ICCV 2021 Papers

In this work, we propose a new feed-forward arbitrary style transfer method, referred to as StyleFormer, which can simultaneously fulfill fine-grained style diversity and semantic content coherency. Specifically, our transformer-inspired feature-level stylization method consists of three modules: (a) the style bank generation module for sparse but compact parametric style pattern extraction, (b) the transformer-driven style composition module for content-guided global style composition, and (c) the parametric content modulation module for flexible but faithful stylization. The output stylized images are impressively coherent with the content structure, sensitive to the detailed style variations, but still holistically adhere to the style distributions from the style images. Qualitative and quantitative comparisons as well as comprehensive user studies demonstrate that our StyleFormer outperforms the existing SOTA methods in generating visually plausible stylization results with real-time efficiency.

Inspired by the ability of StyleGAN to generate highly re-alistic images in a variety of domains, much recent work hasfocused on understanding how to use the latent spaces ofStyleGAN to manipulate generated and real images. How-ever, discovering semantically meaningful latent manipula-tions typically involves painstaking human examination ofthe many degrees of freedom, or an annotated collectionof images for each desired manipulation. In this work, weexplore leveraging the power of recently introduced Con-trastive Language-Image Pre-training (CLIP) models in or-der to develop a text-based interface for StyleGAN imagemanipulation that does not require such manual effort. Wefirst introduce an optimization scheme that utilizes a CLIP-based loss to modify an input latent vector in response to auser-provided text prompt. Next, we describe a latent map-per that infers a text-guided latent manipulation step fora given input image, allowing faster and more stable text-based manipulation. Finally, we present a method for map-ping a text prompts to input-agnostic directions in Style-GAN's style space, enabling interactive text-driven imagemanipulation. Extensive results and comparisons demon-strate the effectiveness of our approaches.

Knowledge distillation (KD) transfers the dark knowledge from cumbersome networks (teacher) to lightweight (student) networks and expects the student to achieve more promising performance than training without the teacher's knowledge. However, a counter-intuitive argument is that better teachers do not make better students due to the capacity mismatch. To this end, we present a novel adaptive knowledge distillation method to complement traditional approaches. The proposed method, named as Student Customized Knowledge Distillation (SCKD), examines the capacity mismatch between teacher and student from the perspective of gradient similarity. We formulate the knowledge distillation as a multi-task learning problem so that the teacher transfers knowledge to the student only if the student can benefit from learning such knowledge. We validate our methods on multiple datasets with various teacher-student configurations on image classification, object detection, and semantic segmentation.

This paper presents an explore-and-classify framework for structured architectural reconstruction from aerial image. Starting from a potentially imperfect building reconstruction by an existing algorithm, our approach 1) explores the space of building models by modifying the reconstruction via heuristic actions; 2) learns to classify the correctness of building models while generating classification labels based on the ground-truth; and 3) repeat. At test time, we iterate exploration and classification, seeking for a result with the best classification score. We evaluate the approach using initial reconstructions by two baselines and two state-of-the-art reconstruction algorithms. Qualitative and quantitative evaluations demonstrate that our approach consistently improves the reconstruction quality from every initial reconstruction.

Autonomous navigation requires structured representation of the road network and instance-wise identification of the other traffic agents. Since the traffic scene is defined on the ground plane, this corresponds to scene understanding in the bird's-eye-view (BEV). However, the onboard cameras of autonomous cars are customarily mounted horizontally for a better view of the surrounding, making this task very challenging. In this work, we study the problem of extracting a directed graph representing the local road network in BEV coordinates, from a single onboard camera image. Moreover, we show that the method can be extended to detect dynamic objects on the BEV plane. The semantics, locations, and orientations of the detected objects together with the road graph facilitates a comprehensive understanding of the scene. Such understanding becomes fundamental for the downstream tasks, such as path planning and navigation. We validate our approach against powerful baselines and show that our network achieves superior performance. We also demonstrate the effects of various design choices through ablation studies.

Pose-guided virtual try-on task aims to modify the fashion item based on pose transfer task. These two tasks that belong to person image synthesis have strong correlations and similarities. However, existing methods treat them as two individual tasks and do not explore correlations between them. Moreover, these two tasks are challenging due to large misalignment and occlusions, thus most of these methods are prone to generate unclear human body structure and blurry fine-grained textures. In this paper, we devise a structure-transformed texture-enhanced network to generate high-quality person images and construct the relationships between two tasks. It consists of two modules: structure-transformed renderer and texture-enhanced stylizer. The structure-transformed renderer is introduced to transform the source person structure to the target one, while the texture-enhanced stylizer is served to enhance detailed textures and controllably inject the fashion style founded on the structural transformation. With the two modules, our model can generate photorealistic person images in diverse poses and even with various fashion styles. Extensive experiments demonstrate that our approach achieves state-of-the-art results on two tasks.

Single image deraining is important for many high-level computer vision tasks since the rain streaks can severely degrade the visibility of images, thereby affecting the recognition and analysis of the image. Recently, many CNN-based methods have been proposed for rain removal. Although these methods can remove part of the rain streaks, it is difficult for them to adapt to real-world scenarios and restore high-quality rain-free images with clear and accurate structures. To solve this problem, we propose a Structure-Preserving Deraining Network (SPDNet) with RCP guidance. SPDNet directly generates high-quality rain-free images with clear and accurate structures under the guidance of RCP but does not rely on any rain-generating assumptions. Specifically, we found that the RCP of images contains more accurate structural information than rainy images. Therefore, we introduced it to our deraining network to protect structure information of the rain-free image. Meanwhile, a Wavelet-based Multi-Level Module (WMLM) is proposed as the backbone for learning the background information of rainy images and an Interactive Fusion Module (IFM) is designed to make full use of RCP information. In addition, an iterative guidance strategy is proposed to gradually improve the accuracy of RCP, refining the result in a progressive path. Extensive experimental results on both synthetic and real-world datasets demonstrate that the proposed model achieves new state-of-the-art results. Code: https://github.com/Joyies/SPDNet

Re-assembling multiple pots accurately from numerous 3D scanned fragments remains a challenging task to this date. Previous methods extract all potential matching pairs of pot sherds and considers them simultaneously to search for an optimal global pot configuration. In this work, we empirically show such global approach greatly suffers from false positive matches between sherds inflicted by indistinctive sharp fracture surfaces in pot fragments. To mitigate this problem, we take inspirations from the field of structure-from-motion (SfM), where many pipelines have matured in reconstructing a 3D scene from multiple images. Motivated by the success of the incremental approach in robust SfM, we present an efficient reassembly method for axially symmetric pots based on iterative registration of one sherd at a time. Our method goes beyond replicating incremental SfM and addresses indistinguishable false matches by embracing beam search to explore multitudes of registration possibilities. Additionally, we utilize multiple roots in each step to allow simultaneous reassembly of multiple pots. The proposed approach shows above 80% reassembly accuracy on a dataset of real 80 fragments mixed from 5 pots, pushing the state-of-the-art and paving the way towards the goal of large-scale pot reassembly. Our code and preprocessed data will be made available for research.

Self-supervised monocular depth estimation has achieved impressive performance on outdoor datasets. Its performance however degrades notably in indoor environments because of the lack of textures. Without rich textures, the photometric consistency is too weak to train a good depth network. Inspired by the early works on indoor modeling, we leverage the structural regularities exhibited in indoor scenes, to train a better depth network. Specifically, we adopt two extra supervisory signals for self-supervised training: 1) the Manhattan normal constraint and 2) the co-planar constraint. The Manhattan normal constraint enforces the major surfaces (the floor, ceiling, and walls) to be aligned with dominant directions. The co-planar constraint states that the 3D points be well fitted by a plane if they are located within the same planar region. To generate the supervisory signals, we adopt two components to classify the major surface normal into dominant directions and detect the planar regions on the fly during training. As the predicted depth becomes more accurate after more training epochs, the supervisory signals also improve and in turn feedback to obtain a better depth model. Through extensive experiments on indoor benchmark datasets, the results show that our network outperforms the state-of-the-art methods. The source code is available at https://github.com/SJTU-ViSYS/StructDepth.

Class-incremental learning (CIL) aims at continuously updating a trained model with new classes (plasticity) without forgetting previously learned old ones (stability). Contemporary studies resort to storing representative exemplars for rehearsal or preventing consolidated model parameters from drifting, but the former requires an additional space for storing exemplars at every incremental phase while the latter usually shows poor model generalization. In this paper, we focus on resolving the stability-plasticity dilemma in class-incremental learning where no exemplars from old classes are stored. To make a trade-off between learning new information and maintaining old knowledge, we reformulate a simple yet effective baseline method based on a cosine classifier framework and reciprocal adaptive weights. With the reformulated baseline, we present two new approaches to CIL by learning class-independent knowledge and multi-perspective knowledge, respectively. The former exploits class-independent knowledge to bridge learning new and old classes, while the latter learns knowledge from different perspectives to facilitate CIL. Extensive experiments on several widely used CIL benchmark datasets show the superiority of our approaches over the state-of-the-art methods.

Automating sign language translation (SLT) is a challenging real-world application. Despite its societal importance, though, research progress in the field remains rather poor. Crucially, existing methods that yield viable performance necessitate the availability of laborious to obtain gloss sequence groundtruth. In this paper, we attenuate this need, by introducing an end-to-end SLT model that does not entail explicit use of glosses; the model only needs text groundtruth. This is in stark contrast to existing end-to-end models that use gloss sequence groundtruth, either in the form of a modality that is recognized at an intermediate model stage, or in the form of a parallel output process, jointly trained with the SLT model. Our approach constitutes a Transformer network with a novel type of layers that combines: (i) local winner-takes-all (LWTA) layers with stochastic winner sampling, instead of conventional ReLU layers, (ii) stochastic weights with posterior distributions estimated via variational inference, and (iii) a weight compression technique at inference time that exploits estimated posterior variance to perform massive, almost lossless compression. We demonstrate that our approach can reach the currently best reported BLEU-4 score on the PHOENIX 2014T benchmark, but without making use of glosses for model training, and with a memory footprint reduced by more than 70%.

A long-standing goal in computer vision is to capture, model, and realistically synthesize human behavior. Specifically, by learning from data, our goal is to enable virtual humans to navigate within cluttered indoor scenes and naturally interact with objects. Such embodied behavior has applications in virtual reality, computer games, and robotics, while synthesized behavior can be used as a source of training data. This is challenging because real human motion is diverse and adapts to the scene. For example, a person can sit or lie on a sofa in many places and with varying styles. It is necessary to model this diversity when synthesizing virtual humans that realistically perform human-scene interactions. We present a novel data-driven, stochastic motion synthesis method that models different styles of performing a given action with a target object. Our method, called SAMP, for Scene-Aware Motion Prediction, generalizes to target objects of various geometries while enabling the character to navigate in cluttered scenes. To train our method, we collected MoCap data covering various sitting, lying down, walking, and running styles. We demonstrate our method on complex indoor scenes and achieve superior performance compared to existing solutions. Our code and data are available for research at https://samp.is.tue.mpg.de.

Learning mid-level representation for fine-grained recognition is easily dominated by a limited number of highly discriminative patterns, degrading its robustness and generalization capability. To this end, we propose a novel Stochastic Partial Swap (SPS) scheme to address this issue. Our method performs element-wise swapping for partial features between samples to inject noise during training. It equips a regularization effect similar to Dropout, which promotes more neurons to represent the concepts. Furthermore, it also exhibits other advantages: 1) suppressing over-activation to some part patterns to improve feature representativeness, and 2) enriching pattern combination and simulating noisy cases to enhance classifier generalization. We verify the effectiveness of our approach through comprehensive experiments across four network backbones and three fine-grained datasets. Moreover, we demonstrate its ability to complement high-level representations, allowing a simple model to achieve performance comparable to the top-performing technologies in fine-grained recognition, indoor scene recognition, and material recognition while improving model interpretability.

We present a large-scale stereo RGB image object pose estimation dataset named the StereOBJ-1M dataset. The dataset is designed to address challenging cases such as object transparency, translucency, and specular reflection, in addition to the common challenges of occlusion, symmetry, and variations in illumination and environments. In order to collect data of sufficient scale for modern deep learning models, we propose a novel method for efficiently annotating pose data in a multi-view fashion that allows data capturing in complex and flexible environments. Fully annotated with 6D object poses, our dataset contains over 396K frames and over 1.5M annotations of 18 objects recorded in 183 scenes constructed in 11 different environments. The 18 objects include 8 symmetric objects, 7 transparent objects, and 8 reflective objects. We benchmark two state-of-the-art pose estimation frameworks on StereOBJ-1M as baselines for future work. We also propose a novel object-level pose optimization method for computing 6D pose from keypoint predictions in multiple images.

The growing use of deep learning for a wide range of data problems has highlighted the need to understand and diagnose these models appropriately, making deep learning interpretation techniques an essential tool for data analysts. The numerous model interpretation methods proposed in recent years are generally based on heuristics, with little or no theoretical guarantees. Here we present a statistical framework for saliency estimation for black-box computer vision models. Our proposed model-agnostic estimation procedure, which is statistically consistent and capable of passing saliency checks, has polynomial-time computational efficiency since it only requires solving a linear program. An upper bound is established on the number of model evaluations needed to recover regions of importance with high probability through our theoretical analysis. Furthermore, a new perturbation scheme is presented for the estimation of local gradients that is more efficient than commonly used random perturbation schemes. The validity and excellence of our new method are demonstrated experimentally using sensitivity analysis on multiple datasets.

Image enhancement is a subjective process whose targets vary with user preferences. In this paper, we propose a deep learning-based image enhancement method covering multiple tonal styles using only a single model dubbed StarEnhancer. It can transform an image from one tonal style to another, even if that style is unseen. With a simple one-time setting, users can customize the model to make the enhanced images more in line with their aesthetics. To make the method more practical, we propose a well-designed enhancer that can process a 4K-resolution image over 200 FPS but surpasses the contemporaneous single style image enhancement methods in terms of PSNR, SSIM, and LPIPS. Finally, our proposed enhancement method has good interactability, which allows the user to fine-tune the enhanced image using intuitive options.

Identifying unexpected objects on roads in semantic segmentation (e.g., identifying dogs on roads) is crucial in safety-critical applications. Existing approaches use images of unexpected objects from external datasets or require additional training (e.g., retraining segmentation networks or training an extra network), which necessitate a non-trivial amount of labor intensity or lengthy inference time. One possible alternative is to use prediction scores of a pre-trained network such as the max logits (i.e., maximum values among classes before the final softmax layer) for detecting such objects. However, the distribution of max logits of each predicted class is significantly different from each other, which degrades the performance of identifying unexpected objects in urban-scene segmentation. To address this issue, we propose a simple yet effective approach that standardizes the max logits in order to align the different distributions and reflect the relative meanings of max logits within each predicted class. Moreover, we consider the local regions from two different perspectives based on the intuition that neighboring pixels share similar semantic information. In contrast to previous approaches, our method does not utilize any external datasets or require additional training, which makes our method widely applicable to existing pre-trained segmentation models. Such a straightforward approach achieves a new state-of-the-art performance on the publicly available Fishyscapes Lost & Found leaderboard with a large margin. Our code is publicly available at this link.

Multi-view pedestrian detection aims to predict a bird's eye view (BEV) occupancy map from multiple camera views. This task is confronted with two challenges: how to establish the 3D correspondences from views to the BEV map and how to assemble occupancy information across views. In this paper, we propose a novel Stacked HOmography Transformations (SHOT) approach, which is motivated by approximating projections in 3D world coordinates via a stack of homographies. We first construct a stack of transformations for projecting views to the ground plane at different height levels. Then we design a soft selection module so that the network learns to predict the likelihood of the stack of transformations. Moreover, we provide an in-depth theoretical analysis on constructing SHOT and how well SHOT approximates projections in 3D world coordinates. SHOT is empirically verified to be capable of estimating accurate correspondences from individual views to the BEV map, leading to new state-of-the-art performance on standard evaluation benchmarks.

In this paper we propose a novel square root sliding-window bundle adjustment suitable for real-time odometry applications. The square root formulation pervades three major aspects of our optimization-based sliding-window estimator: for bundle adjustment we eliminate landmark variables with nullspace projection; to store the marginalization prior we employ a matrix square root of the Hessian; and when marginalizing old poses we avoid forming normal equations and update the square root prior directly with a specialized QR decomposition. We show that the proposed square root marginalization is algebraically equivalent to the conventional use of Schur complement (SC) on the Hessian. Moreover, it elegantly deals with rank-deficient Jacobians producing a prior equivalent to SC with Moore--Penrose inverse. Our evaluation of visual and visual-inertial odometry on real-world datasets demonstrates that the proposed estimator is 36% faster than the baseline. It furthermore shows that in single precision, conventional Hessian-based marginalization leads to numeric failures and reduced accuracy. We analyse numeric properties of the marginalization prior to explain why our square root form does not suffer from the same effect and therefore entails superior performance.