ICCV 2021 论文列表

Video semantic segmentation is an essential task for the analysis and understanding of videos. Recent efforts largely focus on supervised video segmentation by learning from fully annotated data, but the learnt models often experience clear performance drop while applied to videos of a different domain. This paper presents DA-VSN, a domain adaptive video segmentation network that addresses domain gaps in videos by temporal consistency regularization (TCR) for consecutive frames of target-domain videos. DA-VSN consists of two novel and complementary designs. The first is cross-domain TCR that guides the prediction of target frames to have similar temporal consistency as that of source frames (learnt from annotated source data) via adversarial learning. The second is intra-domain TCR that guides unconfident predictions of target frames to have similar temporal consistency as confident predictions of target frames. Extensive experiments demonstrate the superiority of our proposed domain adaptive video segmentation network which outperforms multiple baselines consistently by large margins.

Domain adaptation for semantic segmentation aims to improve the model performance in the presence of a distribution shift between source and target domain. Leveraging the supervision from auxiliary tasks (such as depth estimation) has the potential to heal this shift because many visual tasks are closely related to each other. However, such a supervision is not always available. In this work, we leverage the guidance from self-supervised depth estimation, which is available on both domains, to bridge the domain gap. On the one hand, we propose to explicitly learn the task feature correlation to strengthen the target semantic predictions with the help of target depth estimation. On the other hand, we use the depth prediction discrepancy from source and target depth decoders to approximate the pixel-wise adaptation difficulty. The adaptation difficulty, inferred from depth, is then used to refine the target semantic segmentation pseudo-labels. The proposed method can be easily implemented into existing segmentation frameworks. We demonstrate the effectiveness of our approach on the benchmark tasks SYNTHIA-to-Cityscapes and GTA-to-Cityscapes, on which we achieve the new state-of-the-art performance of 55.0% and 56.6%, respectively. Our code is available at https://qin.ee/corda

We present DocFormer - a multi-modal transformer based architecture for the task of Visual Document Understanding (VDU). VDU is a challenging problem which aims to understand documents in their varied formats(forms, receipts etc.) and layouts. In addition, DocFormer is pre-trained in an unsupervised fashion using carefully designed tasks which encourage multi-modal interaction. DocFormer uses text, vision and spatial features and combines them using a novel multi-modal self-attention layer. DocFormer also shares learned spatial embeddings across modalities which makes it easy for the model to correlate text to visual tokens and vice versa. DocFormer is evaluated on 4 different datasets each with strong baselines. DocFormer achieves state-of-the-art results on all of them, sometimes beating models 4x its size (in no. of parameters)

Vision models notoriously flicker when applied to videos: they correctly recognize objects in some frames, but fail on perceptually similar, nearby frames. In this work, we systematically analyze the robustness of image classifiers to such temporal perturbations in videos. To do so, we construct two new datasets, ImageNet-Vid-Robust and YTBB-Robust, containing a total of 57,897 images grouped into 3,139 sets of perceptually similar images. Our datasets were derived from ImageNet-Vid and YouTube-BB, respectively, and thoroughly re-annotated by human experts for image similarity. We evaluate a diverse array of classifiers pre-trained on ImageNet and show a median classification accuracy drop of 16 and 10 points, respectively, on our two datasets. Additionally, we evaluate three detection models and show that natural perturbations induce both classification as well as localization errors, leading to a median drop in detection mAP of 14 points. Our analysis demonstrates that perturbations occurring naturally in videos pose a substantial and realistic challenge to deploying convolutional neural networks in environments that require both reliable and low-latency predictions.

Recent studies have shown that many deep metric learning loss functions perform very similarly under the same experimental conditions. One potential reason for this unexpected result is that all losses let the network focus on similar image regions or properties. In this paper, we investigate this by conducting a two-step analysis to extract and compare the learned visual features of the same model architecture trained with different loss functions: First, we compare the learned features on the pixel level by correlating saliency maps of the same input images. Second, we compare the clustering of embeddings for several image properties, e.g. object color or illumination. To provide independent control over these properties, photo-realistic 3D car renders similar to images in the Cars196 dataset are generated. In our analysis, we compare 14 pretrained models from a recent study and find that, even though all models perform similarly, different loss functions can guide the model to learn different features. We especially find differences between classification and ranking based losses. Our analysis also shows that some seemingly irrelevant properties can have significant influence on the resulting embedding. We encourage researchers from the deep metric learning community to use our methods to get insights into the features learned by their proposed methods.

We present a novel approach for estimating depth from a monocular camera as it moves through complex and crowded indoor environments, e.g., a department store or a metro station. Our approach predicts absolute scale depth maps over the entire scene consisting of a static background and multiple moving people, by training on dynamic scenes. Since it is difficult to collect dense depth maps from crowded indoor environments, we design our training framework without requiring groundtruth depths produced from depth sensing devices. Our network leverages RGB images and sparse depth maps generated from traditional 3D reconstruction methods to estimate dense depth maps. We use two constraints to handle depth for non-rigidly moving people without tracking their motion explicitly. We demonstrate that our approach offers consistent improvements over recent depth estimation methods on the NAVERLABS dataset, which includes complex and crowded scenes.

Reconstruction-based methods play an important role in unsupervised anomaly detection in images. Ideally, we expect a perfect reconstruction for normal samples and poor reconstruction for abnormal samples. Since the generalizability of deep neural networks is difficult to control, existing models such as autoencoder do not work well. In this work, we interpret the reconstruction of an image as a divide-and-assemble procedure. Surprisingly, by varying the granularity of division on feature maps, we are able to modulate the reconstruction capability of the model for both normal and abnormal samples. That is, finer granularity leads to better reconstruction, while coarser granularity leads to poorer reconstruction. With proper granularity, the gap between the reconstruction error of normal and abnormal samples can be maximized. The divide-and-assemble framework is implemented by embedding a novel multi-scale block-wise memory module into an autoencoder network. Besides, we introduce adversarial learning and explore the semantic latent representation of the discriminator, which improves the detection of subtle anomaly. We achieve state-of-the-art performance on the challenging MVTec AD dataset. Remarkably, we improve the vanilla autoencoder model by 10.1% in terms of the AUROC score.

Self-supervised learning holds promise in leveraging large amounts of unlabeled data, however much of its progress has thus far been limited to highly curated pre-training data such as ImageNet. We explore the effects of contrastive learning from larger, less-curated image datasets such as YFCC, and find there is indeed a large difference in the resulting representation quality. We hypothesize that this curation gap is due to a shift in the distribution of image classes---which is more diverse and heavy-tailed---resulting in less relevant negative samples to learn from. We test this hypothesis with a new approach, Divide and Contrast (DnC), which alternates between contrastive learning and clustering-based hard negative mining. When pretrained on less curated datasets, DnC greatly improves the performance of self-supervised learning on downstream tasks, while remaining competitive with the current state-of-the-art on curated datasets.

Single-frame temporal action localization (STAL) aims to localize actions in untrimmed videos with only one timestamp annotation for each action instance. Existing methods adopt the one-stage framework but couple the counting goal and the localization goal. This paper proposes a novel two-stage framework for the STAL task with the spirit of divide and conquer. The instance counting stage leverages the location supervision to determine the number of action instances and divide a whole video into multiple video clips, so that each video clip contains only one complete action instance; and the location estimation stage leverages the category supervision to localize the action instance in each video clip. To efficiently represent the action instance in each video clip, we introduce the proposal-based representation, and design a novel differentiable mask generator to enable the end-to-end training supervised by category labels. On THUMOS14, GTEA, and BEOID datasets, our method outperforms state-of-the-art methods by 3.5%, 2.7%, 4.8% mAP on average. And extensive experiments verify the effectiveness of our method.

Image style transfer aims to transfer the styles of artworks onto arbitrary photographs to create novel artistic images. Although style transfer is inherently an underdetermined problem, existing approaches usually assume a deterministic solution, thus failing to capture the full distribution of possible outputs. To address this limitation, we propose a Diverse Image Style Transfer (DIST) framework which achieves significant diversity by enforcing an invertible cross-space mapping. Specifically, the framework consists of three branches: disentanglement branch, inverse branch, and stylization branch. Among them, the disentanglement branch factorizes artworks into content space and style space; the inverse branch encourages the invertible mapping between the latent space of input noise vectors and the style space of generated artistic images; the stylization branch renders the input content image with the style of an artist. Armed with these three branches, our approach is able to synthesize significantly diverse stylized images without loss of quality. We conduct extensive experiments and comparisons to evaluate our approach qualitatively and quantitatively. The experimental results demonstrate the effectiveness of our method.

Human-designed data augmentation strategies havebeen replaced by automatically learned augmentation pol-icy in the past two years. Specifically, recent works haveexperimentally shown that the superior performance of theautomated methods stems from increasing the diversity ofaugmented data. However, two factors regard-ing the diversity of augmented data are still missing: 1)the explicit definition (and thus measurement) of diversityand 2) the quantifiable relationship between diversity andits regularization effects. To fill this gap, we propose a di-versity measure called "Variance Diversity" and theoreti-cally show that the regularization effect of data augmenta-tion is promised by Variance Diversity. We confirm in exper-iments that the relative gain from automated data augmen-tation in test accuracy of a given model is highly correlatedto Variance Diversity. To improve the search process ofautomated augmentation, an unsupervised sampling-basedframework,DivAug, is designed to directly optimize Vari-ance Diversity and hence strengthen the regularization ef-fect. Without requiring a separate search process, the per-formance gain from DivAug is comparable with state-of-the-art method with better efficiency. Moreover, under thesemi-supervised setting, our framework can further improvethe performance of semi-supervised learning algorithmsbased on RandAugment, making it highly applicable to real-world problems, where labeled data is scarce. The code is available at https://github.com/warai-0toko/DivAug.

Real-world data is often unbalanced and long-tailed, but deep models struggle to recognize rare classes in the presence of frequent classes. To address unbalanced data, most studies try balancing the data, the loss, or the classifier to reduce classification bias towards head classes. Far less attention has been given to the latent representations learned with unbalanced data. We show that the feature extractor part of deep networks suffers greatly from this bias. We propose a new loss based on robustness theory, which encourages the model to learn high-quality representations for both head and tail classes. While the general form of the robustness loss may be hard to compute, we further derive an easy-to-compute upper bound that can be minimized efficiently. This procedure reduces representation bias towards head classes in the feature space and achieves new SOTA results on CIFAR100-LT, ImageNet-LT, and iNaturalist long-tail benchmarks. We find that training with robustness increases recognition accuracy of tail classes while largely maintaining the accuracy of head classes. The new robustness loss can be combined with various classifier balancing techniques and can be applied to representations at several layers of the deep model.

Recent state-of-the-art learning-based approaches to point cloud registration have largely been based on graph neural networks (GNN). However, these prominent GNN backbones suffer from the indistinguishable features problem associated with over-smoothing and structural ambiguity of the high-level features, a crucial bottleneck to point cloud registration that has evaded scrutiny in the recent relevant literature. To address this issue, we propose the Distinctiveness oriented Positional Equilibrium (DoPE) module, a novel positional embedding scheme that significantly improves the distinctiveness of the high-level features within both the source and target point clouds, resulting in superior point matching and hence registration accuracy. Specifically, we use the DoPE module in an iterative registration framework, whereby the two point clouds are gradually registered via rigid transformations that are computed from DoPE's position-aware features. With every successive iteration, the DoPE module feeds increasingly consistent positional information to would-be corresponding pairs, which in turn enhances the resulting point-to-point correspondence predictions used to estimate the rigid transformation. Within only a few iterations, the network converges to a desired equilibrium, where the positional embeddings given to matching pairs become essentially identical. We validate the effectiveness of DoPE through comprehensive experiments on various registration benchmarks, registration task settings, and prominent backbones, yielding unprecedented performance improvement across all combinations.

We tackle the long-tailed visual recognition problem from the knowledge distillation perspective by proposing a Distill the Virtual Examples (DiVE) method. Specifically, by treating the predictions of a teacher model as virtual exam- ples, we prove that distilling from these virtual examples is equivalent to label distribution learning under certain con- straints. We show that when the virtual example distribu- tion becomes flatter than the original input distribution, the under-represented tail classes will receive significant im- provements, which is crucial in long-tailed recognition. The proposed DiVE method can explicitly tune the virtual exam- ple distribution to become flat. Extensive experiments on three benchmark datasets, including the large-scale iNat- uralist ones, justify that the proposed DiVE method can significantly outperform state-of-the-art methods. Further- more, additional analyses and experiments verify the virtual example interpretation, and demonstrate the effectiveness of tailored designs in DiVE for long-tailed problems.

Current state-of-the-art Neural Architecture Search (NAS) methods neither efficiently scale to many hardware platforms nor handle diverse architectural search-spaces. To remedy this, we present DONNA (Distilling Optimal Neural Network Architectures), a novel pipeline for rapid, scalable and diverse NAS, that scales to many user scenarios. DONNA consists of three phases. First, an accuracy predictor is built using blockwise knowledge distillation from a reference model. This predictor enables searching across diverse networks with varying macro-architectural parameters such as layer types and attention mechanisms, as well as across micro-architectural parameters such as block repeats and expansion rates. Second, a rapid evolutionary search finds a set of pareto-optimal architectures for any scenario using the accuracy predictor and on-device measurements. Third, optimal models are quickly finetuned to training-from-scratch accuracy. DONNA is up to 100x faster than MNasNet in finding state-of-the-art architectures on-device. Classifying ImageNet, DONNA architectures are 20% faster than EfficientNet-B0 and MobileNetV2 on a Nvidia V100 GPU and 10% faster with 0.5% higher accuracy than MobileNetV2-1.4x on a Samsung S20 smartphone. In addition to NAS, DONNA is used for search-space extension and exploration, as well as hardware-aware model compression.

Knowledge Distillation (KD) aims at transferring knowledge from a larger well-optimized teacher network to a smaller learnable student network. Existing KD methods have mainly considered two types of knowledge, namely the individual knowledge and the relational knowledge. However, these two types of knowledge are usually modeled independently while the inherent correlations between them are largely ignored. It is critical for sufficient student network learning to integrate both individual knowledge and relational knowledge while reserving their inherent correlation. In this paper, we propose to distill the novel holistic knowledge based on an attributed graph constructed among instances. The holistic knowledge is represented as a unified graph-based embedding by aggregating individual knowledge from relational neighborhood samples with graph neural networks, the student network is learned by distilling the holistic knowledge in a contrastive manner. Extensive experiments and ablation studies are conducted on benchmark datasets, the results demonstrate the effectiveness of the proposed method. The code has been published in https://github.com/wyc-ruiker/HKD

In prevalent knowledge distillation, logits in most image recognition models are computed by global average pooling, then used to learn to encode the high-level and task-relevant knowledge. In this work, we solve the limitation of this global logit transfer in this distillation context. We point out that it prevents the transfer of informative spatial information, which provides localized knowledge as well as rich relational information across contexts of an input scene. To exploit the rich spatial information, we propose a simple yet effective logit distillation approach. We add a local spatial pooling layer branch to the penultimate layer, thereby our method extends the standard logit distillation and enables learning of both finely-localized knowledge and holistic representation. Our proposed method shows favorable accuracy improvement against the state-of-the-art methods on several image classification datasets. We show that our distilled students trained on the image classification task can be successfully leveraged for object detection and semantic segmentation tasks; this result demonstrates our method's high transferability.

Image inpainting methods have shown significant improvements by using deep neural networks recently. However, many of these techniques often create distorted structures or blurry inconsistent textures. The problem is rooted in the encoder layers' ineffectiveness in building a complete and faithful embedding of the missing regions from scratch. Existing solutions like course-to-fine, progressive refinement, structural guidance, etc., suffer from huge computational overheads owing to multiple generator networks, limited ability of handcrafted features, and sub-optimal utilization of the information present in the ground truth. We propose a distillation-based approach for inpainting, where we provide direct feature-level supervision while training. We deploy cross and self-distillation techniques and design a dedicated completion-block in encoder to produce more accurate encoding of the holes. Next, we demonstrate how an inpainting network's attention module can improve by leveraging a distillation-based attention transfer technique and enhancing coherence by using a pixel-adaptive global-local feature fusion. We conduct extensive evaluations on multiple datasets to validate our method. Along with achieving significant improvements over previous SOTA methods, the proposed approach's effectiveness is also demonstrated through its ability to improve existing inpainting works.

We address the problem of network quantization, that is, reducing bit-widths of weights and/or activations to lighten network architectures. Quantization methods use a rounding function to map full-precision values to the nearest quantized ones, but this operation is not differentiable. There are mainly two approaches to training quantized networks with gradient-based optimizers. First, a straight-through estimator (STE) replaces the zero derivative of the rounding with that of an identity function, which causes a gradient mismatch problem. Second, soft quantizers approximate the rounding with continuous functions at training time, and exploit the rounding for quantization at test time. This alleviates the gradient mismatch, but causes a quantizer gap problem. We alleviate both problems in a unified framework. To this end, we introduce a novel quantizer, dubbed a distance-aware quantizer (DAQ), that mainly consists of a distance-aware soft rounding (DASR) and a temperature controller. To alleviate the gradient mismatch problem, DASR approximates the discrete rounding with the kernel soft argmax, which is based on our insight that the quantization can be formulated as a distance-based assignment problem between full-precision values and quantized ones. The controller adjusts the temperature parameter in DASR adaptively according to the input, addressing the quantizer gap problem. Experimental results on standard benchmarks show that DAQ outperforms the state of the art significantly for various bit-widths without bells and whistles.