ICCV 2021 Papers

Visual localization is the problem of estimating the position and orientation from which a given image (or a sequence of images) is taken in a known scene. It is an important part of a wide range of computer vision and robotics applications, from self-driving cars to augmented/virtual reality systems. Visual localization techniques should work reliably and robustly under a wide range of conditions, including seasonal, weather, illumination and man-made changes. Recent benchmarking efforts model this by providing images under different conditions, and the community has made rapid progress on these datasets since their inception. However, they are limited to a few geographical regions and often recorded with a single device. We propose a new benchmark for visual localization in outdoor scenes, using crowd-sourced data to cover a wide range of geographical regions and camera devices with a focus on the failure cases of current algorithms. Experiments with state-of-the-art localization approaches show that our dataset is very challenging, with all evaluated methods failing on its hardest parts. As part of the dataset release, we provide the tooling used to generate it, enabling efficient and effective 2D correspondence annotation to obtain reference poses.

To fully leverage the data captured from different scenes with different view angles while reducing the annotation cost, this paper studies a novel crowd counting setting, i.e. only using partial annotations in each image as training data. Inspired by the repetitive patterns in the annotated and unannotated regions as well as the ones between them, we design a network with three components to tackle those unannotated regions: i) in an Unannotated Regions Characterization (URC) module, we employ a memory bank to only store the annotated features, which could help the visual features extracted from these annotated regions flow to these unannotated regions; ii) For each image, Feature Distribution Consistency (FDC) regularizes the feature distributions of annotated head and unannotated head regions to be consistent; iii) a Cross-regressor Consistency Regularization (CCR) module is designed to learn the visual features of unannotated regions in a self-supervised style. The experimental results validate the effectiveness of our proposed model under the partial annotation setting for several datasets, such as ShanghaiTech, UCF-CC-50, UCF-QNRF, NWPU-Crowd, and JHU-CROWD++. With only 10% annotated regions in each image, our proposed model achieves better performance than the recent methods and baselines under semi-supervised or active learning settings on all datasets. The code is https://github.com/svip-lab/CrwodCountingPAL.

Modeling temporal visual context across frames is critical for video instance segmentation (VIS) and other video understanding tasks. In this paper, we propose a fast online VIS model termed CrossVIS. For temporal information modeling in VIS, we present a novel crossover learning scheme that uses the instance feature in the current frame to pixel-wisely localize the same instance in other frames. Different from previous schemes, crossover learning does not require any additional network parameters for feature enhancement. By integrating with the instance segmentation loss, crossover learning enables efficient cross-frame instance-to-pixel relation learning and brings cost-free improvement during inference. Besides, a global balanced instance embedding branch is proposed for better and more stable online instance association. We conduct extensive experiments on three challenging VIS benchmarks, i.e., YouTube-VIS-2019, OVIS, and YouTube-VIS-2021 to evaluate our methods. CrossVIS achieves state-of-the-art online VIS performance and shows a decent trade-off between latency and accuracy. Code is available at https://github.com/hustvl/CrossVIS.

The recently developed vision transformer (ViT) has achieved promising results on image classification compared to convolutional neural networks. Inspired by this, in this paper, we study how to learn multi-scale feature representations in transformer models for image classification. To this end, we propose a dual-branch transformer to combine image patches (i.e., tokens in a transformer) of different sizes to produce stronger image features. Our approach processes small-patch and large-patch tokens with two separate branches of different computational complexity and these tokens are then fused purely by attention multiple times to complement each other. Furthermore, to reduce computation, we develop a simple yet effective token fusion module based on cross attention, which uses a single token for each branch as a query to exchange information with other branches. Our proposed cross-attention only requires linear time for both computational and memory complexity instead of quadratic time otherwise. Extensive experiments demonstrate that our approach performs better than or on par with several concurrent works on vision transformer, in addition to efficient CNN models. For example, on the ImageNet1K dataset, with some architectural changes, our approach outperforms the recent DeiT by a large margin of 2% with a small to moderate increase in FLOPs and model parameters. Our source codes and models are available at https://github.com/IBM/CrossViT.

Traditional normalization techniques (e.g., Batch Normalization and Instance Normalization) generally and simplistically assume that training and test data follow the same distribution. As distribution shifts are inevitable in real-world applications, well-trained models with previous normalization methods can perform badly in new environments. Can we develop new normalization methods to improve generalization robustness under distribution shifts? In this paper, we answer the question by proposing CrossNorm and SelfNorm. CrossNorm exchanges channel-wise mean and variance between feature maps to enlarge training distribution, while SelfNorm uses attention to recalibrate the statistics to bridge gaps between training and test distributions. CrossNorm and SelfNorm can complement each other, though exploring different directions in statistics usage. Extensive experiments on different fields (vision and language), tasks (classification and segmentation), settings (supervised and semi-supervised), and distribution shift types (synthetic and natural) show the effectiveness. Code is available at https://github.com/amazon-research/crossnorm-selfnorm

Object detection aims to accurately locate and classify objects in an image, which requires precise object representations. Existing methods usually use rectangular anchor boxes or a set of points to represent objects. However, these methods either introduce background noise or miss the continuous appearance information inside the object, and thus cause incorrect detection results. In this paper, we propose a novel anchor-free object detection network, called CrossDet, which uses a set of growing cross lines along horizontal and vertical axes as object representations. An object can be flexibly represented as cross lines in different combinations. It not only can effectively reduce the interference of noise, but also takes into account the continuous object information, which is useful to enhance the discriminability of object features and find the object boundaries. Based on the learned cross lines, we propose a crossline extraction module to adaptively capture features of cross lines. Furthermore, we design a decoupled regression mechanism to regress the localization along the horizontal and vertical directions respectively, which helps to decrease the optimization difficulty because the optimization space is limited to a specific direction. Our method achieves consistently improvement on the PASCAL VOC and MS-COCO datasets. The experiment results demonstrate the effectiveness of our proposed method.

Contrastive learning allows us to flexibly define powerful losses by contrasting positive pairs from sets of negative samples. Recently, the principle has also been used to learn cross-modal embeddings for video and text, yet without exploiting its full potential. In particular, previous losses do not take the intra-modality similarities into account, which leads to inefficient embeddings, as the same content is mapped to multiple points in the embedding space. With CrossCLR, we present a contrastive loss that fixes this issue. Moreover, we define sets of highly related samples in terms of their input embeddings and exclude them from the negative samples to avoid issues with false negatives. We show that these principles consistently improve the quality of the learned embeddings. The joint embeddings learned with CrossCLR extend the state of the art in video-text retrieval on Youcook2 and LSMDC datasets and in video captioning on the Youcook2 dataset by a large margin. We also demonstrate the generality of the concept by learning improved joint embeddings for other pairs of modalities.

Video activity localisation has recently attained increasing attention due to its practical values in automatically localising the most salient visual segments corresponding to their language descriptions (sentences) from untrimmed and unstructured videos. For supervised model training, a temporal annotation of both the start and end time index of each video segment for a sentence (a video moment) must be given. This is not only very expensive but also sensitive to ambiguity and subjective annotation bias, a much harder task than image labelling. In this work, we develop a more accurate weakly-supervised solution by introducing Cross-Sentence Relations Mining (CRM) in video moment proposal generation and matching when only a paragraph description of activities without per-sentence temporal annotation is available. Specifically, we explore two cross-sentence relational constraints: (1) Temporal ordering and (2) semantic consistency among sentences in a paragraph description of video activities. Existing weakly-supervised techniques only consider within-sentence video segment correlations in training without considering cross-sentence paragraph context. This can mislead due to ambiguous expressions of individual sentences with visually indiscriminate video moment proposals in isolation. Experiments on two publicly available activity localisation datasets show the advantages of our approach over the state-of-the-art weakly supervised methods, especially so when the video activity descriptions become more complex.

Recently, deep learning-based image denoising methods have achieved significant improvements over traditional methods. Due to the hardware limitation, most deep learning-based image denoising methods utilize cropped small patches to train a convolutional neural network to infer the clean images. However, the real noisy images in practical are mostly of high resolution rather than the cropped small patches and the vanilla training strategies ignore the cross-patch contextual dependency in the whole image. In this paper, we propose Cross-Patch Net (CPNet), which is the first deep- learning-based real image denoising method for HR (high resolution) input. Furthermore, we design a novel loss guided by the noise level map to obtain better performance. Compared with the vanilla patch-based training strategies, our approach effectively exploits the cross-patch contextual dependency. effective method to generate realistic sRGB noisy images from their corresponding clean sRGB images for denoiser training. Denoising experiments on real-world sRGB images show the effectiveness of the proposed method. More importantly, our method achieves state-of-the-art performance on practical sRGB noisy image denoising.

Cross-modality person re-identification is a challenging task due to large cross-modality discrepancy and intra-modality variations. Currently, most existing methods focus on learning modality-specific or modality-shareable features by using the identity supervision or modality label. Different from existing methods, this paper presents a novel Modality Confusion Learning Network (MCLNet). Its basic idea is to confuse two modalities, ensuring that the optimization is explicitly concentrated on the modality-irrelevant perspective. Specifically, MCLNet is designed to learn modality-invariant features by simultaneously minimizing inter-modality discrepancy while maximizing cross-modality similarity among instances in a single framework. Furthermore, an identity-aware marginal center aggregation strategy is introduced to extract the centralization features, while keeping diversity with a marginal constraint. Finally, we design a camera-aware learning scheme to enrich the discriminability. Extensive experiments on SYSU-MM01 and RegDB datasets show that MCLNet outperforms the state-of-the-art by a large margin. On the large-scale SYSU-MM01 dataset, our model can achieve 65.40% and 61.98% in terms of Rank-1 accuracy and mAP value.

In order to train 3D gaze estimators without too many annotations, we propose an unsupervised learning framework, Cross-Encoder, to leverage the unlabeled data to learn suitable representation for gaze estimation. To address the issue that the feature of gaze is always intertwined with the appearance of the eye, Cross-Encoder disentangles the features using a latent-code-swapping mechanism on eye-consistent image pairs and gaze-similar ones. Specifically, each image is encoded as a gaze feature and an eye feature. Cross-Encoder is trained to reconstruct each image in the eye-consistent pair according to its gaze feature and the other's eye feature, but to reconstruct each image in the gaze-similar pair according to its eye feature and the other's gaze feature. Experimental results show the validity of our work. First, using the Cross-Encoder-learned gaze representation, the gaze estimator trained with very few samples outperforms the ones using other unsupervised learning methods, under both within-dataset and cross-dataset protocol. Second, ResNet18 pretrained by Cross-Encoder is competitive with state-of-the-art gaze estimation methods. Third, ablation study shows that Cross-Encoder disentangles the gaze feature and eye feature.

Visual localization and mapping is the key technology underlying the majority of mixed reality and robotics systems. Most state-of-the-art approaches rely on local features to establish correspondences between images. In this paper, we present three novel scenarios for localization and mapping which require the continuous update of feature representations and the ability to match across different feature types. While localization and mapping is a fundamental computer vision problem, the traditional setup supposes the same local features are used throughout the evolution of a map. Thus, whenever the underlying features are changed, the whole process is repeated from scratch. However, this is typically impossible in practice, because raw images are often not stored and re-building the maps could lead to loss of the attached digital content. To overcome the limitations of current approaches, we present the first principled solution to cross-descriptor localization and mapping. Our data-driven approach is agnostic to the feature descriptor type, has low computational requirements, and scales linearly with the number of description algorithms. Extensive experiments demonstrate the effectiveness of our approach on state-of-the-art benchmarks for a variety of handcrafted and learned features.

Autonomous highlight detection is crucial for enhancing the efficiency of video browsing on social media platforms. To attain this goal in a data-driven way, one may often face the situation where highlight annotations are not available on the target video category used in practice, while the supervision on another video category (named as source video category) is achievable. In such a situation, one can derive an effective highlight detector on target video category by transferring the highlight knowledge acquired from source video category to the target one. We call this problem cross-category video highlight detection, which has been rarely studied in previous works. For tackling such practical problem, we propose a Dual-Learner-based Video Highlight Detection (DL-VHD) framework. Under this framework, we first design a Set-based Learning module (SL-module) to improve the conventional pair-based learning by assessing the highlight extent of a video segment under a broader context. Based on such learning manner, we introduce two different learners to acquire the basic distinction of target category videos and the characteristics of highlight moments on source video category, respectively. These two types of highlight knowledge are further consolidated via knowledge distillation. Extensive experiments on three benchmark datasets demonstrate the superiority of the proposed SL-module, and the DL-VHD method outperforms five typical Unsupervised Domain Adaptation (UDA) algorithms on various cross-category highlight detection tasks.

We present "Cross-Camera Convolutional Color Constancy" (C5), a learning-based method, trained on images from multiple cameras, that accurately estimates a scene's illuminant color from raw images captured by a new camera previously unseen during training. C5 is a hypernetwork-like extension of the convolutional color constancy (CCC) approach: C5 learns to generate the weights of a CCC model that is then evaluated on the input image, with the CCC weights dynamically adapted to different input content. Unlike prior cross-camera color constancy models, which are usually designed to be agnostic to the spectral properties of test-set images from unobserved cameras, C5 approaches this problem through the lens of transductive inference: additional unlabeled images are provided as input to the model at test time, which allows the model to calibrate itself to the spectral properties of the test-set camera during inference. C5 achieves state-of-the-art accuracy for cross-camera color constancy on several datasets, is fast to evaluate ( 7 and 90 ms per image on a GPU or CPU, respectively), and requires little memory ( 2 MB), and thus is a practical solution to the problem of calibration-free automatic white balance for mobile photography.

Cracks are irregular line structures that are of interest in many computer vision applications. Crack detection (e.g., from pavement images) is a challenging task due to intensity in-homogeneity, topology complexity, low contrast and noisy background. The overall crack detection accuracy can be significantly affected by the detection performance on fine-grained cracks. In this work, we propose a Crack Transformer network (CrackFormer) for fine-grained crack detection. The CrackFormer is composed of novel attention modules in a SegNet-like encoder-decoder architecture. Specifically, it consists of novel self-attention modules with 1x1 convolutional kernels for efficient contextual information extraction across feature-channels, and efficient positional embedding to capture large receptive field contextual information for long range interactions. It also introduces new scaling-attention modules to combine outputs from the corresponding encoder and decoder blocks to suppress non-semantic features and sharpen semantic cracks. The CrackFormer is trained and evaluated on three classical crack datasets. The experimental results show that CrackFormer achieves ODS values of 0.871, 0.877 and 0.881, respectively, on the three datasets and outperforms the state-of-the-art methods.

Attention mechanism has demonstrated great potential in fine-grained visual recognition tasks. In this paper, we present a counterfactual attention learning method to learn more effective attention based on causal inference. Unlike most existing methods that learn visual attention based on conventional likelihood, we propose to learn the attention with counterfactual causality, which provides a tool to measure the attention quality and a powerful supervisory signal to guide the learning process. Specifically, we analyze the effect of the learned visual attention on network prediction through counterfactual intervention and maximize the effect to encourage the network to learn more useful attention for fine-grained image recognition. Empirically, we evaluate our method on a wide range of fine-grained visual recognition tasks where attention plays a crucial role, including fine-grained image categorization, person re-identification, and vehicle re-identification. The consistent improvement on all benchmarks demonstrates the effectiveness of our method.

Shape analysis has been playing an important role in early diagnosis and prognosis of neurodegenerative diseases such as Alzheimer's diseases (AD). However, obtaining effective shape representations remains challenging. This paper proposes to use the Alexandrov polyhedra as surface-based shape signatures for cortical morphometry analysis. Given a closed genus-0 surface, its Alexandrov polyhedron is a convex representation that encodes its intrinsic geometry information. We propose to compute the polyhedra via a novel spherical optimal transport (OT) computation. In our experiments, we observe that the Alexandrov polyhedra of cortical surfaces between pathology-confirmed AD and cognitively unimpaired individuals are significantly different. Moreover, we propose a visualization method by comparing local geometry differences across cortical surfaces. We show that the proposed method is effective in pinpointing regional cortical structural changes impacted by AD.

This paper proposes a method for representation learning of multimodal data using contrastive losses. A traditional approach is to contrast different modalities to learn the information shared between them. However, that approach could fail to learn the complementary synergies between modalities that might be useful for downstream tasks. Another approach is to concatenate all the modalities into a tuple and then contrast positive and negative tuple correspondences. However, that approach could consider only the stronger modalities while ignoring the weaker ones. To address these issues, we propose a novel contrastive learning objective, TupleInfoNCE. It contrasts tuples based not only on positive and negative correspondences, but also by composing new negative tuples using modalities describing different scenes. Training with these additional negatives encourages the learning model to examine the correspondences among modalities in the same tuple, ensuring that weak modalities are not ignored. We provide a theoretical justification based on mutual-information for why this approach works, and we propose a sample optimization algorithm to generate positive and negative samples to maximize training efficacy. We find that TupleInfoNCE significantly outperforms previous state of the arts on three different downstream tasks.

Recent works have advanced the performance of self-supervised representation learning by a large margin. The core among these methods is intra-image invariance learning. Two different transformations of one image instance are considered as a positive sample pair, where various tasks are designed to learn invariant representations by comparing the pair. Analogically, for video data, representations of frames from the same video are trained to be closer than frames from other videos, i.e. intra-video invariance. However, cross-video relation has barely been explored for visual representation learning. Unlike intra-video invariance, ground-truth labels of cross-video relation is usually unavailable without human labors. In this paper, we propose a novel contrastive learning method which explores the cross-video relation by using cycle-consistency for general image representation learning. This allows to collect positive sample pairs across different video instances, which we hypothesize will lead to higher-level semantics. We validate our method by transferring our image representation to multiple downstream tasks including visual object tracking, image classification, and action recognition. We show significant improvement over state-of-the-art contrastive learning methods. Project page is available at https://happywu.github.io/cycle_contrast_video.