporates blur-kernel modeling in the training. Our proposed KMSR consists of two stages: we first build a pool of realistic blur-kernels with a generative adversarial network (GAN) and then we train a super-resolution network with HR and corresponding LR images constructed with the generated kernels. Our extensive experimental validations demonstrate the effectiveness of our single-image super-resolution approach on photographs with unknown blur-kernels.

We present Kernel Point Convolution (KPConv), a new design of point convolution, i.e. that operates on point clouds without any intermediate representation. The convolution weights of KPConv are located in Euclidean space by kernel points, and applied to the input points close to them. Its capacity to use any number of kernel points gives KPConv more flexibility than fixed grid convolutions. Furthermore, these locations are continuous in space and can be learned by the network. Therefore, KPConv can be extended to deformable convolutions that learn to adapt kernel points to local geometry. Thanks to a regular subsampling strategy, KPConv is also efficient and robust to varying densities. Whether they use deformable KPConv for complex tasks, or rigid KPconv for simpler tasks, our networks outperform state-of-the-art classification and segmentation approaches on several datasets. We also offer ablation studies and visualizations to provide understanding of what has been learned by KPConv and to validate the descriptive power of deformable KPConv.

In this paper, we introduce the problem of K-best transformation synchronization for the purpose of multiple scan matching. Given noisy pair-wise transformations computed between a subset of depth scan pairs, K-best transformation synchronization seeks to output multiple consistent relative transformations. This problem naturally arises in many geometry reconstruction applications, where the underlying object possesses self-symmetry. For approximately symmetric or even non-symmetric objects, K-best solutions offer an intermediate presentation for recovering the underlying single-best solution. We introduce a simple yet robust iterative algorithm for K-best transformation synchronization, which alternates between transformation propagation and transformation clustering. We present theoretical guarantees on the robust and exact recoveries of our algorithm. Experimental results demonstrate the advantage of our approach against state-of-the-art transformation synchronization techniques on both synthetic and real datasets.

Extrapolating fine-grained pixel-level correspondences in a fully unsupervised manner from a large set of misaligned images can benefit several computer vision and graphics problems, e.g. co-segmentation, super-resolution, image edit propagation, structure-from-motion, and 3D reconstruction. Several joint image alignment and congealing techniques have been proposed to tackle this problem, but robustness to initialisation, ability to scale to large datasets, and alignment accuracy seem to hamper their wide applicability. To overcome these limitations, we propose an unsupervised joint alignment method leveraging a densely fused spatial transformer network to estimate the warping parameters for each image and a low-capacity auto-encoder whose reconstruction error is used as an auxiliary measure of joint alignment. Experimental results on digits from multiple versions of MNIST (i.e., original, perturbed, affNIST and infiMNIST) and faces from LFW, show that our approach is capable of aligning millions of images with high accuracy and robustness to different levels and types of perturbation. Moreover, qualitative and quantitative results suggest that the proposed method outperforms state-of-the-art approaches both in terms of alignment quality and robustness to initialisation.

Video captioning is a challenging task that involves not only visual perception but also syntax representation learning. Recent progress in video captioning has been achieved through visual perception, but syntax representation learning is still under-explored. We propose a novel video captioning approach that takes into account both visual perception and syntax representation learning to generate accurate descriptions of videos. Specifically, we use sentence templates composed of Part-of-Speech (POS) tags to represent the syntax structure of captions, and accordingly, syntax representation learning is performed by directly inferring POS tags from videos. The visual perception is implemented by a mixture model which translates visual cues into lexical words that are conditional on the learned syntactic structure of sentences. Thus, a video captioning task consists of two sub-tasks: video POS tagging and visual cue translation, which are jointly modeled and trained in an end-to-end fashion. Evaluations on three public benchmark datasets demonstrate that our proposed method achieves substantially better performance than the state-of-the-art methods, which validates the superiority of joint modeling of syntax representation learning and visual perception for video captioning.

Trajectory prediction for objects is challenging and critical for various applications (e.g., autonomous driving, and anomaly detection). Most of the existing methods focus on homogeneous pedestrian trajectories prediction, where pedestrians are treated as particles without size. However, they fall short of handling crowded vehicle-pedestrian-mixed scenes directly since vehicles, limited with kinematics in reality, should be treated as rigid, non-particle objects ideally. In this paper, we tackle this problem using separate LSTMs for heterogeneous vehicles and pedestrians. Specifically, we use an oriented bounding box to represent each vehicle, calculated based on its position and orientation, to denote its kinematic trajectories. We then propose a framework called VP-LSTM to predict the kinematic trajectories of both vehicles and pedestrians simultaneously. In order to evaluate our model, a large dataset containing the trajectories of both vehicles and pedestrians in vehicle-pedestrian-mixed scenes is specially built. Through comparisons between our method with state-of-the-art approaches, we show the effectiveness and advantages of our method on kinematic trajectories prediction in vehicle-pedestrian-mixed scenes.

When describing images with natural language, descriptions can be made more informative if tuned for downstream tasks. This can be achieved by training two networks: a "speaker" that generates sentences given an image and a "listener" that uses them to perform a task. Unfortunately, training multiple networks jointly to communicate, faces two major challenges. First, the descriptions generated by a speaker network are discrete and stochastic, making optimization very hard and inefficient. Second, joint training usually causes the vocabulary used during communication to drift and diverge from natural language. To address these challenges, we present an effective optimization technique based on partial-sampling from a multinomial distribution combined with straight-through gradient updates, which we name PSST for Partial-Sampling Straight-Through. We then show that the generated descriptions can be kept close to natural by constraining them to be similar to human descriptions. Together, this approach creates descriptions that are both more discriminative and more natural than previous approaches. Evaluations on the COCO benchmark show that PSST improve the recall@10 from 60% to 86% maintaining comparable language naturalness. Human evaluations show that it also increases naturalness while keeping the discriminative power of generated captions.

Vehicle 3D extents and trajectories are critical cues for predicting the future location of vehicles and planning future agent ego-motion based on those predictions. In this paper, we propose a novel online framework for 3D vehicle detection and tracking from monocular videos. The framework can not only associate detections of vehicles in motion over time, but also estimate their complete 3D bounding box information from a sequence of 2D images captured on a moving platform. Our method leverages 3D box depth-ordering matching for robust instance association and utilizes 3D trajectory prediction for re-identification of occluded vehicles. We also design a motion learning module based on an LSTM for more accurate long-term motion extrapolation. Our experiments on simulation, KITTI, and Argoverse datasets show that our 3D tracking pipeline offers robust data association and tracking. On Argoverse, our image-based method is significantly better for tracking 3D vehicles within 30 meters than the LiDAR-centric baseline methods.

Convolutional neural networks (CNNs) based approaches for semantic alignment and object landmark detection have improved their performance significantly. Current efforts for the two tasks focus on addressing the lack of massive training data through weakly- or unsupervised learning frameworks. In this paper, we present a joint learning approach for obtaining dense correspondences and discovering object landmarks from semantically similar images. Based on the key insight that the two tasks can mutually provide supervisions to each other, our networks accomplish this through a joint loss function that alternatively imposes a consistency constraint between the two tasks, thereby boosting the performance and addressing the lack of training data in a principled manner. To the best of our knowledge, this is the first attempt to address the lack of training data for the two tasks through the joint learning. To further improve the robustness of our framework, we introduce a probabilistic learning formulation that allows only reliable matches to be used in the joint learning process. With the proposed method, state-of-the-art performance is attained on several benchmarks for semantic matching and landmark detection.

Existing weakly supervised semantic segmentation (WSSS) methods usually utilize the results of pre-trained saliency detection (SD) models without explicitly modelling the connections between the two tasks, which is not the most efficient configuration. Here we propose a unified multi-task learning framework to jointly solve WSSS and SD using a single network, i.e. saliency and segmentation network (SSNet). SSNet consists of a segmentation network (SN) and a saliency aggregation module (SAM). For an input image, SN generates the segmentation result and, SAM predicts the saliency of each category and aggregating the segmentation masks of all categories into a saliency map. The proposed network is trained end-to-end with image-level category labels and class-agnostic pixel-level saliency labels. Experiments on PASCAL VOC 2012 segmentation dataset and four saliency benchmark datasets show the performance of our method compares favorably against state-of-the-art weakly supervised segmentation methods and fully supervised saliency detection methods.

We propose a new Group Feature Selection method for Discriminative Correlation Filters (GFS-DCF) based visual object tracking. The key innovation of the proposed method is to perform group feature selection across both channel and spatial dimensions, thus to pinpoint the structural relevance of multi-channel features to the filtering system. In contrast to the widely used spatial regularisation or feature selection methods, to the best of our knowledge, this is the first time that channel selection has been advocated for DCF-based tracking. We demonstrate that our GFS-DCF method is able to significantly improve the performance of a DCF tracker equipped with deep neural network features. In addition, our GFS-DCF enables joint feature selection and filter learning, achieving enhanced discrimination and interpretability of the learned filters. To further improve the performance, we adaptively integrate historical information by constraining filters to be smooth across temporal frames, using an efficient low-rank approximation. By design, specific temporal-spatial-channel configurations are dynamically learned in the tracking process, highlighting the relevant features, and alleviating the performance degrading impact of less discriminative representations and reducing information redundancy. The experimental results obtained on OTB2013, OTB2015, VOT2017, VOT2018 and TrackingNet demonstrate the merits of our GFS-DCF and its superiority over the state-of-the-art trackers. The code is publicly available at https://github.com/XU-TIANYANG/GFS-DCF.

3D scan geometry and CAD models often contain complementary information towards understanding environments, which could be leveraged through establishing a mapping between the two domains. However, this is a challenging task due to strong, lower-level differences between scan and CAD geometry. We propose a novel approach to learn a joint embedding space between scan and CAD geometry, where semantically similar objects from both domains lie close together. To achieve this, we introduce a new 3D CNN-based approach to learn a joint embedding space representing object similarities across these domains. To learn a shared space where scan objects and CAD models can interlace, we propose a stacked hourglass approach to separate foreground and background from a scan object, and transform it to a complete, CAD-like representation to produce a shared embedding space. This embedding space can then be used for CAD model retrieval; to further enable this task, we introduce a new dataset of ranked scan-CAD similarity annotations, enabling new, fine-grained evaluation of CAD model retrieval to cluttered, noisy, partial scans. Our learned joint embedding outperforms current state of the art for CAD model retrieval by 12% in instance retrieval accuracy.

Demosaicking and denoising are the first steps of any camera image processing pipeline and are key for obtaining high quality RGB images. A promising current research trend aims at solving these two problems jointly using convolutional neural networks. Due to the unavailability of ground truth data these networks cannot be currently trained using real RAW images. Instead, they resort to simulated data. In this paper we present a method to learn demosaicking directly from mosaicked images, without requiring ground truth RGB data. We apply this to learn joint demosaicking and denoising only from RAW images, thus enabling the use of real data. In addition we show that for this application fine-tuning a network to a specific burst improves the quality of restoration for both demosaicking and denoising.

Accurate grading of skin disease severity plays a crucial role in precise treatment for patients. Acne vulgaris, the most common skin disease in adolescence, can be graded by evidence-based lesion counting as well as experience-based global estimation in the medical field. However, due to the appearance similarity of acne with close severity, it is challenging to count and grade acne accurately. In this paper, we address the problem of acne image analysis via Label Distribution Learning (LDL) considering the ambiguous information among acne severity. Based on the professional grading criterion, we generate two acne label distributions considering the relationship between the similar number of lesions and severity of acne, respectively. We also propose a unified framework for joint acne image grading and counting, which is optimized by the multi-task learning loss. In addition, we further build the ACNE04 dataset with annotations of acne severity and lesion number of each image for evaluation. Experiments demonstrate that our proposed framework performs favorably against state-of-the-art methods. We make the code and dataset publicly available at https://github.com/xpwu95/ldl.

To effectively reduce JPEG compression artifacts, we propose a deep convolutional sparse coding (DCSC) network architecture. We design our DCSC in the framework of classic learned iterative shrinkage-threshold algorithm. To focus on recognizing and separating artifacts only, we sparsely code the feature maps instead of the raw image. The final de-blocked image is directly reconstructed from the coded features. We use dilated convolution to extract multi-scale image features, which allows our single model to simultaneously handle multiple JPEG compression levels. Since our method integrates model-based convolutional sparse coding with a learning-based deep neural network, the entire network structure is compact and more explainable. The resulting lightweight model generates comparable or better de-blocking results when compared with state-of-the-art methods.

Subspace clustering methods based on expressing each data point as a linear combination of other data points have achieved great success in computer vision applications such as motion segmentation, face and digit clustering. In face clustering, the subspaces are linear and subspace clustering methods can be applied directly. In motion segmentation, the subspaces are affine and an additional affine constraint on the coefficients is often enforced. However, since affine subspaces can always be embedded into linear subspaces of one extra dimension, it is unclear if the affine constraint is really necessary. This paper shows, both theoretically and empirically, that when the dimension of the ambient space is high relative to the sum of the dimensions of the affine subspaces, the affine constraint has a negligible effect on clustering performance. Specifically, our analysis provides conditions that guarantee the correctness of affine subspace clustering methods both with and without the affine constraint, and shows that these conditions are satisfied for high-dimensional data. Underlying our analysis is the notion of affinely independent subspaces, which not only provides geometrically interpretable correctness conditions, but also clarifies the relationships between existing results for affine subspace clustering.

Visual localization in large and complex indoor scenes, dominated by weakly textured rooms and repeating geometric patterns, is a challenging problem with high practical relevance for applications such as Augmented Reality and robotics. To handle the ambiguities arising in this scenario, a common strategy is, first, to generate multiple estimates for the camera pose from which a given query image was taken. The pose with the largest geometric consistency with the query image, e.g., in the form of an inlier count, is then selected in a second stage. While a significant amount of research has concentrated on the first stage, there has been considerably less work on the second stage. In this paper, we thus focus on pose verification. We show that combining different modalities, namely appearance, geometry, and semantics, considerably boosts pose verification and consequently pose accuracy. We develop multiple hand-crafted as well as a trainable approach to join into the geometric-semantic verification and show significant improvements over state-of-the-art on a very challenging indoor dataset.

We present a novel clustering objective that learns a neural network classifier from scratch, given only unlabelled data samples. The model discovers clusters that accurately match semantic classes, achieving state-of-the-art results in eight unsupervised clustering benchmarks spanning image classification and segmentation. These include STL10, an unsupervised variant of ImageNet, and CIFAR10, where we significantly beat the accuracy of our closest competitors by 6.6 and 9.5 absolute percentage points respectively. The method is not specialised to computer vision and operates on any paired dataset samples; in our experiments we use random transforms to obtain a pair from each image. The trained network directly outputs semantic labels, rather than high dimensional representations that need external processing to be usable for semantic clustering. The objective is simply to maximise mutual information between the class assignments of each pair. It is easy to implement and rigorously grounded in information theory, meaning we effortlessly avoid degenerate solutions that other clustering methods are susceptible to. In addition to the fully unsupervised mode, we also test two semi-supervised settings. The first achieves 88.8% accuracy on STL10 classification, setting a new global state-of-the-art over all existing methods (whether supervised, semi-supervised or unsupervised). The second shows robustness to 90% reductions in label coverage, of relevance to applications that wish to make use of small amounts of labels. github.com/xu-ji/IIC