ICCV 2021 Papers

Multi-object tracking and segmentation (MOTS) is important for understanding dynamic scenes in video data. Existing methods perform well on multi-object detection and segmentation for independent video frames, but tracking of objects over time remains a challenge. MOTS methods formulate tracking locally, i.e., frame-by-frame, leading to sub-optimal results. Classical global methods on tracking operate directly on object detections, which leads to a combinatorial growth in the detection space. In contrast, we formulate a global method for MOTS over the space of assignments rather than detections: First, we find all top-k assignments of objects detected and segmented between any two consecutive frames and develop a structured prediction formulation to score assignment sequences across any number of consecutive frames. We use dynamic programming to find the global optimizer of this formulation in polynomial time. Second, we connect objects which reappear after having been out of view for some time. For this we formulate an assignment problem. On the challenging KITTI-MOTS and MOTSChallenge datasets, this achieves state-of-the-art results among methods which don't use depth data.

Text-based image retrieval has seen considerable progress in recent years. However, the performance of existing methods suffers in real life since the user is likely to provide an incomplete description of an image, which often leads to results filled with false positives that fit the incomplete description. In this work, we introduce the partial-query problem and extensively analyze its influence on text-based image retrieval. Previous interactive methods tackle the problem by passively receiving users' feedback to supplement the incomplete query iteratively, which is time-consuming and requires heavy user effort. Instead, we propose a novel retrieval framework that conducts the interactive process in an Ask-and-Confirm fashion, where AI actively searches for discriminative details missing in the current query, and users only need to confirm AI's proposal. Specifically, we propose an object-based interaction to make the interactive retrieval more user-friendly and present a reinforcement-learning-based policy to search for discriminative objects. Furthermore, since fully-supervised training is often infeasible due to the difficulty of obtaining human-machine dialog data, we present a weakly-supervised training strategy that needs no human-annotated dialogs other than a text-image dataset. Experiments show that our framework significantly improves the performance of text-based image retrieval. Code is available at https://github.com/CuthbertCai/Ask-Confirm.

Photorealistic image generation has reached a new level of quality due to the breakthroughs of generative adversarial networks (GANs). Yet, the dark side of such deepfakes, the malicious use of generated media, raises concerns about visual misinformation. While existing research work on deepfake detection demonstrates high accuracy, it is subject to advances in generation techniques and adversarial iterations on detection countermeasure techniques. Thus, we seek a proactive and sustainable solution on deepfake detection, that is agnostic to the evolution of generative models, by introducing artificial fingerprints into the models. Our approach is simple and effective. We first embed artificial fingerprints into training data, then validate a surprising discovery on the transferability of such fingerprints from training data to generative models, which in turn appears in the generated deepfakes. Experiments show that our fingerprinting solution (1) holds for a variety of cutting-edge generative models, (2) leads to a negligible side effect on generation quality, (3) stays robust against image-level and model-level perturbations, (4) stays hard to be detected by adversaries, and (5) converts deepfake detection and attribution into trivial tasks and outperforms the recent state-of-the-art baselines. Our solution closes the responsibility loop between publishing pre-trained generative model inventions and their possible misuses, which makes it independent of the current arms race.

Pseudo-LiDAR-based methods for monocular 3D object detection have received considerable attention in the community due to the performance gains exhibited on the KITTI3D benchmark, in particular on the commonly reported validation split. This generated a distorted impression about the superiority of Pseudo-LiDAR-based (PL-based) approaches over methods working with RGB images only. Our first contribution consists in rectifying this view by pointing out and showing experimentally that the validation results published by PL-based methods are substantially biased. The source of the bias resides in an overlap between the KITTI3D object detection validation set and the training/validation sets used to train depth predictors feeding PL-based methods. Surprisingly, the bias remains also after geographically removing the overlap. This leaves the test set as the only reliable set for comparison, where published PL-based methods do not excel. Our second contribution brings PL-based methods back up in the ranking with the design of a novel deep architecture which introduces a 3D confidence prediction module. We show that 3D confidence estimation techniques derived from RGB-only 3D detection approaches can be successfully integrated into our framework and, more importantly, that improved performance can be obtained with a newly designed 3D confidence measure, leading to state-of-the-art performance on the KITTI3D benchmark.

Understanding the inner workings of deep neural networks (DNNs) is essential to provide trustworthy artificial intelligence techniques for practical applications. Existing studies typically involve linking semantic concepts to units or layers of DNNs, but fail to explain the inference process. In this paper, we introduce neural architecture disentanglement (NAD) to fill the gap. Specifically, NAD learns to disentangle a pre-trained DNN into sub-architectures according to independent tasks, forming information flows that describe the inference processes. We investigate whether, where, and how the disentanglement occurs through experiments conducted with handcrafted and automatically-searched network architectures, on both object-based and scene-based datasets. Based on the experimental results, we present three new findings that provide fresh insights into the inner logic of DNNs. First, DNNs can be divided into sub-architectures for independent tasks. Second, deeper layers do not always correspond to higher semantics. Third, the connection type in a DNN affects how the information flows across layers, leading to different disentanglement behaviors. With NAD, we further explain why DNNs sometimes give wrong predictions. Experimental results show that misclassified images have a high probability of being assigned to task sub-architectures similar to the correct ones. Our code is available at https://github.com/hujiecpp/NAD.

We propose Anticipative Video Transformer (AVT), an end-to-end attention-based video modeling architecture that attends to the previously observed video in order to anticipate future actions. We train the model jointly to predict the next action in a video sequence, while also learning frame feature encoders that are predictive of successive future frames' features. Compared to existing temporal aggregation strategies, AVT has the advantage of both maintaining the sequential progression of observed actions while still capturing long-range dependencies--both critical for the anticipation task. Through extensive experiments, we show that AVT obtains the best reported performance on four popular action anticipation benchmarks: EpicKitchens-55, EpicKitchens-100, EGTEA Gaze+, and 50-Salads; and it wins first place in the EpicKitchens-100 CVPR'21 challenge.

In egocentric videos, the face of a wearer capturing the video is never captured. This gives a false sense of security that the wearer's privacy is preserved while sharing such videos. However, egocentric cameras are typically harnessed to wearer's head, and hence, also capture wearer's gait. Recent works have shown that wearer gait signatures can be extracted from egocentric videos, which can be used to determine if two egocentric videos have the same wearer. In a more damaging scenario, one can even recognize a wearer using hand gestures from egocentric videos, or identify a wearer in third person videos such as from a surveillance camera. We believe, this could be a death knell in sharing of egocentric videos, and fatal for egocentric vision research. In this work, we suggest a novel technique to anonymize egocentric videos, which create carefully crafted, but small, and imperceptible optical flow perturbations in an egocentric video's frames. Importantly, these perturbations do not affect object detection or action/activity recognition from egocentric videos but are strong enough to dis-balance the gait recovery process. In our experiments on benchmark \epic dataset, the proposed perturbation degrades the wearer recognition performance of [??], from 66.3% to 13.4%, while preserving the activity recognition performance of [??] from 89.6% to 87.4%. To test our anonymization with more wearer recognition techniques, we also developed a stronger, and more generalizable wearer recognition method based on camera egomotion cues. The approach achieves state-of-the-art (SOTA) performance of 59.67% on \epicns, compared to 55.06% by [??]. However, the accuracy of our recognition technique also drops to 12% using the proposed anonymizing perturbations.

This paper addresses the challenge of reconstructing an animatable human model from a multi-view video. Some recent works have proposed to decompose a non-rigidly deforming scene into a canonical neural radiance field and a set of deformation fields that map observation-space points to the canonical space, thereby enabling them to learn the dynamic scene from images. However, they represent the deformation field as translational vector field or SE(3) field, which makes the optimization highly under-constrained. Moreover, these representations cannot be explicitly controlled by input motions. Instead, we introduce neural blend weight fields to produce the deformation fields. Based on the skeleton-driven deformation, blend weight fields are used with 3D human skeletons to generate observation-to-canonical and canonical-to-observation correspondences. Since 3D human skeletons are more observable, they can regularize the learning of deformation fields. Moreover, the learned blend weight fields can be combined with input skeletal motions to generate new deformation fields to animate the human model. Experiments show that our approach significantly outperforms recent human synthesis methods. The code and supplementary materials are available at \href https://zju3dv.github.io/animatable_nerf/ https://zju3dv.github.io/animatable_nerf/ .

We propose 3DETR, an end-to-end Transformer based object detection model for 3D point clouds. Compared to existing detection methods that employ a number of 3D-specific inductive biases, 3DETR requires minimal modifications to the vanilla Transformer block. Specifically, we find that a standard Transformer with non-parametric queries and Fourier positional embeddings is competitive with specialized architectures that employ libraries of 3D-specific operators with hand-tuned hyperparameters. Nevertheless, 3DETR is conceptually simple and easy to implement, enabling further improvements by incorporating 3D domain knowledge. Through extensive experiments, we show 3DETR outperforms the well-established and highly optimized VoteNet baselines on the challenging ScanNetV2 dataset by 9.5%. Furthermore, we show 3DETR is applicable to 3D tasks beyond detection, and can serve as a building block for future research.

The state-of-the-art semi-supervised learning models are consistency-based which learn about unlabeled images by maximizing the similarity between different augmentations of an image. But when we apply the methods to human pose estimation which has extremely imbalanced class distribution, the models often collapse and predict every pixel in unlabeled images as background. This is because the decision boundary may pass through the high-density area of the minor class so more and more pixels are gradually mis-classified as the background class. In this work, we present a surprisingly simple approach to drive the model to learn in the correct direction. For each image, it composes a pair of easy and hard augmentations and uses the more accurate predictions on the easy image to teach the network to learn about the hard one. The accuracy superiority of teaching signals allows the network to be "monotonically" improved which effectively avoids collapsing. We apply our method to recent pose estimators and find that they achieve significantly better performances than their supervised counterparts on three public datasets.

This paper does not describe a novel method. Instead, it studies a straightforward, incremental, yet must-know baseline given the recent progress in computer vision: self-supervised learning for Vision Transformers (ViT). While the training recipes for standard convolutional networks have been highly mature and robust, the recipes for ViT are yet to be built, especially in the self-supervised scenarios where training becomes more challenging. In this work, we go back to basics and investigate the effects of several fundamental components for training self-supervised ViT. We observe that instability is a major issue that degrades accuracy, and it can be hidden by apparently good results. We reveal that these results are indeed partial failure, and they can be improved when training is made more stable. We benchmark ViT results in MoCo v3 and several other self-supervised frameworks, with ablations in various aspects. We discuss the currently positive evidence as well as challenges and open questions. We hope that this work will provide useful data points and experience for future research.

The minimal geodesic models based on the Eikonal equations are capable of finding suitable solutions in various image segmentation scenarios. Existing geodesic-based segmentation approaches usually exploit the image features in conjunction with geometric regularization terms (such as curve length or elastica length) for computing geodesic paths. In this paper, we consider a more complicated problem: finding simple and closed geodesic curves which are imposed a convexity shape prior. The proposed approach relies on an orientation-lifting strategy, by which a planar curve can be mapped to an high-dimensional orientation space. The convexity shape prior serves as a constraint for the construction of local metrics. The geodesic curves in the lifted space then can be efficiently computed through the fast marching method. In addition, we introduce a way to incorporate region-based homogeneity features into the proposed geodesic model so as to solve the region-based segmentation issues with shape prior constraints.

Event cameras are ideally suited to capture HDR visual information without blur but perform poorly on static or slowly changing scenes. Conversely, conventional image sensors measure absolute intensity of slowly changing scenes effectively but do poorly on high dynamic range or quickly changing scenes. In this paper, we present an event-based video reconstruction pipeline for High Dynamic Range (HDR) scenarios. The proposed algorithm includes a frame augmentation pre-processing step that deblurs and temporally interpolates frame data using events. The augmented frame and event data are then fused using a novel asynchronous Kalman filter under a unifying uncertainty model for both sensors. Our experimental results are evaluated on both publicly available datasets with challenging lighting conditions and fast motions and our new dataset with HDR reference. The proposed algorithm outperforms state-of-the-art methods in both absolute intensity error (48% reduction) and image similarity indexes (average 11% improvement).

Recently, the generalization behavior of Convolutional Neural Networks (CNN) is gradually transparent through explanation techniques with the frequency components decomposition. However, the importance of the phase spectrum of the image for a robust vision system is still ignored. In this paper, we notice that the CNN tends to converge at the local optimum which is closely related to the high-frequency components of the training images, while the amplitude spectrum is easily disturbed such as noises or common corruptions. In contrast, more empirical studies found that humans rely on more phase components to achieve robust recognition. This observation leads to more explanations of the CNN's generalization behaviors in both robustness to common perturbations and out-of-distribution detection, and motivates a new perspective on data augmentation designed by re-combing the phase spectrum of the current image and the amplitude spectrum of the distracter image. That is, the generated samples force the CNN to pay more attention to the structured information from phase components and keep robust to the variation of the amplitude. Experiments on several image datasets indicate that the proposed method achieves state-of-the-art performances on multiple generalizations and calibration tasks, including adaptability for common corruptions and surface variations, out-of-distribution detection, and adversarial attack.

Modern computer vision applications suffer from catastrophic forgetting when incrementally learning new concepts over time. The most successful approaches to alleviate this forgetting require extensive replay of previously seen data, which is problematic when memory constraints or data legality concerns exist. In this work, we consider the high-impact problem of Data-Free Class-Incremental Learning (DFCIL), where an incremental learning agent must learn new concepts over time without storing generators or training data from past tasks. One approach for DFCIL is to replay synthetic images produced by inverting a frozen copy of the learner's classification model, but we show this approach fails for common class-incremental benchmarks when using standard distillation strategies. We diagnose the cause of this failure and propose a novel incremental distillation strategy for DFCIL, contributing a modified cross-entropy training and importance-weighted feature distillation, and show that our method results in up to a 25.1% increase in final task accuracy (absolute difference) compared to SOTA DFCIL methods for common class-incremental benchmarks. Our method even outperforms several standard replay based methods which store a coreset of images.

The goal of this work is to temporally align asynchronous subtitles in sign language videos. In particular, we focus on sign-language interpreted TV broadcast data comprising (i) a video of continuous signing, and (ii) subtitles corresponding to the audio content. Previous work exploiting such weakly-aligned data only considered finding keyword-sign correspondences, whereas we aim to localise a complete subtitle text in continuous signing. We propose a Transformer architecture tailored for this task, which we train on manually annotated alignments covering over 15K subtitles that span 17.7 hours of video. We use BERT subtitle embeddings and CNN video representations learned for sign recognition to encode the two signals, which interact through a series of attention layers. Our model outputs frame-level predictions, i.e., for each video frame, whether it belongs to the queried subtitle or not. Through extensive evaluations, we show substantial improvements over existing alignment baselines that do not make use of subtitle text embeddings for learning. Our automatic alignment model opens up possibilities for advancing machine translation of sign languages via providing continuously synchronized video-text data.

In this work, we develop a method to generate infinite high-resolution images with diverse and complex content. It is based on a perfectly equivariant patch-wise generator with synchronous interpolations in the image and latent spaces. Latent codes, when sampled, are positioned on the coordinate grid, and each pixel is computed from an interpolation of the neighboring codes. We modify the AdaIN mechanism to work in such a setup and train a GAN model to generate images positioned between any two latent vectors. At test time, this allows for generating infinitely large images of diverse scenes that transition naturally from one into another. Apart from that, we introduce LHQ: a new dataset of 90k high-resolution nature landscapes. We test the approach on LHQ, LSUN Tower and LSUN Bridge and outperform the baselines by at least 4 times in terms of quality and diversity of the produced infinite images. The project website is located at https://universome.github.io/alis.

Vision-and-language navigation (VLN) aims to enable embodied agents to navigate in realistic environments using natural language instructions. Given the scarcity of domain-specific training data and the high diversity of image and language inputs, the generalization of VLN agents to unseen environments remains challenging. Recent methods explore pretraining to improve generalization, however, the use of generic image-caption datasets or existing small-scale VLN environments is suboptimal and results in limited improvements. In this work, we introduce BnB, a large-scale and diverse in-domain VLN dataset. We first collect image-caption (IC) pairs from hundreds of thousands of listings from online rental marketplaces. Using IC pairs we next propose automatic strategies to generate millions of VLN path-instruction (PI) pairs. We further propose a shuffling loss that improves the learning of temporal order inside PI pairs. We use BnB to pretrain our Airbert model that can be adapted to discriminative and generative settings and show that it outperforms state of the art for Room-to-Room (R2R) navigation and Remote Referring Expression (REVERIE) benchmarks. Moreover, our in-domain pretraining significantly increases performance on a challenging few-shot VLN evaluation, where we train the model only on VLN instructions from a few houses.

The decision-based black-box attack means to craft adversarial examples with only the top-1 label of the victim model available. A common practice is to start from a large perturbation and then iteratively reduce it with a deterministic direction and a random one while keeping it adversarial. The limited information obtained from each query and inefficient direction sampling impede attack efficiency, making it hard to obtain a small enough perturbation within a limited number of queries. To tackle this problem, we propose a novel attack method termed Adaptive History-driven Attack (AHA) which gathers information from all historical queries as the prior for current sampling. Moreover, to balance between the deterministic direction and the random one, we dynamically adjust the coefficient according to the ratio of the actual magnitude reduction to the expected one. Such a strategy improves the success rate of queries during optimization, letting adversarial examples move swiftly along the decision boundary. Our method can also integrate with subspace optimization like dimension reduction to further improve efficiency. Extensive experiments on both ImageNet and CelebA datasets demonstrate that our method achieves at least 24.3% lower magnitude of perturbation on average with the same number of queries. Finally, we prove the practical potential of our method by evaluating it on popular defense methods and a real-world system provided by MEGVII Face++.