Computer Vision – ECCV 2020

In crowd counting datasets, the location labels are costly, yet, they are not taken into the evaluation metrics. Besides, existing multi-task approaches employ high-level tasks to improve counting accuracy. This research tendency increases the demand for more annotations. In this paper, we propose a weakly-supervised counting network, which directly regresses the crowd numbers without the location supervision. Moreover, we train the network to count by exploiting the relationship among the images. We propose a soft-label sorting network along with the counting network, which sorts the given images by their crowd numbers. The sorting network drives the shared backbone CNN model to obtain density-sensitive ability explicitly. Therefore, the proposed method improves the counting accuracy by utilizing the information hidden in crowd numbers, rather than learning from extra labels, such as locations and perspectives. We evaluate our proposed method on three crowd counting datasets, and the performance of our method plays favorably against the fully supervised state-of-the-art approaches.

Caricature attributes provide distinctive facial features to help research in Psychology and Neuroscience. However, unlike the facial photo attribute datasets that have a quantity of annotated images, the annotations of caricature attributes are rare. To facility the research in attribute learning of caricatures, we propose a caricature attribute dataset, namely WebCariA. Moreover, to utilize models that trained by face attributes, we propose a novel unsupervised domain adaptation framework for cross-modality (i.e., photos to caricatures) attribute recognition, with an integrated inter- and intra-domain consistency learning scheme. Specifically, the inter-domain consistency learning scheme consisting an image-to-image translator to first fill the domain gap between photos and caricatures by generating intermediate image samples, and a label consistency learning module to align their semantic information. The intra-domain consistency learning scheme integrates the common feature consistency learning module with a novel attribute-aware attention-consistency learning module for a more efficient alignment. We did an extensive ablation study to show the effectiveness of the proposed method. And the proposed method also outperforms the state-of-the-art methods by a margin. The implementation of the proposed method is available at https://github.com/KeleiHe/DAAN .

Object detection has witnessed significant progress by relying on large, manually annotated datasets. Annotating such datasets is highly time consuming and expensive, which motivates the development of weakly supervised and few-shot object detection methods. However, these methods largely underperform with respect to their strongly supervised counterpart, as weak training signals often result in partial or oversized detections. Towards solving this problem we introduce, for the first time, an online annotation module (OAM) that learns to generate a many-shot set of reliable annotations from a larger volume of weakly labelled images. Our OAM can be jointly trained with any fully supervised two-stage object detection method, providing additional training annotations on the fly. This results in a fully end-to-end strategy that only requires a low-shot set of fully annotated images. The integration of the OAM with Fast(er) R-CNN improves their performance by $$17\%$$ 17 % mAP, $$9\%$$ 9 % AP50 on PASCAL VOC 2007 and MS-COCO benchmarks, and significantly outperforms competing methods using mixed supervision.

When learning to sketch, beginners start with simple and flexible shapes, and then gradually strive for more complex and accurate ones in the subsequent training sessions. In this paper, we design a “shape curriculum” for learning continuous Signed Distance Function (SDF) on shapes, namely Curriculum DeepSDF. Inspired by how humans learn, Curriculum DeepSDF organizes the learning task in ascending order of difficulty according to the following two criteria: surface accuracy and sample difficulty. The former considers stringency in supervising with ground truth, while the latter regards the weights of hard training samples near complex geometry and fine structure. More specifically, Curriculum DeepSDF learns to reconstruct coarse shapes at first, and then gradually increases the accuracy and focuses more on complex local details. Experimental results show that a carefully-designed curriculum leads to significantly better shape reconstructions with the same training data, training epochs and network architecture as DeepSDF. We believe that the application of shape curricula can benefit the training process of a wide variety of 3D shape representation learning methods.

We are interested in reconstructing the mesh representation of object surfaces from point clouds. Surface reconstruction is a prerequisite for downstream applications such as rendering, collision avoidance for planning, animation, etc. However, the task is challenging if the input point cloud has a low resolution, which is common in real-world scenarios (e.g., from LiDAR or Kinect sensors). Existing learning-based mesh generative methods mostly predict the surface by first building a shape embedding that is at the whole object level, a design that causes issues in generating fine-grained details and generalizing to unseen categories. Instead, we propose to leverage the input point cloud as much as possible, by only adding connectivity information to existing points. Particularly, we predict which triplets of points should form faces. Our key innovation is a surrogate of local connectivity, calculated by comparing the intrinsic/extrinsic metrics. We learn to predict this surrogate using a deep point cloud network and then feed it to an efficient post-processing module for high-quality mesh generation. We demonstrate that our method can not only preserve details, handle ambiguous structures, but also possess strong generalizability to unseen categories by experiments on synthetic and real data.

Adversarial attack has recently become a tremendous threat to deep learning models. To improve the robustness of machine learning models, adversarial training, formulated as a minimax optimization problem, has been recognized as one of the most effective defense mechanisms. However, the non-convex and non-concave property poses a great challenge to the minimax training. In this paper, we empirically demonstrate that the commonly used PGD attack may not be optimal for inner maximization, and improved inner optimizer can lead to a more robust model. Then we leverage a learning-to-learn (L2L) framework to train an optimizer with recurrent neural networks, providing update directions and steps adaptively for the inner problem. By co-training optimizer’s parameters and model’s weights, the proposed framework consistently improves over PGD-based adversarial training and TRADES.

In this paper, we present a large-scale Diverse Real-world image Super-Resolution dataset, i.e., DRealSR, as well as a divide-and-conquer Super-Resolution (SR) network, exploring the utility of guiding SR model with low-level image components. DRealSR establishes a new SR benchmark with diverse real-world degradation processes, mitigating the limitations of conventional simulated image degradation. In general, the targets of SR vary with image regions with different low-level image components, e.g., smoothness preserving for flat regions, sharpening for edges, and detail enhancing for textures. Learning an SR model with conventional pixel-wise loss usually is easily dominated by flat regions and edges, and fails to infer realistic details of complex textures. We propose a Component Divide-and-Conquer (CDC) model and a Gradient-Weighted (GW) loss for SR. Our CDC parses an image with three components, employs three Component-Attentive Blocks (CABs) to learn attentive masks and intermediate SR predictions with an intermediate supervision learning strategy, and trains an SR model following a divide-and-conquer learning principle. Our GW loss also provides a feasible way to balance the difficulties of image components for SR. Extensive experiments validate the superior performance of our CDC and the challenging aspects of our DRealSR dataset related to diverse real-world scenarios. Our dataset and codes are publicly available at https://github.com/xiezw5/Component-Divide-and-Conquer-for-Real-World-Image-Super-Resolution .

Weakly supervised object detection (WSOD) has attracted extensive research attention due to its great flexibility of exploiting large-scale image-level annotation for detector training. Whilst deep residual networks such as ResNet and DenseNet have become the standard backbones for many computer vision tasks, the cutting-edge WSOD methods still rely on plain networks, e.g., VGG, as backbones. It is indeed not trivial to employ deep residual networks for WSOD, which even shows significant deterioration of detection accuracy and non-convergence. In this paper, we discover the intrinsic root with sophisticated analysis and propose a sequence of design principles to take full advantages of deep residual learning for WSOD from the perspectives of adding redundancy, improving robustness and aligning features. First, a redundant adaptation neck is key for effective object instance localization and discriminative feature learning. Second, small-kernel convolutions and MaxPool down-samplings help improve the robustness of information flow, which gives finer object boundaries and make the detector more sensitivity to small objects. Third, dilated convolution is essential to align the proposal features and exploit diverse local information by extracting high-resolution feature maps. Extensive experiments show that the proposed principles enable deep residual networks to establishes new state-of-the-arts on PASCAL VOC and MS COCO.

This paper presents a near-light photometric stereo method for spatially varying reflectances. Recent studies in photometric stereo proposed learning-based approaches to handle diverse real-world reflectances and achieve high accuracy compared to conventional methods. However, they assume distant (i.e., parallel) lights, which can in practical settings only be approximately realized, and they fail in near-light conditions. Near-light photometric stereo methods address near-light conditions but previous works are limited to over-simplified reflectances, such as Lambertian reflectance. The proposed method takes a hybrid approach of distant- and near-light models, where the surface normal of a small area (corresponding to a pixel) is computed locally with a distant light assumption, and the reconstruction error is assessed based on a near-light image formation model. This paper is the first work to solve unknown, spatially varying, diverse reflectances in near-light photometric stereo.

Pretraining general-purpose visual features has become a crucial part of tackling many computer vision tasks. While one can learn such features on the extensively-annotated ImageNet dataset, recent approaches have looked at ways to allow for noisy, fewer, or even no annotations to perform such pretraining. Starting from the observation that captioned images are easily crawlable, we argue that this overlooked source of information can be exploited to supervise the training of visual representations. To do so, motivated by the recent progresses in language models, we introduce image-conditioned masked language modeling (ICMLM) – a proxy task to learn visual representations over image-caption pairs. ICMLM consists in predicting masked words in captions by relying on visual cues. To tackle this task, we propose hybrid models, with dedicated visual and textual encoders, and we show that the visual representations learned as a by-product of solving this task transfer well to a variety of target tasks. Our experiments confirm that image captions can be leveraged to inject global and localized semantic information into visual representations. Project website: https://europe.naverlabs.com/ICMLM .

Long-tail recognition tackles the natural non-uniformly distributed data in real-world scenarios. While modern classifiers perform well on populated classes, its performance degrades significantly on tail classes. Humans, however, are less affected by this since, when confronted with uncertain examples, they simply opt to provide coarser predictions. Motivated by this, a deep realistic taxonomic classifier (Deep-RTC) is proposed as a new solution to the long-tail problem, combining realism with hierarchical predictions. The model has the option to reject classifying samples at different levels of the taxonomy, once it cannot guarantee the desired performance. Deep-RTC is implemented with a stochastic tree sampling during training to simulate all possible classification conditions at finer or coarser levels and a rejection mechanism at inference time. Experiments on the long-tailed version of four datasets, CIFAR100, AWA2, Imagenet, and iNaturalist, demonstrate that the proposed approach preserves more information on all classes with different popularity levels. Deep-RTC also outperforms the state-of-the-art methods in longtailed recognition, hierarchical classification, and learning with rejection literature using the proposed correctly predicted bits (CPB) metric.

This paper proposes a straightforward, intuitive deep learning approach for (biomedical) image segmentation tasks. Different from the existing dense pixel classification methods, we develop a novel multi-level aggregation network to directly regress the coordinates of the boundary of instances in an end-to-end manner. The network seamlessly combines standard convolution neural network (CNN) with Attention Refinement Module (ARM) and Graph Convolution Network (GCN). By iteratively and hierarchically fusing the features across different layers of the CNN, our approach gains sufficient semantic information from the input image and pays special attention to the local boundaries with the help of ARM and GCN. In particular, thanks to the proposed aggregation GCN, our network benefits from direct feature learning of the instances’ boundary locations and the spatial information propagation across the image. Experiments on several challenging datasets demonstrate that our method achieves comparable results with state-of-the-art approaches but requires less inference time on the segmentation of fetal head in ultrasound images and of optic disc and optic cup in color fundus images.

This paper presents a new task named weakly-supervised group activity recognition (GAR) which differs from conventional GAR tasks in that only video-level labels are available, yet the important persons within each frame are not provided even in the training data. This eases us to collect and annotate a large-scale NBA dataset and thus raise new challenges to GAR. To mine useful information from weak supervision, we present a key insight that key instances are likely to be related to each other, and thus design a social adaptive module (SAM) to reason about key persons and frames from noisy data. Experiments show significant improvement on the NBA dataset as well as the popular volleyball dataset. In particular, our model trained on video-level annotation achieves comparable accuracy to prior algorithms which required strong labels.

We present an effective method to progressively integrate and refine the cross-modality complementarities for RGB-D salient object detection (SOD). The proposed network mainly solves two challenging issues: 1) how to effectively integrate the complementary information from RGB image and its corresponding depth map, and 2) how to adaptively select more saliency-related features. First, we propose a cross-modality feature modulation (cmFM) module to enhance feature representations by taking the depth features as prior, which models the complementary relations of RGB-D data. Second, we propose an adaptive feature selection (AFS) module to select saliency-related features and suppress the inferior ones. The AFS module exploits multi-modality spatial feature fusion with the self-modality and cross-modality interdependencies of channel features are considered. Third, we employ a saliency-guided position-edge attention (sg-PEA) module to encourage our network to focus more on saliency-related regions. The above modules as a whole, called cmMS block, facilitates the refinement of saliency features in a coarse-to-fine fashion. Coupled with a bottom-up inference, the refined saliency features enable accurate and edge-preserving SOD. Extensive experiments demonstrate that our network outperforms state-of-the-art saliency detectors on six popular RGB-D SOD benchmarks.

Image generation from scene description is a cornerstone technique for the controlled generation, which is beneficial to applications such as content creation and image editing. In this work, we aim to synthesize images from scene description with retrieved patches as reference. We propose a differentiable retrieval module. With the differentiable retrieval module, we can (1) make the entire pipeline end-to-end trainable, enabling the learning of better feature embedding for retrieval; (2) encourage the selection of mutually compatible patches with additional objective functions. We conduct extensive quantitative and qualitative experiments to demonstrate that the proposed method can generate realistic and diverse images, where the retrieved patches are reasonable and mutually compatible.

In this paper, we propose a general and efficient pre-training paradigm, Montage pre-training, for object detection. Montage pre-training needs only the target detection dataset while taking only 1/4 computational resources compared to the widely adopted ImageNet pre-training. To build such an efficient paradigm, we reduce the potential redundancy by carefully extracting useful samples from the original images, assembling samples in a Montage manner as input, and using an ERF-adaptive dense classification strategy for model pre-training. These designs include not only a new input pattern to improve the spatial utilization but also a novel learning objective to expand the effective receptive field of the pre-trained model. The efficiency and effectiveness of Montage pre-training are validated by extensive experiments on the MS-COCO dataset, where the results indicate that the models using Montage pre-training are able to achieve on-par or even better detection performances compared with the ImageNet pre-training.

Fast person re-identification (ReID) aims to search person images quickly and accurately. The main idea of recent fast ReID methods is the hashing algorithm, which learns compact binary codes and performs fast Hamming distance and counting sort. However, a very long code is needed for high accuracy (e.g. 2048), which compromises search speed. In this work, we introduce a new solution for fast ReID by formulating a novel Coarse-to-Fine (CtF) hashing code search strategy, which complementarily uses short and long codes, achieving both faster speed and better accuracy. It uses shorter codes to coarsely rank broad matching similarities and longer codes to refine only a few top candidates for more accurate instance ReID. Specifically, we design an All-in-One (AiO) framework together with a Distance Threshold Optimization (DTO) algorithm. In AiO, we simultaneously learn and enhance multiple codes of different lengths in a single model. It learns multiple codes in a pyramid structure, and encourage shorter codes to mimic longer codes by self-distillation. DTO solves a complex threshold search problem by a simple optimization process, and the balance between accuracy and speed is easily controlled by a single parameter. It formulates the optimization target as a $$F_{\beta }$$ F β score that can be optimised by Gaussian cumulative distribution functions. Experimental results on 2 datasets show that our proposed method (CtF) is not only $$8\%$$ 8 % more accurate but also $$5\times $$ 5 × faster than contemporary hashing ReID methods. Compared with non-hashing ReID methods, CtF is $$50\times $$ 50 × faster with comparable accuracy. Code is available at https://github.com/wangguanan/light-reid .

The JPEG image compression algorithm is the most popular method of image compression because of it’s ability for large compression ratios. However, to achieve such high compression, information is lost. For aggressive quantization settings, this leads to a noticeable reduction in image quality. Artifact correction has been studied in the context of deep neural networks for some time, but the current methods delivering state-of-the-art results require a different model to be trained for each quality setting, greatly limiting their practical application. We solve this problem by creating a novel architecture which is parameterized by the JPEG file’s quantization matrix. This allows our single model to achieve state-of-the-art performance over models trained for specific quality settings. ...

A transmission matrix (TM) describes the linear relationship between input and output phasor fields when a coherent wave passes through a scattering medium. Measurement of the TM enables numerous applications, but is challenging and time-intensive for an arbitrary medium. State-of-the-art methods, including phase-shifting holography and double phase retrieval, require significant amounts of measurements, and post-capture reconstruction that is often computationally intensive. In this paper, we propose 3PointTM, an approach for sensing TMs that uses a minimal number of measurements per pixel—reducing the measurement budget by a factor of two as compared to state of the art in phase-shifting holography for measuring TMs—and has a low computational complexity as compared to phase retrieval. We validate our approach on real and simulated data, and show successful focusing of light and image reconstruction on dense scattering media.

Photorealistic style transfer is the task of transferring the artistic style of an image onto a content target, producing a result that is plausibly taken with a camera. Recent approaches, based on deep neural networks, produce impressive results but are either too slow to run at practical resolutions, or still contain objectionable artifacts. We propose a new end-to-end model for photorealistic style transfer that is both fast and inherently generates photorealistic results. The core of our approach is a feed-forward neural network that learns local edge-aware affine transforms that automatically obey the photorealism constraint. When trained on a diverse set of images and a variety of styles, our model can robustly apply style transfer to an arbitrary pair of input images. Compared to the state of the art, our method produces visually superior results and is three orders of magnitude faster, enabling real-time performance at 4K on a mobile phone. We validate our method with ablation and user studies.

Recently, deep learning based 3D face reconstruction methods have shown promising results in both quality and efficiency. However, most of their training data is constructed by 3D Morphable Model, whose space spanned is only a small part of the shape space. As a result, the reconstruction results lose the fine-grained geometry and look different from real faces. To alleviate this issue, we first propose a solution to construct large-scale fine-grained 3D data from RGB-D images, which are expected to be massively collected as the proceeding of hand-held depth camera. A new dataset Fine-Grained 3D face (FG3D) with 200k samples is constructed to provide sufficient data for neural network training. Secondly, we propose a Fine-Grained reconstruction Network (FGNet) that can concentrate on shape modification by warping the network input and output to the UV space. Through FG3D and FGNet, we successfully generate reconstruction results with fine-grained geometry. The experiments on several benchmarks validate the effectiveness of our method compared to several baselines and other state-of-the-art methods. The proposed method and code will be available at https://github.com/XiangyuZhu-open/Beyond3DMM .

Video-to-video synthesis (vid2vid) aims for converting high-level semantic inputs to photorealistic videos. While existing vid2vid methods can achieve short-term temporal consistency, they fail to ensure the long-term one. This is because they lack knowledge of the 3D world being rendered and generate each frame only based on the past few frames. To address the limitation, we introduce a novel vid2vid framework that efficiently and effectively utilizes all past generated frames during rendering. This is achieved by condensing the 3D world rendered so far into a physically-grounded estimate of the current frame, which we call the guidance image. We further propose a novel neural network architecture to take advantage of the information stored in the guidance images. Extensive experimental results on several challenging datasets verify the effectiveness of our approach in achieving world consistency—the output video is consistent within the entire rendered 3D world.

Partially supervised instance segmentation aims to perform learning on limited mask-annotated categories of data thus eliminating expensive and exhaustive mask annotation. The learned models are expected to be generalizable to novel categories. Existing methods either learn a transfer function from detection to segmentation, or cluster shape priors for segmenting novel categories. We propose to learn the underlying class-agnostic commonalities that can be generalized from mask-annotated categories to novel categories. Specifically, we parse two types of commonalities: 1) shape commonalities which are learned by performing supervised learning on instance boundary prediction; and 2) appearance commonalities which are captured by modeling pairwise affinities among pixels of feature maps to optimize the separability between instance and the background. Incorporating both the shape and appearance commonalities, our model significantly outperforms the state-of-the-art methods on both partially supervised setting and few-shot setting for instance segmentation on COCO dataset. The code is available at https://github.com/fanq15/FewX .

The semantic segmentation of parts of objects in the wild is a challenging task in which multiple instances of objects and multiple parts within those objects must be detected in the scene. This problem remains nowadays very marginally explored, despite its fundamental importance towards detailed object understanding. In this work, we propose a novel framework combining higher object-level context conditioning and part-level spatial relationships to address the task. To tackle object-level ambiguity, a class-conditioning module is introduced to retain class-level semantics when learning parts-level semantics. In this way, mid-level features carry also this information prior to the decoding stage. To tackle part-level ambiguity and localization we propose a novel adjacency graph-based module that aims at matching the relative spatial relationships between ground truth and predicted parts. The experimental evaluation on the Pascal-Part dataset shows that we achieve state-of-the-art results on this task.

Event cameras are bio-inspired sensors that respond to per-pixel brightness changes in the form of asynchronous and sparse “events”. Recently, pattern recognition algorithms, such as learning-based methods, have made significant progress with event cameras by converting events into synchronous dense, image-like representations and applying traditional machine learning methods developed for standard cameras. However, these approaches discard the spatial and temporal sparsity inherent in event data at the cost of higher computational complexity and latency. In this work, we present a general framework for converting models trained on synchronous image-like event representations into asynchronous models with identical output, thus directly leveraging the intrinsic asynchronous and sparse nature of the event data. We show both theoretically and experimentally that this drastically reduces the computational complexity and latency of high-capacity, synchronous neural networks without sacrificing accuracy. In addition, our framework has several desirable characteristics: (i) it exploits spatio-temporal sparsity of events explicitly, (ii) it is agnostic to the event representation, network architecture, and task, and (iii) it does not require any train-time change, since it is compatible with the standard neural networks’ training process. We thoroughly validate the proposed framework on two computer vision tasks: object detection and object recognition. In these tasks, we reduce the computational complexity up to 20 times with respect to high-latency neural networks. At the same time, we outperform state-of-the-art asynchronous approaches up to $$24\%$$ 24 % in prediction accuracy.

We introduce a novel end-to-end approach to predict a 3D room layout from a single panoramic image. Compared to recent state-of-the-art works, our method is not limited to Manhattan World environments, and can reconstruct rooms bounded by vertical walls that do not form right angles or are curved – i.e., Atlanta World models. In our approach, we project the original gravity-aligned panoramic image on two horizontal planes, one above and one below the camera. This representation encodes all the information needed to recover the Atlanta World 3D bounding surfaces of the room in the form of a 2D room footprint on the floor plan and a room height. To predict the 3D layout, we propose an encoder-decoder neural network architecture, leveraging Recurrent Neural Networks (RNNs) to capture long-range geometric patterns, and exploiting a customized training strategy based on domain-specific knowledge. The experimental results demonstrate that our method outperforms state-of-the-art solutions in prediction accuracy, in particular in cases of complex wall layouts or curved wall footprints.

Convolutional operations have two limitations: (1) do not explicitly model where to focus as the same filter is applied to all the positions, and (2) are unsuitable for modeling long-range dependencies as they only operate on a small neighborhood. While both limitations can be alleviated by attention operations, many design choices remain to be determined to use attention, especially when applying attention to videos. Towards a principled way of applying attention to videos, we address the task of spatiotemporal attention cell search. We propose a novel search space for spatiotemporal attention cells, which allows the search algorithm to flexibly explore various design choices in the cell. The discovered attention cells can be seamlessly inserted into existing backbone networks, e.g., I3D or S3D, and improve video classification accuracy by more than 2% on both Kinetics-600 and MiT datasets. The discovered attention cells outperform non-local blocks on both datasets, and demonstrate strong generalization across different modalities, backbones, and datasets. Inserting our attention cells into I3D-R50 yields state-of-the-art performance on both datasets.

People learn throughout life. However, incrementally updating conventional neural networks leads to catastrophic forgetting. A common remedy is replay, which is inspired by how the brain consolidates memory. Replay involves fine-tuning a network on a mixture of new and old instances. While there is neuroscientific evidence that the brain replays compressed memories, existing methods for convolutional networks replay raw images. Here, we propose REMIND, a brain-inspired approach that enables efficient replay with compressed representations. REMIND is trained in an online manner, meaning it learns one example at a time, which is closer to how humans learn. Under the same constraints, REMIND outperforms other methods for incremental class learning on the ImageNet ILSVRC-2012 dataset. We probe REMIND’s robustness to data ordering schemes known to induce catastrophic forgetting. We demonstrate REMIND’s generality by pioneering online learning for Visual Question Answering (VQA) ( https://github.com/tyler-hayes/REMIND ).

State-of-the-art image classifiers are trained and tested using well-illuminated images. These images are typically captured by CMOS image sensors with at least tens of photons per pixel. However, in dark environments when the photon flux is low, image classification becomes difficult because the measured signal is suppressed by noise. In this paper, we present a new low-light image classification solution using Quanta Image Sensors (QIS). QIS are a new type of image sensors that possess photon-counting ability without compromising on pixel size and spatial resolution. Numerous studies over the past decade have demonstrated the feasibility of QIS for low-light imaging, but their usage for image classification has not been studied. This paper fills the gap by presenting a student-teacher learning scheme which allows us to classify the noisy QIS raw data. We show that with student-teacher learning, we can achieve image classification at a photon level of one photon per pixel or lower. Experimental results verify the effectiveness of the proposed method compared to existing solutions.

Benefiting from deep learning research and large-scale datasets, saliency prediction has achieved significant success in the past decade. However, it still remains challenging to predict saliency maps on images in new domains that lack sufficient data for data-hungry models. To solve this problem, we propose a few-shot transfer learning paradigm for saliency prediction, which enables efficient transfer of knowledge learned from the existing large-scale saliency datasets to a target domain with limited labeled samples. Specifically, few target domain samples are used as the reference to train a model with a source domain dataset such that the training process can converge to a local minimum in favor of the target domain. Then, the learned model is further fine-tuned with the reference. The proposed framework is gradient-based and model-agnostic. We conduct comprehensive experiments and ablation study on various source domain and target domain pairs. The results show that the proposed framework achieves a significant performance improvement. The code is publicly available at https://github.com/luoyan407/n-reference .

In this paper, we aim to develop an efficient and compact deep network for RGB-D salient object detection, where the depth image provides complementary information to boost performance in complex scenarios. Starting from a coarse initial prediction by a multi-scale residual block, we propose a progressively guided alternate refinement network to refine it. Instead of using ImageNet pre-trained backbone network, we first construct a lightweight depth stream by learning from scratch, which can extract complementary features more efficiently with less redundancy. Then, different from the existing fusion based methods, RGB and depth features are fed into proposed guided residual (GR) blocks alternately to reduce their mutual degradation. By assigning progressive guidance in the stacked GR blocks within each side-output, the false detection and missing parts can be well remedied. Extensive experiments on seven benchmark datasets demonstrate that our model outperforms existing state-of-the-art approaches by a large margin, and also shows superiority in efficiency (71 FPS) and model size (64.9 MB).

Recently, temporal action localization (TAL), i.e., finding specific action segments in untrimmed videos, has attracted increasing attentions of the computer vision community. State-of-the-art solutions for TAL involves evaluating the frame-level probabilities of three action-indicating phases, i.e. starting, continuing, and ending; and then post-processing these predictions for the final localization. This paper delves deep into this mechanism, and argues that existing methods, by modeling these phases as individual classification tasks, ignored the potential temporal constraints between them. This can lead to incorrect and/or inconsistent predictions when some frames of the video input lack sufficient discriminative information. To alleviate this problem, we introduce two regularization terms to mutually regularize the learning procedure: the Intra-phase Consistency (IntraC) regularization is proposed to make the predictions verified inside each phase; and the Inter-phase Consistency (InterC) regularization is proposed to keep consistency between these phases. Jointly optimizing these two terms, the entire framework is aware of these potential constraints during an end-to-end optimization process. Experiments are performed on two popular TAL datasets, THUMOS14 and ActivityNet1.3. Our approach clearly outperforms the baseline both quantitatively and qualitatively. The proposed regularization also generalizes to other TAL methods (e.g., TSA-Net and PGCN). Code: https://github.com/PeisenZhao/Bottom-Up-TAL-with-MR .

Inspired by optimization techniques, we propose a novel meta-learning algorithm with gradient modulation to encourage fast-adaptation of neural networks in the absence of abundant data. Our method, termed ModGrad, is designed to circumvent the noisy nature of the gradients which is prevalent in low-data regimes. Furthermore and having the scalability concern in mind, we formulate ModGrad via low-rank approximations, which in turn enables us to employ ModGrad to adapt hefty neural networks. We thoroughly assess and contrast ModGrad against a large family of meta-learning techniques and observe that the proposed algorithm outperforms baselines comfortably while enjoying faster convergence.

Style transfer generates an image whose content comes from one image and style from the other. Image-to-image translation approaches with disentangled representations have been shown effective for style transfer between two image categories. However, previous methods often assume a shared domain-invariant content space, which could compromise the content representation power. For addressing this issue, this paper leverages domain-specific mappings for remapping latent features in the shared content space to domain-specific content spaces. This way, images can be encoded more properly for style transfer. Experiments show that the proposed method outperforms previous style transfer methods, particularly on challenging scenarios that would require semantic correspondences between images. Code and results are available at https://github.com/acht7111020/DSMAP .

Visual similarity plays an important role in many computer vision applications. Deep metric learning (DML) is a powerful framework for learning such similarities which not only generalize from training data to identically distributed test distributions, but in particular also translate to unknown test classes. However, its prevailing learning paradigm is class-discriminative supervised training, which typically results in representations specialized in separating training classes. For effective generalization, however, such an image representation needs to capture a diverse range of data characteristics. To this end, we propose and study multiple complementary learning tasks, targeting conceptually different data relationships by only resorting to the available training samples and labels of a standard DML setting. Through simultaneous optimization of our tasks we learn a single model to aggregate their training signals, resulting in strong generalization and state-of-the-art performance on multiple established DML benchmark datasets.

Network pruning has been the driving force for the acceleration of neural networks and the alleviation of model storage/transmission burden. With the advent of AutoML and neural architecture search (NAS), pruning has become topical with automatic mechanism and searching based architecture optimization. Yet, current automatic designs rely on either reinforcement learning or evolutionary algorithm. Due to the non-differentiability of those algorithms, the pruning algorithm needs a long searching stage before reaching the convergence.To circumvent this problem, this paper introduces a differentiable pruning method via hypernetworks for automatic network pruning. The specifically designed hypernetworks take latent vectors as input and generate the weight parameters of the backbone network. The latent vectors control the output channels of the convolutional layers in the backbone network and act as a handle for the pruning of the layers. By enforcing $$\ell _1$$ ℓ 1 sparsity regularization to the latent vectors and utilizing proximal gradient solver, sparse latent vectors can be obtained. Passing the sparsified latent vectors through the hypernetworks, the corresponding slices of the generated weight parameters can be removed, achieving the effect of network pruning. The latent vectors of all the layers are pruned together, resulting in an automatic layer configuration. Extensive experiments are conducted on various networks for image classification, single image super-resolution, and denoising. And the experimental results validate the proposed method. Code is available at https://github.com/ofsoundof/dhp .

Network quantization is an effective method for network compression. Existing methods train a low-precision network by fine-tuning from a pre-trained model. However, training a low-precision network often requires large-scale labeled data to achieve superior performance. In many real-world scenarios, only limited labeled data are available due to expensive labeling costs or privacy protection. With limited training data, fine-tuning methods may suffer from the overfitting issue and substantial accuracy loss. To alleviate these issues, we introduce transfer learning into network quantization to obtain an accurate low-precision model. Specifically, we propose a method named deep transferring quantization (DTQ) to effectively exploit the knowledge in a pre-trained full-precision model. To this end, we propose a learnable attentive transfer module to identify the informative channels for alignment. In addition, we introduce the Kullback-Leibler (KL) divergence to further help train a low-precision model. Extensive experiments on both image classification and face recognition demonstrate the effectiveness of DTQ.

The trained person re-identification systems fundamentally need to be deployed on different target environments. Learning the cross-domain model has great potential for the scalability of real-world applications. In this paper, we propose a deep credible metric learning (DCML) method for unsupervised domain adaptation person re-identification. Unlike existing methods that directly finetune the model in the target domain with pseudo labels generated by the source pre-trained model, our DCML method adaptively mines credible samples for training to avoid the misleading from noise labels. Specifically, we design two credibility metrics for sample mining including the k-Nearest Neighbor similarity for density evaluation and the prototype similarity for centrality evaluation. As the increasing of the pseudo label credibility, we progressively adjust the sampling strategy in the training process. In addition, we propose an instance margin spreading loss to further increase instance-wise discrimination. Experimental results demonstrate that our DCML method explores credible and valuable training data and improves the performance of unsupervised domain adaptation.

Video-based person re-identification aims to match pedestrians with the consecutive video sequences. While a rich line of work focuses solely on extracting the motion features from pedestrian videos, we show in this paper that the temporal coherence plays a more critical role. To distill the temporal coherence part of video representation from frame representations, we propose a simple yet effective Adversarial Feature Augmentation (AFA) method, which highlights the temporal coherence features by introducing adversarial augmented temporal motion noise. Specifically, we disentangle the video representation into the temporal coherence and motion parts and randomly change the scale of the temporal motion features as the adversarial noise. The proposed AFA method is a general lightweight component that can be readily incorporated into various methods with negligible cost. We conduct extensive experiments on three challenging datasets including MARS, iLIDS-VID, and DukeMTMC-VideoReID, and the experimental results verify our argument and demonstrate the effectiveness of the proposed method.

Arbitrary-oriented object detection has recently attracted increasing attention in vision for their importance in aerial imagery, scene text, and face etc. In this paper, we show that existing regression-based rotation detectors suffer the problem of discontinuous boundaries, which is directly caused by angular periodicity or corner ordering. By a careful study, we find the root cause is that the ideal predictions are beyond the defined range. We design a new rotation detection baseline, to address the boundary problem by transforming angular prediction from a regression problem to a classification task with little accuracy loss, whereby high-precision angle classification is devised in contrast to previous works using coarse-granularity in rotation detection. We also propose a circular smooth label (CSL) technique to handle the periodicity of the angle and increase the error tolerance to adjacent angles. We further introduce four window functions in CSL and explore the effect of different window radius sizes on detection performance. Extensive experiments and visual analysis on two large-scale public datasets for aerial images i.e. DOTA, HRSC2016, as well as scene text dataset ICDAR2015 and MLT, show the effectiveness of our approach. The code is public available at https://github.com/Thinklab-SJTU/CSL_RetinaNet_Tensorflow .

Event-based sensors, which have a response if the change of pixel intensity exceeds a triggering threshold, can capture high-speed motion with microsecond accuracy. Assisted by an event camera, we can generate high frame-rate sharp videos from low frame-rate blurry ones captured by an intensity camera. In this paper, we propose an effective event-driven video deblurring and interpolation algorithm based on deep convolutional neural networks (CNNs). Motivated by the physical model that the residuals between a blurry image and sharp frames are the integrals of events, the proposed network uses events to estimate the residuals for the sharp frame restoration. As the triggering threshold varies spatially, we develop an effective method to estimate dynamic filters to solve this problem. To utilize the temporal information, the sharp frames restored from the previous blurry frame are also considered. The proposed algorithm achieves superior performance against state-of-the-art methods on both synthetic and real datasets.

This paper tackles a 2D architecture vectorization problem, whose task is to infer an outdoor building architecture as a 2D planar graph from a single RGB image. We provide a new benchmark with ground-truth annotations for 2,001 complex buildings across the cities of Atlanta, Paris, and Las Vegas. We also propose a novel algorithm utilizing 1) convolutional neural networks (CNNs) that detects geometric primitives and infers their relationships and 2) an integer programming (IP) that assembles the information into a 2D planar graph. While being a trivial task for human vision, the inference of a graph structure with an arbitrary topology is still an open problem for computer vision. Qualitative and quantitative evaluations demonstrate that our algorithm makes significant improvements over the current state-of-the-art, towards an intelligent system at the level of human perception. We will share code and data.

Transferring knowledges learned from multiple source domains to target domain is a more practical and challenging task than conventional single-source domain adaptation. Furthermore, the increase of modalities brings more difficulty in aligning feature distributions among multiple domains. To mitigate these problems, we propose a Learning to Combine for Multi-Source Domain Adaptation (LtC-MSDA) framework via exploring interactions among domains. In the nutshell, a knowledge graph is constructed on the prototypes of various domains to realize the information propagation among semantically adjacent representations. On such basis, a graph model is learned to predict query samples under the guidance of correlated prototypes. In addition, we design a Relation Alignment Loss (RAL) to facilitate the consistency of categories’ relational interdependency and the compactness of features, which boosts features’ intra-class invariance and inter-class separability. Comprehensive results on public benchmark datasets demonstrate that our approach outperforms existing methods with a remarkable margin. Our code is available at https://github.com/ChrisAllenMing/LtC-MSDA .

Unsupervised domain adaptation without consuming annotation process for unlabeled target data attracts appealing interests in semantic segmentation. However, 1) existing methods neglect that not all semantic representations across domains are transferable, which cripples domain-wise transfer with untransferable knowledge; 2) they fail to narrow category-wise distribution shift due to category-agnostic feature alignment. To address above challenges, we develop a new Critical Semantic-Consistent Learning (CSCL) model, which mitigates the discrepancy of both domain-wise and category-wise distributions. Specifically, a critical transfer based adversarial framework is designed to highlight transferable domain-wise knowledge while neglecting untransferable knowledge. Transferability-critic guides transferability-quantizer to maximize positive transfer gain under reinforcement learning manner, although negative transfer of untransferable knowledge occurs. Meanwhile, with the help of confidence-guided pseudo labels generator of target samples, a symmetric soft divergence loss is presented to explore inter-class relationships and facilitate category-wise distribution alignment. Experiments on several datasets demonstrate the superiority of our model.

Few-shot segmentation is challenging because objects within the support and query images could significantly differ in appearance and pose. Using a single prototype acquired directly from the support image to segment the query image causes semantic ambiguity. In this paper, we propose prototype mixture models (PMMs), which correlate diverse image regions with multiple prototypes to enforce the prototype-based semantic representation. Estimated by an Expectation-Maximization algorithm, PMMs incorporate rich channel-wised and spatial semantics from limited support images. Utilized as representations as well as classifiers, PMMs fully leverage the semantics to activate objects in the query image while depressing background regions in a duplex manner. Extensive experiments on Pascal VOC and MS-COCO datasets show that PMMs significantly improve upon state-of-the-arts. Particularly, PMMs improve 5-shot segmentation performance on MS-COCO by up to 5.82% with only a moderate cost for model size and inference speed (Code is available at github.com/Yang-Bob/PMMs .).

This paper focuses on webly supervised learning (WSL), where datasets are built by crawling samples from the Internet and directly using search queries as web labels. Although WSL benefits from fast and low-cost data collection, noises in web labels hinder better performance of the image classification model. To alleviate this problem, in recent works, self-label supervised loss $$\mathcal {L}_s$$ L s is utilized together with webly supervised loss $$\mathcal {L}_w$$ L w . $$\mathcal {L}_s$$ L s relies on pseudo labels predicted by the model itself. Since the correctness of the web label or pseudo label is usually on a case-by-case basis for each web sample, it is desirable to adjust the balance between $$\mathcal {L}_s$$ L s and $$\mathcal {L}_w$$ L w on sample level. Inspired by the ability of Deep Neural Networks (DNNs) in confidence prediction, we introduce Self-Contained Confidence (SCC) by adapting model uncertainty for WSL setting, and use it to sample-wisely balance $$\mathcal {L}_s$$ L s and $$\mathcal {L}_w$$ L w . Therefore, a simple yet effective WSL framework is proposed. A series of SCC-friendly regularization approaches are investigated, among which the proposed graph-enhanced mixup is the most effective method to provide high-quality confidence to enhance our framework. The proposed WSL framework has achieved the state-of-the-art results on two large-scale WSL datasets, WebVision-1000 and Food101-N. Code is available at https://github.com/bigvideoresearch/SCC .