CVPR 2021 论文列表

We present M3P, a Multitask Multilingual Multimodal Pre-trained model that combines multilingual pre-training and multimodal pre-training into a unified framework via multitask pre-training. Our goal is to learn universal representations that can map objects occurred in different modalities or texts expressed in different languages into a common semantic space. In addition, to explicitly encourage fine-grained alignment between images and non-English languages, we also propose Multimodal Code-switched Training (MCT) to combine monolingual pre-training and multimodal pre-training via a code-switch strategy. Experiments are performed on the multilingual image retrieval task across two benchmark datasets, including MSCOCO and Multi30K. M3P can achieve comparable results for English and new state-of-the-art results for non-English languages.

In this paper, we propose a Monocular 3D Single Stage object Detector (M3DSSD) with feature alignment and asymmetric non-local attention. Current anchor-based monocular 3D object detection methods suffer from feature mismatching. To overcome this, we propose a two-step feature alignment approach. In the first step, the shape alignment is performed to enable the receptive field of the feature map to focus on the pre-defined anchors with high confidence scores. In the second step, the center alignment is used to align the features at 2D/3D centers. Further, it is often difficult to learn global information and capture long-range relationships, which are important for the depth prediction of objects. Therefore, we propose a novel asymmetric non-local attention block with multi-scale sampling to extract depth-wise features. The proposed M3DSSD achieves significantly better performance than the monocular 3D object detection methods on the KITTI dataset, in both 3D object detection and bird's eye view tasks. The code is released at https://github.com/mumianyuxin/M3DSSD.

In this paper, we address the problem of separating individual speech signals from videos using audio-visual neural processing. Most conventional approaches utilize frame-wise matching criteria to extract shared information between co-occurring audio and video. Thus, their performance heavily depends on the accuracy of audio-visual synchronization and the effectiveness of their representations. To overcome the frame discontinuity problem between two modalities due to transmission delay mismatch or jitter, we propose a cross-modal affinity network (CaffNet) that learns global correspondence as well as locally-varying affinities between audio and visual streams. Given that the global term provides stability over a temporal sequence at the utterance-level, this resolves the label permutation problem characterized by inconsistent assignments. By extending the proposed cross-modal affinity on the complex network, we further improve the separation performance in the complex spectral domain. Experimental results verify that the proposed methods outperform conventional ones on various datasets, demonstrating their advantages in real-world scenarios.

Tremendous efforts have been made on instance segmentation but the mask quality is still not satisfactory. The boundaries of predicted instance masks are usually imprecise due to the low spatial resolution of feature maps and the imbalance problem caused by the extremely low proportion of boundary pixels. To address these issues, we propose a conceptually simple yet effective post-processing refinement framework to improve the boundary quality based on the results of any instance segmentation model, termed BPR. Following the idea of looking closer to segment boundaries better, we extract and refine a series of small boundary patches along the predicted instance boundaries. The refinement is accomplished by a boundary patch refinement network at higher resolution. The proposed BPR framework yields significant improvements over the Mask R-CNN baseline on Cityscapes benchmark, especially on the boundary-aware metrics. Moreover, by applying the BPR framework to the PolyTransform + SegFix baseline, we reached 1st place on the Cityscapes leaderboard.

While most conversational AI systems focus on textual dialogue only, conditioning utterances on visual context (when it's available) can lead to more realistic conversations. Unfortunately, a major challenge for incorporating visual context into conversational dialogue is the lack of large-scale labeled datasets. We provide a solution in the form of a new visually conditioned Future Utterance Prediction task. Our task involves predicting the next utterance in a video, using both visual frames and transcribed speech as context. By exploiting the large number of instructional videos online, we train a model to solve this task at scale, without the need for manual annotations. Leveraging recent advances in multimodal learning, our model consists of a novel co-attentional multimodal video transformer, and when trained on both textual and visual context, outperforms baselines that use textual inputs alone. Further, we demonstrate that our model trained for this task on unlabelled videos achieves state-of-the-art performance on a number of downstream VideoQA benchmarks such as MSRVTT-QA, MSVD-QA, ActivityNet-QA and How2QA.

An LBYL ( 'Look Before You Leap' ) Network is proposed for end-to-end trainable one-stage visual grounding. The idea behind LBYL-Net is intuitive and straightforward: we follow a language's description to localize the target object based on its relative spatial relation to 'Landmarks', which is characterized by some spatial positional words and some descriptive words about the object. The core of our LBYL-Net is a landmark feature convolution module that transmits the visual features with the guidance of linguistic description along with different directions. Consequently, such a module encodes the relative spatial positional relations between the current object and its context. Then we combine the contextual information from the landmark feature convolution module with the target's visual features for grounding. To make this landmark feature convolution light-weight, we introduce a dynamic programming algorithm (termed dynamic max pooling) with low complexity to extract the landmark feature. Thanks to the landmark feature convolution module, we mimic the human behavior of `Look Before You Leap` to design an LBYL-Net, which takes full consideration of contextual information. Extensive experiments show our method's effectiveness in four grounding datasets. Specifically, our LBYL-Net outperforms all state-of-the-art two-stage and one-stage methods on ReferitGame. On RefCOCO and RefCOCO+, Our LBYL-Net also achieves comparable results or even better results than existing one-stage methods. Code is available at https://github.com/svip-lab/LBYLNet.

Long-tailed data distribution is common in many multi-label visual recognition tasks and the direct use of these data for training usually leads to relatively low performance on tail classes. While re-balanced data sampling can improve the performance on tail classes, it may also hurt the performance on head classes in training due to label co-occurrence. In this paper, we propose a new approach to train on both uniform and re-balanced samplings in a collaborative way, resulting in performance improvement on both head and tail classes. More specifically, we design a visual recognition network with two branches: one takes the uniform sampling as input while the other takes the re-balanced sampling as the input. For each branch, we conduct visual recognition using a binary-cross-entropy-based classification loss with learnable logit compensation. We further define a new cross-branch loss to enforce the consistency when the same input image goes through the two branches. We conduct extensive experiments on VOC-LT and COCO-LT datasets. The results show that the proposed method significantly outperforms previous state-of-the-art methods on long-tailed multi-label visual recognition.

Referring image segmentation aims to segment the objects referred by a natural language expression. Previous methods usually focus on designing an implicit and recurrent feature interaction mechanism to fuse the visual-linguistic features to directly generate the final segmentation mask without explicitly modeling the localization of the referent guided by language expression and designing a powerful segmentation module. To tackle these problems, we view this task from another perspective by decoupling it into a "locate-then-segment" (LTS) scheme. Given a language expression, people generally first perform attention to the corresponding target image regions, then generate a segmentation mask about the object based on its context. The LTS first extracts and fuses both visual and textual features to get a cross-modal representation, then applies a cross-model interaction on the visual-textual features to locate the referred object with position prior, and finally generates the segmentation result with a light-weight network. Our LTS is simple but surprisingly effective. On three popular benchmark datasets, the LTS outperforms all the previous state-of-the-arts methods by a large margin (e.g., +3.2% on RefCOCO+ and +3.4% on RefCOCOg). In addition, our model is more interpretable with explicitly locating the object, which is also proved by visualization experiments. Accordingly, this framework is very promising to serve as a pipeline for referring image segmentation.

Registering point clouds of dressed humans to parametric human models is a challenging task in computer vision. Traditional approaches often rely on heavily engineered pipelines that require accurate manual initialization of human poses and tedious post-processing. More recently, learning-based methods are proposed in hope to automate this process. We observe that pose initialization is key to accurate registration but existing methods often fail to provide accurate pose initialization. One major obstacle is that, despite recent effort on rotation representation learning in neural networks, regressing joint rotations from point clouds or images of humans is still very challenging. To this end, we propose novel piecewise transformation fields (PTF), a set of functions that learn 3D translation vectors to map any query point in posed space to its correspond position in rest-pose space. We combine PTF with multi-class occupancy networks, obtaining a novel learning-based framework that learns to simultaneously predict shape and per-point correspondences between the posed space and the canonical space for clothed human. Our key insight is that the translation vector for each query point can be effectively estimated using the point-aligned local features; consequently, rigid per bone transformations and joint rotations can be obtained efficiently via a least-square fitting given the estimated point correspondences, circumventing the challenging task of directly regressing joint rotations from neural networks. Furthermore, the proposed PTF facilitate canonicalized occupancy estimation, which greatly improves generalization capability and result in more accurate surface reconstruction with only half of the parameters compared with the state-of-the-art. Both qualitative and quantitative studies show that fitting parametric models with poses initialized by our network results in much better registration quality, especially for extreme poses.

The objective of this work is to localize sound sources that are visible in a video without using manual annotations. Our key technical contribution is to show that, by training the network to explicitly discriminate challenging image fragments, even for images that do contain the object emitting the sound, we can significantly boost the localization performance. We do so elegantly by introducing a mechanism to mine hard samples and add them to a contrastive learning formulation automatically. We show that our algorithm achieves state-of-the-art performance on the popular Flickr SoundNet dataset. Furthermore, we introduce the VGG-Sound Source (VGG-SS) benchmark, a new set of annotations for the recently-introduced VGG-Sound dataset, where the sound sources visible in each video clip are explicitly marked with bounding box annotations. This dataset is 20 times larger than analogous existing ones, contains 5K videos spanning over 200 categories, and, differently from Flickr SoundNet, is video-based. On VGG-SS, we also show that our algorithm achieves state-of-the-art performance against several baselines. Code and datasets can be found at http://www.robots.ox.ac.uk/ vgg/research/lvs/

We present a novel method for local image feature matching. Instead of performing image feature detection, description, and matching sequentially, we propose to first establish pixel-wise dense matches at a coarse level and later refine the good matches at a fine level. In contrast to dense methods that use a cost volume to search correspondences, we use self and cross attention layers in Transformer to obtain feature descriptors that are conditioned on both images. The global receptive field provided by Transformer enables our method to produce dense matches in low-texture areas, where feature detectors usually struggle to produce repeatable interest points. The experiments on indoor and outdoor datasets show that LoFTR outperforms state-of-the-art methods by a large margin. LoFTR also ranks first on two public benchmarks of visual localization among the published methods. Code is available at our project page: https://zju3dv.github.io/loftr/.

We present an efficient high-resolution network, Lite-HRNet, for human pose estimation. We start by simply applying the efficient shuffle block in ShuffleNet to HRNet (high-resolution network), yielding stronger performance over popular lightweight networks, such as MobileNet, ShuffleNet, and Small HRNet. We find that the heavily-used pointwise (1x1) convolutions in shuffle blocks become the computational bottleneck. We introduce a lightweight unit, conditional channel weighting, to replace costly pointwise (1x1) convolutions in shuffle blocks. The complexity of channel weighting is linear w.r.t the number of channels and lower than the quadratic time complexity for pointwise convolutions. Our solution learns the weights from all the channels and over multiple resolutions that are readily available in the parallel branches in HRNet. It uses the weights as the bridge to exchange information across channels and resolutions, compensating the role played by the pointwise (1x1) convolution. Lite-HRNet demonstrates superior results on human pose estimation over popular lightweight networks. Moreover, Lite-HRNet can be easily applied to semantic segmentation task in the same lightweight manner. The code and models have been publicly available at https://github.com/HRNet/Lite-HRNet.

Makeup transfer is the task of applying on a source face the makeup style from a reference image. Real-life makeups are diverse and wild, which cover not only color-changing but also patterns, such as stickers, blushes, and jewelries. However, existing works overlooked the latter components and confined makeup transfer to color manipulation, focusing only on light makeup styles. In this work, we propose a holistic makeup transfer framework that can handle all the mentioned makeup components. It consists of an improved color transfer branch and a novel pattern transfer branch to learn all makeup properties, including color, shape, texture, and location. To train and evaluate such a system, we also introduce new makeup datasets for real and synthetic extreme makeup. Experimental results show that our framework achieves the state of the art performance on both light and extreme makeup styles.

Although current deep learning-based face forgery detectors achieve impressive performance in constrained scenarios, they are vulnerable to samples created by unseen manipulation methods. Some recent works show improvements in generalisation but rely on cues that are easily corrupted by common post-processing operations such as compression. In this paper, we propose LipForensics, a detection approach capable of both generalising to novel manipulations and withstanding various distortions. LipForensics targets high-level semantic irregularities in mouth movements, which are common in many generated videos. It consists in first pretraining a spatio-temporal network to perform visual speech recognition (lipreading), thus learning rich internal representations related to natural mouth motion. A temporal network is subsequently finetuned on fixed mouth embeddings of real and forged data in order to detect fake videos based on mouth movements without overfitting to low-level, manipulation-specific artefacts. Extensive experiments show that this simple approach significantly surpasses the state-of-the-art in terms of generalisation to unseen manipulations and robustness to perturbations, as well as shed light on the factors responsible for its performance.

In this paper, we present a video-based learning framework for animating personalized 3D talking faces from audio. We introduce two training-time data normalizations that significantly improve data sample efficiency. First, we isolate and represent faces in a normalized space that decouples 3D geometry, head pose, and texture. This decomposes the prediction problem into regressions over the 3D face shape and the corresponding 2D texture atlas. Second, we leverage facial symmetry and approximate albedo constancy of skin to isolate and remove spatiotemporal lighting variations. Together, these normalizations allow simple networks to generate high fidelity lip-sync videos under novel ambient illumination while training with just a single video (of usually < 5 minutes). Further, to stabilize temporal dynamics, we introduce an auto-regressive approach that conditions the model on its previous visual state. Human ratings and objective metrics demonstrate that our method outperforms contemporary state-of-the-art audio-driven video reenactment benchmarks in terms of realism, lip-sync and visual quality scores. We illustrate several applications enabled by our framework.

Prior work in scene graph generation requires categorical supervision at the level of triplets---subjects and objects, and predicates that relate them, either with or without bounding box information. However, scene graph generation is a holistic task: thus holistic, contextual supervision should intuitively improve performance. In this work, we explore how linguistic structures in captions can benefit scene graph generation. Our method captures the information provided in captions about relations between individual triplets, and context for subjects and objects (e.g. visual properties are mentioned). Captions are a weaker type of supervision than triplets since the alignment between the exhaustive list of human-annotated subjects and objects in triplets, and the nouns in captions, is weak. However, given the large and diverse sources of multimodal data on the web (e.g. blog posts with images and captions), linguistic supervision is more scalable than crowdsourced triplets. We show extensive experimental comparisons against prior methods which leverage instance- and image-level supervision, and ablate our method to show the impact of leveraging phrasal and sequential context, and techniques to improve localization of subjects and objects.

Generative Adversarial Networks (GANs) are able to generate high-quality images, but it remains difficult to explicitly specify the semantics of synthesized images. In this work, we aim to better understand the semantic representation of GANs, and thereby enable semantic control in GAN's generation process. Interestingly, we find that a well-trained GAN encodes image semantics in its internal feature maps in a surprisingly simple way: a linear transformation of feature maps suffices to extract the generated image semantics. To verify this simplicity, we conduct extensive experiments on various GANs and datasets; and thanks to this simplicity, we are able to learn a semantic segmentation model for a trained GAN from a small number (e.g., 8) of labeled images. Last but not least, leveraging our finding, we propose two few-shot image editing approaches, namely Semantic-Conditional Sampling and Semantic Image Editing. Given a trained GAN and as few as eight semantic annotations, the user is able to generate diverse images subject to a user-provided semantic layout, and control the synthesized image semantics. We have made the code publicly available.

In this paper, we present a joint end-to-end line segment detection algorithm using Transformers that is post-processing and heuristics-guided intermediate processing (edge/junction/region detection) free. Our method, named LinE segment TRansformers (LETR), takes advantages of having integrated tokenized queries, a self-attention mechanism, and encoding-decoding strategy within Transformers by skipping standard heuristic designs for the edge element detection and perceptual grouping processes. We equip Transformers with a multi-scale encoder/decoder strategy to perform fine-grained line segment detection under a direct endpoint distance loss. This loss term is particularly suitable for detecting geometric structures such as line segments that are not conveniently represented by the standard bounding box representations. The Transformers learn to gradually refine line segments through layers of self-attention. In our experiments, we show state-of-the-art results on Wireframe and YorkUrban benchmarks.

Feature alignment is an approach to improving robustness to distribution shift that matches the distribution of feature activations between the training distribution and test distribution. A particularly simple but effective approach to feature alignment involves aligning the batch normalization statistics between the two distributions in a trained neural network. This technique has received renewed interest lately because of its impressive performance on robustness benchmarks. However, when and why this method works is not well understood. We investigate the approach in more detail and identify several limitations. We show that it only significantly helps with a narrow set of distribution shifts and we identify several settings in which it even degrades performance. We also explain why these limitations arise by pinpointing why this approach can be so effective in the first place. Our findings call into question the utility of this approach and Unsupervised Domain Adaptation more broadly for improving robustness in practice.