ICCV 2021 Papers

The instance discrimination paradigm has become dominant in unsupervised learning. It always adopts a teacher-student framework, in which the teacher provides embedded knowledge as a supervision signal for the student. The student learns meaningful representations by enforcing instance spatial consistency with the views from the teacher. However, the outputs of the teacher can vary dramatically on the same instance during different training stages, introducing unexpected noise and leading to catastrophic forgetting caused by inconsistent objectives. In this paper, we first integrate instance temporal consistency into current instance discrimination paradigms, and propose a novel and strong algorithm named Temporal Knowledge Consistency (TKC). Specifically, our TKC dynamically ensembles the knowledge of temporal teachers and adaptively selects useful information according to its importance to learning instance temporal consistency. Experimental result shows that TKC can learn better visual representations on both ResNet and AlexNet on linear evaluation protocol while transfer well to downstream tasks. All experiments suggest the good effectiveness and generalization of our method. Code will be made available.

Video highlight detection plays an increasingly important role in social media content filtering, however, it remains highly challenging to develop automated video highlight detection methods because of the lack of temporal annotations (i.e., where the highlight moments are in long videos) for supervised learning. In this paper, we propose a novel weakly supervised method that can learn to detect highlights by mining video characteristics with video level annotations (topic tags) only. Particularly, we exploit audio-visual features to enhance video representation and take temporal cues into account for improving detection performance. Our contributions are threefold: 1) we propose an audio-visual tensor fusion mechanism that efficiently models the complex association between two modalities while reducing the gap of the heterogeneity between the two modalities; 2) we introduce a novel hierarchical temporal context encoder to embed local temporal clues in between neighboring segments; 3) finally, we alleviate the gradient vanishing problem theoretically during model optimization with attention-gated instance aggregation. Extensive experiments on two benchmark datasets (YouTube Highlights and TVSum) have demonstrated our method outperforms other state-of-the-art methods with remarkable improvements.

Training temporal action detection in videos requires large amounts of labeled data, yet such annotation is expensive to collect. Incorporating unlabeled or weakly-labeled data to train action detection model could help reduce annotation cost. In this work, we first introduce the Semi-supervised Action Detection (SSAD) task with a mixture of labeled and unlabeled data and analyze different types of errors in the proposed SSAD baselines which are directly adapted from the semi-supervised classification literature. Identifying that the main source of error is action incompleteness (i.e., missing parts of actions), we alleviate it by designing an unsupervised foreground attention (UFA) module utilizing the conditional independence between foreground and background motion. Then we incorporate weakly-labeled data into SSAD and propose Omni-supervised Action Detection (OSAD) with three levels of supervision. To overcome the accompanying action-context confusion problem in OSAD baselines, an information bottleneck (IB) is designed to suppress the scene information in non-action frames while preserving the action information. We extensively benchmark against the baselines for SSAD and OSAD on our created data splits in THUMOS14 and ActivityNet1.2, and demonstrate the effectiveness of the proposed UFA and IB methods. Lastly, the benefit of our full OSAD-IB model under limited annotation budgets is shown by exploring the optimal annotation strategy for labeled, unlabeled and weakly-labeled data.

Online semantic segmentation on a time series of point cloud frames is an essential task in autonomous driving. Existing models focus on single-frame segmentation, which cannot achieve satisfactory segmentation accuracy and offer unstably flicker among frames. In this paper, we propose a light-weight semantic segmentation framework for large-scale point cloud series, called TempNet, which can improve both the accuracy and the stability of existing semantic segmentation models by combining a novel frame aggregation scheme. To be computational cost efficient, feature extraction and aggregation are only conducted on a small portion of key frames via a temporal feature aggregation (TFA) network using an attentional pooling mechanism, and such enhanced features are propagated to the intermediate non-key frames. To avoid information loss from non-key frames, a partial feature update (PFU) network is designed to partially update the propagated features with the local features extracted on a non-key frame if a large disparity between the two is quickly assessed. As a result, consistent and information-rich features can be obtained for each frame. We implement TempNet on five state-of-the-art (SOTA) point cloud segmentation models and conduct extensive experiments on the SemanticKITTI dataset. Results demonstrate that TempNet outperforms SOTA competitors by wide margins with little extra computational cost.

Most existing image retrieval systems use text queries as a way for the user to express what they are looking for. However, fine-grained image retrieval often requires the ability to also express where in the image the content they are looking for is. The text modality can only cumbersomely express such localization preferences, whereas pointing is a more natural fit. In this paper, we propose an image retrieval setup with a new form of multimodal queries, where the user simultaneously uses both spoken natural language (the what) and mouse traces over an empty canvas (the where) to express the characteristics of the desired target image. We then describe simple modifications to an existing image retrieval model, enabling it to operate in this setup. Qualitative and quantitative experiments show that our model effectively takes this spatial guidance into account, and provides significantly more accurate retrieval results compared to text-only equivalent systems.

We present the Teacher-Student Generative Adversarial Network (TS-GAN) to generate depth images from single RGB images in order to boost the performance of face recognition systems. For our method to generalize well across unseen datasets, we design two components in the architecture, a teacher and a student. The teacher, which itself consists of a generator and a discriminator, learns a latent mapping between input RGB and paired depth images in a supervised fashion. The student, which consists of two generators (one shared with the teacher) and a discriminator, learns from new RGB data with no available paired depth information, for improved generalization. The fully trained shared generator can then be used in runtime to hallucinate depth from RGB for downstream applications such as face recognition. We perform rigorous experiments to show the superiority of TS-GAN over other methods in generating synthetic depth images. Moreover, face recognition experiments demonstrate that our hallucinated depth along with the input RGB images boost performance across various architectures when compared to a single RGB modality by average values of +1.2%, +2.6%, and +2.6% for IIIT-D, EURECOM, and LFW datasets respectively. We make our implementation public at: https://github.com/hardik-uppal/teacher-student-gan.git.

In recent years, considerable progress on the task of text-video retrieval has been achieved by leveraging large-scale pretraining on visual and audio datasets to construct powerful video encoders. By contrast, despite the natural symmetry, the design of effective algorithms for exploiting large-scale language pretraining remains under-explored. In this work, we are the first to investigate the design of such algorithms and propose a novel generalized distillation method,TeachText, which leverages complementary cues from multiple text encoders to provide an enhanced supervisory signal to the retrieval model. Moreover, we extend our method to video side modalities and show that we can effectively reduce the number of used modalities at test time without compromising performance. Our approach advances the state of the art on several video retrieval benchmarks by a significant margin and adds no computational overhead at test time. Last but not least, we show an effective application of our method for eliminating noise from retrieval datasets. Code and data can be found at https://www.robots.ox.ac.uk/ vgg/research/teachtext/.

Few-Shot Learning (FSL) aims at classifying samples into new unseen classes with only a handful of labeled samples available. However, most of the existing methods are based on the image-level pooled representation, yet ignore considerable local clues that are transferable across tasks. To address this issue, we propose an end-to-end Task-aware Part Mining Network (TPMN) by integrating an automatic part mining process into the metric-based model for FSL. The proposed TPMN model enjoys several merits. First, we design a meta filter learner to generate task-aware part filters based on the task embedding in a meta-learning way. The task-aware part filters can adapt to any individual task and automatically mine task-related local parts even for an unseen task. Second, an adaptive importance generator is proposed to identify key local parts and assign adaptive importance weights to different parts. To the best of our knowledge, this is the first work to automatically exploit the task-aware local parts in a meta-learning way for FSL. Extensive experimental results on four standard benchmarks demonstrate that the proposed model performs favorably against state-of-the-art FSL methods.

We introduce Task Switching Networks (TSNs), a task-conditioned architecture with a single unified encoder/decoder for efficient multi-task learning. Multiple tasks are performed by switching between them, performing one task at a time. TSNs have a constant number of parameters irrespective of the number of tasks. This scalable yet conceptually simple approach circumvents the overhead and intricacy of task-specific network components in existing works. In fact, we demonstrate for the first time that multi-tasking can be performed with a single task-conditioned decoder. We achieve this by learning task-specific conditioning parameters through a jointly trained task embedding network, encouraging constructive interaction between tasks. Experiments validate the effectiveness of our approach, achieving state-of-the-art results on two challenging multi-task benchmarks, PASCAL-Context and NYUD. Our analysis of the learned task embeddings further indicates a connection to task relationships studied in the recent literature.

This paper deals with a challenging task of video scene graph generation (VidSGG), which could serve as a structured video representation for high-level understanding tasks. We present a new detect-to-track paradigm for this task by decoupling the context modeling for relation prediction from the complicated low-level entity tracking. Specifically, we design an efficient method for frame-level VidSGG, termed as Target Adaptive Context Aggregation Network (TRACE), with a focus on capturing spatio-temporal context information for relation recognition. Our TRACE framework streamlines the VidSGG pipeline with a modular design, and presents two unique blocks of Hierarchical Relation Tree (HRTree) construction and Target-adaptive Context Aggregation. More specific, our HRTree first provides an adpative structure for organizing possible relation candidates efficiently, and guides context aggregation module to effectively capture spatio-temporal structure information. Then, we obtain a contextualized feature representation for each relation candidate and build a classification head to recognize its relation category. Finally, we provide a simple temporal association strategy to track TRACE detected results to yield the video-level VidSGG. We perform experiments on two VidSGG benchmarks: ImageNet-VidVRD and Action Genome, and the results demonstrate that our TRACE achieves the state-of-the-art performance. The code and models are made available at https://github.com/MCG-NJU/TRACE.

Facial editing is an important task in vision and graphics with numerous applications. However, existing works are incapable to deliver a continuous and fine-grained editing mode (e.g., editing a slightly smiling face to a big laughing one) with natural interactions with users. In this work, we propose Talk-to-Edit, an interactive facial editing framework that performs fine-grained attribute manipulation through dialog between the user and the system. Our key insight is to model a continual ""semantic field"" in the GAN latent space. 1) Unlike previous works that regard the editing as traversing straight lines in the latent space, here the fine-grained editing is formulated as finding a curving trajectory that respects fine-grained attribute landscape on the semantic field. 2) The curvature at each step is location-specific and determined by the input image as well as the users' language requests. 3) To engage the users in a meaningful dialog, our system generates language feedback by considering both the user request and the current state of the semantic field. We also contribute CelebA-Dialog, a visual-language facial editing dataset to facilitate large-scale study. Specifically, each image has manually annotated fine-grained attribute annotations as well as template-based textual descriptions in natural language. Extensive quantitative and qualitative experiments demonstrate the superiority of our framework in terms of 1) the smoothness of fine-grained editing, 2) the identity/attribute preservation, and 3) the visual photorealism and dialog fluency. Notably, user study validates that our overall system is consistently favored by around 80% of the participants.

Weakly supervised object localization (WSOL) is a challenging problem when given image category labels but requires to learn object localization models. Optimizing a convolutional neural network (CNN) for classification tends to activate local discriminative regions while ignoring complete object extent, causing the partial activation issue. In this paper, we argue that partial activation is caused by the intrinsic characteristics of CNN, where the convolution operations produce local receptive fields and experience difficulty to capture long-range feature dependency among pixels. We introduce the token semantic coupled attention map (TS-CAM) to take full advantage of the self-attention mechanism in visual transformer for long-range dependency extraction. TS-CAM first splits an image into a sequence of patch tokens for spatial embedding, which produce attention maps of long-range visual dependency to avoid partial activation. TS-CAM then re-allocates category-related semantics for patch tokens, enabling each of them to be aware of object categories. TS-CAM finally couples the patch tokens with the semantic-agnostic attention map to achieve semantic-aware localization. Experiments on the ILSVRC/CUB-200-2011 datasets show that TS-CAM outperforms its CNN-CAM counterparts by7.1%/27.1%for WSOL, achieving state-of-the-art performance. Code is available at https://github.com/vasgaowei/TS-CAM

Joint forecasting of human trajectory and pose dynamics is a fundamental building block of various applications ranging from robotics and autonomous driving to surveillance systems. Predicting body dynamics requires capturing subtle information embedded in the humans' interactions with each other and with the objects present in the scene. In this paper, we propose a novel TRajectory and POse Dynamics (nicknamed TRiPOD) method based on graph attentional networks to model the human-human and human-object interactions both in the input space and the output space (decoded future output). The model is supplemented by a message passing interface over the graphs to fuse these different levels of interactions efficiently. Furthermore, to incorporate a real-world challenge, we propound to learn an indicator representing whether an estimated body joint is visible/invisible at each frame, e.g. due to occlusion or being outside the sensor field of view. Finally, we introduce a new benchmark for this joint task based on two challenging datasets (PoseTrack and 3DPW) and propose evaluation metrics to measure the effectiveness of predictions in the global space, even when there are invisible cases of joints. Our evaluation shows that TRiPOD outperforms all prior work and state-of-the-art specifically designed for each of the trajectory and pose forecasting tasks.

Due to the superior ability of global dependency modeling, Transformer and its variants have become the primary choice of many vision-and-language tasks. However, in tasks like Visual Question Answering (VQA) and Referring Expression Comprehension (REC), the multimodal prediction often requires visual information from macro- to micro-views. Therefore, how to dynamically schedule the global and local dependency modeling in Transformer has become an emerging issue. In this paper, we propose an example-dependent routing scheme called TRAnsformer Routing (TRAR) to address this issue. Specifically, in TRAR, each visual Transformer layer is equipped with a routing module with different attention spans. The model can dynamically select the corresponding attentions based on the output of the previous inference step, so as to formulate the optimal routing path for each example. Notably, with careful designs, TRAR can reduce the additional computation and memory overhead to almost negligible. To validate TRAR, we conduct extensive experiments on five benchmark datasets of VQA and REC, and achieve superior performance gains than the standard Transformers and a bunch of state-of-the-art methods.

One-stage object detection is commonly implemented by optimizing two sub-tasks: object classification and localization, using heads with two parallel branches, which might lead to a certain level of spatial misalignment in predictions between the two tasks. In this work, we propose a Task-aligned One-stage Object Detection (TOOD) that explicitly aligns the two tasks in a learning-based manner. First, we design a novel Task-aligned Head (T-Head) which offers a better balance between learning task-interactive and task-specific features, as well as a greater flexibility to learn the alignment via a task-aligned predictor. Second, we propose Task Alignment Learning (TAL) to explicitly pull closer (or even unify) the optimal anchors for the two tasks during training via a designed sample assignment scheme and a task-aligned loss. Extensive experiments are conducted on MS-COCO, where TOOD achieves a 51.1 AP at single-model single-scale testing. This surpasses the recent one-stage detectors by a large margin, such as ATSS (47.7 AP), GFL (48.2 AP), and PAA (49.0 AP), with fewer parameters and FLOPs. Qualitative results also demonstrate the effectiveness of TOOD for better aligning the tasks of object classification and localization. Code is available at https://github.com/fcjian/TOOD.

Training vision-based Autonomous driving models is a challenging problem with enormous practical implications. One of the main challenges is the requirement of storage and processing of vast volumes of (possibly redundant) driving video data. In this paper, we study the problem of data-efficient training of autonomous driving systems. We argue that in the context of an edge-device deployment, multi-criteria online video frame subset selection is an appropriate technique for developing such frameworks. We study existing convex optimization based solutions and show that they are unable to provide solution with high weightage to loss of selected video frames. We design a novel multi-criteria online subset selection algorithm, TMCOSS, which uses a thresholded concave function of selection variables. Extensive experiments using driving simulator CARLA show that we are able to drop 80% of the frames, while succeeding to complete 100% of the episodes. We also show that TMCOSS improves performance on the crucial affordance 'Relative Angle' during turns, on inclusion of bucket-specific relative angle loss (BL), leading to selection of more frames in those parts. TMCOSS also achieves an 80% reduction in number of training video frames, on real-world videos from the standard BDD and Cityscapes datasets, for the tasks of drivable area segmentation, and semantic segmentation.

We present THUNDR, a transformer-based deep neural network methodology to reconstruct the 3d pose and shape of people, given monocular RGB images. Key to our methodology is an intermediate 3d marker representation, where we aim to combine the predictive power of model-free-output architectures and the regularizing, anthropometrically-preserving properties of a statistical human surface model like GHUM---a recently introduced, expressive full body statistical 3d human model, trained end-to-end. Our novel transformer-based prediction pipeline can focus on image regions relevant to the task, supports self-supervised regimes, and ensures that solutions are consistent with human anthropometry. We show state-of-the-art results on Human3.6M and 3DPW, for both the fully-supervised and the self-supervised models, for the task of inferring 3d human shape, joint positions, and global translation. Moreover, we observe very solid 3d reconstruction performance for difficult human poses collected in the wild.

Can general-purpose neural models learn to navigate? For PointGoal navigation (""go to x, y""), the answer is a clear `yes' -- mapless neural models composed of task-agnostic components (CNNs and RNNs) trained with large-scale model-free reinforcement learning achieve near-perfect performance. However, for ObjectGoal navigation (""find a TV""), this is an open question; one we tackle in this paper. The current best-known result on ObjectNav with general-purpose models is 6% success rate. First, we show that the key problem is overfitting. Large-scale training results in 94% success rate on training environments and only 8% in validation. We observe that this stems from agents memorizing environment layouts during training -- sidestepping the need for exploration and directly learning shortest paths to nearby goal objects. We show that this is a natural consequence of optimizing for the task metric (which in fact penalizes exploration), is enabled by powerful observation encoders, and is possible due to the finite set of training environment configurations. Informed by our findings, we introduce Treasure Hunt Data Augmentation (THDA) to address overfitting in ObjectNav. THDA inserts 3D scans of household objects at arbitrary scene locations and uses them as ObjectNav goals -- augmenting and greatly expanding the set of training layouts. Taken together with our other proposed changes, we improve the state of art on the Habitat ObjectGoal Navigation benchmark by 90% (from 14% success rate to 27%) and path efficiency by 48% (from 7.5 SPL to 11.1 SPL).

A table arranging data in rows and columns is a very effective data structure, which has been widely used in business and scientific research. Considering large-scale tabular data in online and offline documents, automatic table recognition has attracted increasing attention from the document analysis community. Though human can easily understand the structure of tables, it remains a challenge for machines to understand that, especially due to a variety of different table layouts and styles. Existing methods usually model a table as either the markup sequence or the adjacency matrix between different table cells, failing to address the importance of the logical location of table cells, e.g., a cell is located in the first row and the second column of the table. In this paper, we reformulate the problem of table structure recognition as the table graph reconstruction, and propose an end-to-end trainable table graph reconstruction network (TGRNet) for table structure recognition. Specifically, the proposed method has two main branches, a cell detection branch and a cell logical location branch, to jointly predict the spatial location and the logical location of different cells. Experimental results on three popular table recognition datasets and a new dataset with table graph annotations (TableGraph-350K) demonstrate the effectiveness of the proposed TGRNet for table structure recognition. Code and annotations will be made publicly available.