ECCV 2022 Papers

"The current modus operandi in adapting pre-trained models involves updating all the backbone parameters, i.e. full fine-tuning. This paper introduces Visual Prompt Tuning (VPT) as an efficient and effective alternative to full fine-tuning for large-scale Transformer models in vision. Taking inspiration from recent advances in efficiently tuning large language models, VPT introduces only a small amount (less than 1% of model parameters) of trainable parameters in the input space while keeping the model backbone frozen. Via extensive experiments on a wide variety of downstream recognition tasks, we show that VPT achieves significant performance gains compared to other parameter-efficient tuning protocols. Most importantly, VPT even outperforms full fine-tuning in many cases across model capacities and training scales, while reducing per-task storage cost. Code is available at https://github.com/kmnp/vpt."

"Each year, thousands of people learn new visual categorization tasks - radiologists learn to recognize tumors, birdwatchers learn to distinguish similar species, and crowd workers learn how to annotate valuable data for applications like autonomous driving. As humans learn, their brain updates the visual features it extracts and attend to, which ultimately informs their final classification decisions. In this work, we propose a novel task of tracing the evolving classification behavior of human learners as they engage in challenging visual classification tasks. We propose models that jointly extract the visual features used by learners as well as predicting the classification functions they utilize. We collect three challenging new datasets from real human learners in order to evaluate the performance of different visual knowledge tracing methods. Our results show that our recurrent models are able to predict the classification behavior of human learners on three challenging medical image and species identification tasks."

"This work addresses visual cross-view metric localization for outdoor robotics. Given a ground-level color image and a satellite patch that contains the local surroundings, the task is to identify the location of the ground camera within the satellite patch. Related work addressed this task for range-sensors (LiDAR, Radar), but for vision, only as a secondary regression step after an initial cross-view image retrieval step. Since the local satellite patch could also be retrieved through any rough localization prior (e.g. from GPS/GNSS, temporal filtering), we drop the image retrieval objective and focus on the metric localization only. We devise a novel network architecture with denser satellite descriptors, similarity matching at the bottleneck (rather than at the output as in image retrieval), and a dense spatial distribution as output to capture multi-modal localization ambiguities. We compare against a state-of-the-art regression baseline that uses global image descriptors. Quantitative and qualitative experimental results on the recently proposed VIGOR and the Oxford RobotCar datasets validate our design. The produced probabilities are correlated with localization accuracy, and can even be used to roughly estimate the ground camera’s heading when its orientation is unknown. Overall, our method reduces the median metric localization error by 51%, 37%, and 28% compared to the state-of-the-art when generalizing respectively in the same area, across areas, and across time."

"The goal of this work is to reconstruct speech from a silent talking face video. Recent studies have shown impressive performance on synthesizing speech from silent talking face videos. However, they have not explicitly considered on varying identity characteristics of different speakers, which place a challenge in the video-to-speech synthesis, and this becomes more critical in unseen-speaker settings. Our approach is to separate the speech content and the visage-style from a given silent talking face video. By guiding the model to independently focus on modeling the two representations, we can obtain the speech of high intelligibility from the model even when the input video of an unseen subject is given. To this end, we introduce speech-visage selection that separates the speech content and the speaker identity from the visual features of the input video. The disentangled representations are jointly incorporated to synthesize speech through visage-style based synthesizer which generates speech by coating the visage-styles while maintaining the speech content. Thus, the proposed framework brings the advantage of synthesizing the speech containing the right content even with the silent talking face video of an unseen subject. We validate the effectiveness of the proposed framework on the GRID, TCD-TIMIT volunteer, and LRW datasets."

"While monocular 3D pose estimation seems to have achieved very accurate results on the public datasets, their generalization ability is largely overlooked. In this work, we perform a systematic evaluation of the existing methods and find that they get notably larger errors when tested on different cameras, human poses and appearance. To address the problem, we introduce VirtualPose, a two-stage learning framework to exploit the hidden ""free lunch"" specific to this task, i.e. generating infinite number of poses and cameras for training models at no cost. To that end, the first stage transforms images to abstract geometry representations (AGR), and then the second maps them to 3D poses. It addresses the generalization issue from two aspects: (1) the first stage can be trained on diverse 2D datasets to reduce the risk of over-fitting to limited appearance; (2) the second stage can be trained on diverse AGR synthesized from a large number of virtual cameras and poses. It outperforms the SOTA methods without using any paired images and 3D poses from the benchmarks, which paves the way for practical applications. Code is available at https://github.com/wkom/VirtualPose."

"Novel view synthesis is a long-standing problem. In this work, we consider a variant of the problem where we are given only a few context views sparsely covering a scene or an object. The goal is to predict novel viewpoints in the scene, which requires learning priors. The current state of the art is based on Neural Radiance Field (NeRF), and while achieving impressive results, the methods suffer from long training times as they require evaluating millions of 3D point samples via a neural network for each image. We propose a 2D-only method that maps multiple context views and a query pose to a new image in a single pass of a neural network. Our model uses a two-stage architecture consisting of a codebook and a transformer model. The codebook is used to embed individual images into a smaller latent space, and the transformer solves the view synthesis task in this more compact space. To train our model efficiently, we introduce a novel branching attention mechanism that allows us to use the same model not only for neural rendering but also for camera pose estimation. Experimental results on real-world scenes show that our approach is competitive compared to NeRF-based methods while not reasoning explicitly in 3D, and it is faster to train."

"Understanding and forecasting future trajectories of agents are critical for behavior analysis, robot navigation, autonomous cars, and other related applications. Previous methods mostly treat trajectory prediction as time sequence generation. Different from them, this work studies agents’ trajectories in a ""vertical"" view, i.e., modeling and forecasting trajectories from the spectral domain. Different frequency bands in the trajectory spectrums could hierarchically reflect agents’ motion preferences at different scales. The low-frequency and high-frequency portions could represent their coarse motion trends and fine motion variations, respectively. Accordingly, we propose a hierarchical network V$^2$-Net, which contains two sub-networks, to hierarchically model and predict agents’ trajectories with trajectory spectrums. The coarse-level keypoints estimation sub-network first predicts the ""minimal"" spectrums of agents’ trajectories on several ""key"" frequency portions. Then the fine-level spectrum interpolation sub-network interpolates the spectrums to reconstruct the final predictions. Experimental results display the competitiveness and superiority of V$^2$-Net on both ETH-UCY benchmark and the Stanford Drone Dataset."

"Restoration of weather degraded videos is a challenging problem due to diverse weather conditions e.g., rain, haze, snow, etc. Existing works handle video restoration for each weather using a different custom-designed architecture. This approach has many limitations. First, a custom-designed architecture for each weather condition requires domain-specific knowledge. Second, disparate network architectures across weather conditions prevent easy knowledge transfer to novel weather conditions where we do not have a lot of data to train a model from scratch. For example, while there is a lot of common knowledge to exploit between the models of different weather conditions at day or night time, it is difficult to do such adaptation. To this end, we propose a generic architecture that is effective for any weather condition due to the ability to extract robust feature maps without any domain-specific knowledge. This is achieved by novel components: spatio-temporal feature modulation, multi-level feature aggregation, and recurrent guidance decoder. Next, we propose a meta-learning based adaptation of our deep architecture to the restoration of videos in data-poor conditions (night-time videos). We show comprehensive results on video de-hazing and de-raining datasets in addition to the meta-learning based adaptation results on night-time video restoration tasks. Our results clearly outperform the state-of-the-art weather degraded video restoration methods."

"Video question answering is a challenging task that requires understanding jointly the language input, the visual information in individual video frames, as well as the temporal information about the events occurring in the video. In this paper, we propose a novel multi-stream video encoder for video question answering that uses multiple video inputs and a new video-text iterative co-tokenization approach to answer a variety of questions related to videos. We experimentally evaluate the model on several datasets, such as MSRVTT-QA, MSVD-QA, IVQA, outperforming the previous state-of-the-art by large margins. Simultaneously, our model reduces the required GFLOPs from 150-360 to only 67, producing a highly efficient video question answering model."

"While Video Instance Segmentation (VIS) has seen rapid progress, current approaches struggle to predict high-quality masks with accurate boundary details. Moreover, the predicted segmentations often fluctuate over time, suggesting that temporal consistency cues are neglected or not fully utilized. In this paper, we set out to tackle these issues, with the aim of achieving highly detailed and more temporally stable mask predictions for VIS. We first propose the Video Mask Transfiner (VMT) method, capable of leveraging fine-grained high-resolution features thanks to a highly efficient video transformer structure. Our VMT detects and groups sparse error-prone spatio-temporal regions of each tracklet in the video segment, which are then refined using both local and instance-level cues. Second, we identify that the coarse boundary annotations of the popular YouTube-VIS dataset constitute a major limiting factor. Based on our VMT architecture, we therefore design an automated annotation refinement approach by iterative training and self-correction. To benchmark high-quality mask predictions for VIS, we introduce the HQ-YTVIS dataset, consisting of a manually re-annotated test set and our automatically refined training data. We compare VMT with the most recent state-of-the-art methods on the HQ-YTVIS, as well as the Youtube-VIS, OVIS and BDD100K MOTS benchmarks. Experimental results clearly demonstrate the efficacy and effectiveness of our method on segmenting complex and dynamic objects, by capturing precise details."

"Video frame interpolation is a challenging task due to the ever-changing real-world scene. Previous methods often calculate the bi-directional optical flows and then predict the intermediate optical flows under the linear motion assumptions, leading to isotropic intermediate flow generation. Follow-up research obtained anisotropic adjustment through estimated higher-order motion information with extra frames. Based on the motion assumptions, their methods are hard to model the complicated motion in real scenes. In this paper, we propose an end-to-end training method A^2OF for video frame interpolation with event-driven Anisotropic Adjustment of Optical Flows. Specifically, we use events to generate optical flow distribution masks for the intermediate optical flow, which can model the complicated motion between two frames. Our proposed method outperforms the previous methods in video frame interpolation, taking supervised event-based video interpolation to a higher stage."

"State-of-the-art transformer-based video instance segmentation (VIS) approaches typically utilize either single-scale spatio-temporal features or per-frame multi-scale features during the attention computations. We argue that such an attention computation ignores the multi-scale spatio-temporal feature relationships that are crucial to tackle target appearance deformations in videos. To address this issue, we propose a transformer-based VIS framework, named MS-STS VIS, that comprises a novel multi-scale spatio-temporal split (MS-STS) attention module in the encoder. The proposed MS-STS module effectively captures spatio-temporal feature relationships at multiple scales across frames in a video. We further introduce an attention block in the decoder to enhance the temporal consistency of the detected instances in different frames of a video. Moreover, an auxiliary discriminator is introduced during training to ensure better foreground-background separability within the multi-scale spatio-temporal feature space. We conduct extensive experiments on two benchmarks: Youtube-VIS (2019 and 2021). Our MS-STS VIS achieves state-of-the-art performance on both benchmarks. When using the ResNet50 backbone, our MS-STS achieves a mask AP of 50.1%, outperforming the best reported results in literature by 2.7% and by 4.8% at higher overlap threshold of AP75, while being comparable in model size and speed on Youtube-VIS 2019 val. set. When using the Swin Transformer backbone, MS-STS VIS achieves mask AP of 61.0% on Youtube-VIS 2019 val. set. Our code and models will be made public."

"This paper proposes a Video Graph Transformer (VGT) model for Video Quetion Answering (VideoQA). VGT’s uniqueness are two-fold: 1) it designs a dynamic graph transformer module which encodes video by explicitly capturing the visual objects, their relations, and dynamics for complex spatio-temporal reasoning; and 2) it exploits disentangled video and text Transformers for relevance comparison between the video and text to perform QA, instead of entangled cross-modal Transformer for answer classification. Vision-text communication is done by additional cross-modal interaction modules. With more reasonable video encoding and QA solution, we show that VGT can achieve much better performances on VideoQA tasks that challenge dynamic relation reasoning than prior arts in the pretraining-free scenario. Its performances even surpass those models that are pretrained with millions of external data. We further show that VGT can also benefit a lot from self-supervised cross-modal pretraining, yet with orders of magnitude smaller data. These results clearly demonstrate the effectiveness and superiority of VGT, and reveal its potential for more data-efficient pretraining. With comprehensive analyses and some heuristic observations, we hope that VGT can promote VQA research beyond coarse recognition/description towards fine-grained relation reasoning in realistic videos."

"Novel view synthesis (NVS) and video prediction (VP) are typically considered disjoint tasks in computer vision. However, they can both be seen as ways to observe the spatial-temporal world: NVS aims to synthesize a scene from a new point of view, while VP aims to see a scene from a new point of time. These two tasks provide complementary signals to obtain a scene representation, as viewpoint changes from spatial observations inform depth, and temporal observations inform the motion of cameras and individual objects. Inspired by these observations, we propose to study the problem of Video Extrapolation in Space and Time (VEST). We propose a model that tackles this problem and leverages the self-supervision from both tasks, while existing methods are designed to solve one of them. Experiments show that our method achieves performance better than or comparable to several state-of-the-art NVS and VP methods on indoor and outdoor real-world datasets."

"It would be a technological feat to be able to create a system that can hold a meaningful conversation with humans about what they watch. A setup toward that goal is presented as a video dialog task, where the system is asked to generate natural utterances in response to a question in an ongoing dialog. The task poses great visual, linguistic, and reasoning challenges that cannot be easily overcome without an appropriate representation scheme over video and dialog that supports high-level reasoning. To tackle these challenges we present a new object-centric framework for video dialog that supports neural reasoning dubbed COST - which stands for Conversation about Objects in Space-Time. Here dynamic space-time visual content in videos is first parsed into object trajectories. Given this video abstraction, COST maintains and tracks object-associated dialog states, which are updated upon receiving new questions. Object interactions are dynamically and conditionally inferred for each question, and these serve as the basis for relational reasoning among them. COST also maintains a history of previous answers, and this allows retrieval of relevant object-centric information to enrich the answer forming process. Language production then proceeds in a step-wise manner, taking into the context of the current utterance, the existing dialog, the current question. We evaluate COST on the DSTC7 and DSTC8 benchmarks, demonstrating its competitiveness against state-of-the-arts."

"Video Anomaly Detection (VAD) is an important topic in computer vision. Motivated by the recent advances in self-supervised learning, this paper addresses VAD by solving an intuitive yet challenging pretext task, i.e., spatio-temporal jigsaw puzzles, which is cast as a multi-label fine-grained classification problem. Our method exhibits several advantages over existing works: 1) the spatio-temporal jigsaw puzzles are decoupled in terms of spatial and temporal dimensions, responsible for capturing highly discriminative appearance and motion features, respectively; 2) full permutations are used to provide abundant jigsaw puzzles covering various difficulty levels, allowing the network to distinguish subtle spatio-temporal differences between normal and abnormal events; and 3) the pretext task is tackled in an end-to-end manner without relying on any pre-trained models. Our method outperforms state-of-the-art counterparts on three public benchmarks. Especially on ShanghaiTech Campus, the result is superior to reconstruction and prediction-based methods by a large margin."

"Current methods for video activity localisation over time assume implicitly that activity temporal boundaries labelled for model training are determined and precise. However, in unscripted natural videos, different activities mostly transit smoothly, so that it is intrinsically ambiguous to determine in labelling precisely when an activity starts and ends over time. Such uncertainties in temporal labelling are currently ignored in model training, resulting in learning mis-matched video-text correlation with poor generalisation in test. In this work, we solve this problem by introducing Elastic Moment Bounding (EMB) to accommodate flexible and adaptive activity temporal boundaries towards modelling universally interpretable video-text correlation with tolerance to underlying temporal uncertainties in pre-fixed annotations. Specifically, we construct elastic boundaries adaptively by mining and discovering frame-wise temporal endpoints that can maximise the alignment between video segments and query sentences. To enable both more accurate matching (segment content attention) and more robust localisation (segment elastic boundaries), we optimise the selection of frame-wise endpoints subject to segment-wise contents by a novel Guided Attention mechanism. Extensive experiments on three video activity localisation benchmarks demonstrate compellingly the EMB’s advantages over existing methods without modelling uncertainty."

"We investigate the problem of estimating uncertainty for training data, so that deep neural networks can make use of the results for learning from limited supervision. However, both prediction probability and entropy estimate uncertainty from the instantaneous information. In this paper, we present a novel approach that measures uncertainty from the vibration of sequential data, e.g., the output probability during the training procedure. The key observation is that, a training sample that suffers heavier vibration often offers richer information when it is manually labeled. Motivated by Bayesian theory, we sample the sequences from the latter part of training. We make use of the Fourier Transformation to measure the extent of vibration, deriving a powerful tool that can be used for semi-supervised, active learning, and one-bit supervision. Experiments on the CIFAR10, CIFAR100, mini-ImageNet and ImageNet datasets validate the effectiveness of our approach."

"Vision transformers (ViTs) inherited the success of NLP but their structures have not been sufficiently investigated and optimized for visual tasks. One of the simplest solutions is to directly search the optimal one via the widely used neural architecture search (NAS) in CNNs. However, we empirically find this straightforward adaptation would encounter catastrophic failures and be frustratingly unstable for the training of superformer. In this paper, we argue that since ViTs mainly operate on token embeddings with little inductive bias, imbalance of channels for different architectures would worsen the weight-sharing assumption and cause the training instability as a result. Therefore, we develop a new cyclic weight-sharing mechanism for token embeddings of the ViTs, which enables each channel could more evenly contribute to all candidate architectures. Besides, we also propose identity shifting to alleviate the many-to-one issue in superformer and leverage weak augmentation and regularization techniques for more steady training empirically. Based on these, our proposed method, ViTAS, has achieved significant superiority in both DeiT- and Twins-based ViTs. For example, with only $1.4$G FLOPs budget, our searched architecture has $3.3\%$ ImageNet-$1$k accuracy than the baseline DeiT. With $3.0$G FLOPs, our results achieve $82.0\%$ accuracy on ImageNet-$1$k, and $45.9\%$ mAP on COCO$2017$ which is $2.4\%$ superior than other ViTs."