ICCV 2021 Papers

Generative models for 3D shapes represented by hierarchies of parts can generate realistic and diverse sets of outputs. However, existing models suffer from the key practical limitation of modelling shapes holistically and thus cannot perform conditional sampling, i.e. they are not able to generate variants on individual parts of generated shapes without modifying the rest of the shape. This is limiting for applications such as 3D CAD design that involve adjusting created shapes at multiple levels of detail. To address this, we introduce LSD-StructureNet, an augmentation to the StructureNet architecture that enables re-generation of parts situated at arbitrary positions in the hierarchies of its outputs. We achieve this by learning individual, probabilistic conditional decoders for each hierarchy depth. We evaluate LSD-StructureNet on the PartNet dataset, the largest dataset of 3D shapes represented by hierarchies of parts. Our results show that contrarily to existing methods, LSD-StructureNet can perform conditional sampling without impacting inference speed or the realism and diversity of its outputs.

Recent advances in trajectory prediction have shown that explicit reasoning about agents' intent is important to accurately forecast their motion. However, the current research activities are not directly applicable to intelligent and safety critical systems. This is mainly because very few public datasets are available, and they only consider pedestrian-specific intents for a short temporal horizon from a restricted egocentric view. To this end, we propose LOKI (LOng term and Key Intentions), a novel large-scale dataset that is designed to tackle joint trajectory and intention prediction for heterogeneous traffic agents (pedestrians and vehicles) in an autonomous driving setting. The LOKI dataset is created to discover several factors that may affect intention, including i) agent's own will, ii) social interactions, iii) environmental constraints, and iv) contextual information. We also propose a model that jointly performs trajectory and intention prediction, showing that recurrently reasoning about intention can assist with trajectory prediction. We show our method outperforms state-of-the-art trajectory prediction methods by upto 27% and also provide a baseline for frame-wise intention estimation. The dataset is available at https://usa.honda-ri.com/loki

Recently, machine learning models have demonstrated to be vulnerable to backdoor attacks, primarily due to the lack of transparency in black-box models such as deep neural networks. A third-party model can be poisoned such that it works adequately in normal conditions but behaves maliciously on samples with specific trigger patterns. However, the trigger injection function is manually defined in most existing backdoor attack methods, e.g., placing a small patch of pixels on an image or slightly deforming the image before poisoning the model. This results in a two-stage approach with a sub-optimal attack success rate and a lack of complete stealthiness under human inspection. In this paper, we propose a novel and stealthy backdoor attack framework, LIRA, which jointly learns the optimal, stealthy trigger injection function and poisons the model. We formulate such an objective as a non-convex, constrained optimization problem. Under this optimization framework, the trigger generator function will learn to manipulate the input with imperceptible noise to preserve the model performance on the clean data and maximize the attack success rate on the poisoned data. Then, we solve this challenging optimization problem with an efficient, two-stage stochastic optimization procedure. Finally, the proposed attack framework achieves 100% success rates in several benchmark datasets, including MNIST, CIFAR10, GTSRB, and T-ImageNet, while simultaneously bypassing existing backdoor defense methods and human inspection.

Stereo-based 3D detection aims at detecting 3D object bounding boxes from stereo images using intermediate depth maps or implicit 3D geometry representations, which provides a low-cost solution for 3D perception. However, its performance is still inferior compared with LiDAR-based detection algorithms. To detect and localize accurate 3D bounding boxes, LiDAR-based models can encode accurate object boundaries and surface normal directions from LiDAR point clouds. However, the detection results of stereo-based detectors are easily affected by the erroneous depth features due to the limitation of stereo matching. To solve the problem, we propose LIGA-Stereo (LiDAR Geometry Aware Stereo Detector) to learn stereo-based 3D detectors under the guidance of high-level geometry-aware representations of LiDAR-based detection models. In addition, we found existing voxel-based stereo detectors failed to learn semantic features effectively from indirect 3D supervisions. We attach an auxiliary 2D detection head to provide direct 2D semantic supervisions. Experiment results show that the above two strategies improved the geometric and semantic representation capabilities. Compared with the state-of-the-art stereo detector, our method has improved the 3D detection performance of cars, pedestrians, cyclists by 10.44%, 5.69%, 5.97% mAP respectively on the official KITTI benchmark. The gap between stereo-based and LiDAR-based 3D detectors is further narrowed.

Class Activation Mapping (CAM) is a powerful technique used to understand the decision making of Convolutional Neural Network (CNN) in computer vision. Recently, there have been attempts not only to generate better visual explanations, but also to improve classification performance using visual explanations. However, previous works still have their own drawbacks. In this paper, we propose a novel architecture, LFI-CAM***(Learning Feature Importance Class Activation Mapping), which is trainable for image classification and visual explanation in an end-to-end manner. LFI-CAM generates attention map for visual explanation during forward propagation, and simultaneously uses attention map to improve classification performance through the attention mechanism. Feature Importance Network (FIN) focuses on learning the feature importance instead of directly learning the attention map to obtain a more reliable and consistent attention map. We confirmed that LFI-CAM is optimized not only by learning the feature importance but also by enhancing the backbone feature representation to focus more on important features of the input image. Experiments show that LFI-CAM outperforms baseline models' accuracy on classification tasks as well as significantly improves on previous works in terms of attention map quality and stability over different hyper-parameters.

A variety of effective face-swap and face-reenactment methods have been publicized in recent years, democratizing the face synthesis technology to a great extent. Videos generated as such have come to be called deepfakes with a negative connotation, for various social problems they have caused. Facing the emerging threat of deepfakes, we have built the Korean DeepFake Detection Dataset (KoDF), a large-scale collection of synthesized and real videos focused on Korean subjects. In this paper, we provide a detailed description of methods used to construct the dataset, experimentally show the discrepancy between the distributions of KoDF and existing deepfake detection datasets, and underline the importance of using multiple datasets for real-world generalization. KoDF is publicly available at https://moneybrain-research.github.io/kodf in its entirety (i.e. real clips, synthesized clips, clips with adversarial attack, and metadata).

Model inversion (MI) attacks are aimed at reconstructing training data from model parameters. Such attacks have triggered increasing concerns about privacy, especially given the growing number of online model repositories. However, existing MI attacks against deep neural networks (DNNs) have a large room for performance improvement. We present a novel inversion-specific GAN that can better distill knowledge useful for performing attacks on private models from public data. In particular, we train the discriminator to differentiate not only the real and fake samples but the soft-labels provided by the target model. Moreover, unlike previous work that directly searches for a single data point to represent a target class, we propose to model a private data distribution for each target class. Our experiments show that the combination of these techniques can significantly boost the success rate of the state-of-the-art MI attacks by 150%, and generalize better to a variety of datasets and models. Our code is available at https://github.com/SCccc21/Knowledge-Enriched-DMI.

With the thriving of deep learning, CNN-based object detectors have made great progress in the past decade. However, the domain gap between training and testing data leads to a prominent performance degradation and thus hinders their application in the real world. To alleviate this problem, Knowledge Transfer Network (KTNet) is proposed as a new paradigm for domain adaption. Specifically, KTNet is constructed on a base detector with intrinsic knowledge mining and relational knowledge constraints. First, we design a foreground/background classifier shared by source domain and target domain to extract the common attribute knowledge of objects in different scenarios. Second, we model the relational knowledge graph and explicitly constrain the consistency of category correlation under source domain, target domain, as well as cross-domain conditions. As a result, the detector is guided to learn object-related and domain-independent representation. Extensive experiments and visualizations confirm that transferring object-specific knowledge can yield notable performance gains. The proposed KTNet achieves state-of-the-art results on three cross-domain detection benchmarks.

NeRF synthesizes novel views of a scene with unprecedented quality by fitting a neural radiance field to RGB images. However, NeRF requires querying a deep Multi-Layer Perceptron (MLP) millions of times, leading to slow rendering times, even on modern GPUs. In this paper, we demonstrate that real-time rendering is possible by utilizing thousands of tiny MLPs instead of one single large MLP. In our setting, each individual MLP only needs to represent parts of the scene, thus smaller and faster-to-evaluate MLPs can be used. By combining this divide-and-conquer strategy with further optimizations, rendering is accelerated by three orders of magnitude compared to the original NeRF model without incurring high storage costs. Further, using teacher-student distillation for training, we show that this speed-up can be achieved without sacrificing visual quality.

We present a fast bottom-up method that jointly detects over 100 keypoints on humans or objects, also referred to as human/object pose estimation. We model all keypoints belonging to a human or an object --the pose-- as a graph and leverage insights from community detection to quantify the independence of keypoints. We use a graph centrality measure to assign training weights to different parts of a pose. Our proposed measure quantifies how tightly a keypoint is connected to its neighborhood. Our experiments show that our method outperforms all previous methods for human pose estimation with fine-grained keypoint annotations on the face, the hands and the feet with a total of 133 keypoints. We also show that our method generalizes to car poses.

Embedding data in hyperbolic spaces has proven beneficial for many advanced machine learning applications such as image classification and word embeddings. However, working in hyperbolic spaces is not without difficulties as a result of its curved geometry (e.g., computing the Frechet mean of a set of points requires an iterative algorithm). Furthermore, in Euclidean spaces, one can resort to kernel machines that not only enjoy rich theoretical properties but that can also lead to superior representational power (e.g., infinite-width neural networks). In this paper, we introduce positive definite kernel functions for hyperbolic spaces. This brings in two major advantages, 1. kernelization will pave the way to seamlessly benefit from kernel machines in conjunction with hyperbolic embeddings, and 2. the rich structure of the Hilbert spaces associated with kernel machines enables us to simplify various operations involving hyperbolic data. That said, identifying valid kernel functions on curved spaces is not straightforward and is indeed considered an open problem in the learning community. Our work addresses this gap and develops several valid positive definite kernels in hyperbolic spaces, including the universal ones (e.g., RBF). We comprehensively study the proposed kernels on a variety of challenging tasks including few-shot learning, zero-shot learning, person re-identification and knowledge distillation, showing the superiority of the kernelization for hyperbolic representations.

The class activation mapping, or CAM, has been the cornerstone of feature attribution methods for multiple vision tasks. Its simplicity and effectiveness have led to wide applications in the explanation of visual predictions and weakly-supervised localization tasks. However, CAM has its own shortcomings. The computation of attribution maps relies on ad-hoc calibration steps that are not part of the training computational graph, making it difficult for us to understand the real meaning of the attribution values. In this paper, we improve CAM by explicitly incorporating a latent variable encoding the location of the cue for recognition in the formulation, thereby subsuming the attribution map into the training computational graph. The resulting model, class activation latent mapping, or CALM, is trained with the expectation-maximization algorithm. Our experiments show that CALM identifies discriminative attributes for image classifiers more accurately than CAM and other visual attribution baselines. CALM also shows performance improvements over prior arts on the weakly-supervised object localization benchmarks. Our code is available at https://github.com/naver-ai/calm.

While learning-based multi-view stereo (MVS) methods have recently shown successful performances in quality and efficiency, limited MVS data hampers generalization to unseen environments. A simple solution is to generate various large-scale MVS datasets, but generating dense ground truth for 3D structure requires a huge amount of time and resources. On the other hand, if the reliance on dense ground truth is relaxed, MVS systems will generalize more smoothly to new environments. To this end, we first introduce a novel semi-supervised multi-view stereo framework called a Sparse Ground truth-based MVS Network (SGT-MVSNet) that can reliably reconstruct the 3D structures even with a few ground truth 3D points. Our strategy is to divide the accurate and erroneous regions and individually conquer them based on our observation that a probability map can separate these regions. We propose a self-supervision loss called the 3D Point Consistency Loss to enhance the 3D reconstruction performance, which forces the 3D points back-projected from the corresponding pixels by the predicted depth values to meet at the same 3D coordinates. Finally, we propagate these improved depth predictions toward edges and occlusions by the Coarse-to-fine Reliable Depth Propagation module. We generate the spare ground truth of the DTU dataset for evaluation and extensive experiments verify that our SGT-MVSNet outperforms the state-of-the-art MVS methods on the sparse ground truth setting. Moreover, our method shows comparable reconstruction results to the supervised MVS methods though we only used tens and hundreds of ground truth 3D points.

The video-based action recognition task has been extensively studied in recent years. In this paper, we study the structural vulnerability of deep learning-based action recognition models against the adversarial attack using the one frame attack that adds an inconspicuous perturbation to only a single frame of a given video clip. Our analysis shows that the models are highly vulnerable against the one frame attack due to their structural properties. Experiments demonstrate high fooling rates and inconspicuous characteristics of the attack. Furthermore, we show that strong universal one frame perturbations can be obtained under various scenarios. Our work raises the serious issue of adversarial vulnerability of the state-of-the-art action recognition models in various perspectives.

Recent methods for visual question answering rely on large-scale annotated datasets. Manual annotation of questions and answers for videos, however, is tedious, expensive and prevents scalability. In this work, we propose to avoid manual annotation and generate a large-scale training dataset for video question answering making use of automatic cross-modal supervision. We leverage a question generation transformer trained on text data and use it to generate question-answer pairs from transcribed video narrations. Given narrated videos, we then automatically generate the HowToVQA69M dataset with 69M video-question-answer triplets. To handle the open vocabulary of diverse answers in this dataset, we propose a training procedure based on a contrastive loss between a video-question multi-modal transformer and an answer transformer. We introduce the zero-shot VideoQA task and show excellent results, in particular for rare answers. Furthermore, we demonstrate our method to significantly outperform the state of the art on MSRVTT-QA, MSVD-QA, ActivityNet-QA and How2QA. Finally, for a detailed evaluation we introduce iVQA, a new VideoQA dataset with reduced language biases and high-quality redundant manual annotations.

In video highlight detection, the goal is to identify the interesting moments within an unedited video. Although the audio component of the video provides important cues for highlight detection, the majority of existing efforts focus almost exclusively on the visual component. In this paper, we argue that both audio and visual components of a video should be modeled jointly to retrieve its best moments. To this end, we propose an audio-visual network for video highlight detection. At the core of our approach lies a bimodal attention mechanism, which captures the interaction between the audio and visual components of a video, and produces fused representations to facilitate highlight detection. Furthermore, we introduce a noise sentinel technique to adaptively discount a noisy visual or audio modality. Empirical evaluations on two benchmark datasets demonstrate the superior performance of our approach over the state-of-the-art methods.

Although text recognition has significantly evolved over the years, state-of the-art (SOTA) models still struggle in the wild scenarios due to complex backgrounds, varying fonts, uncontrolled illuminations, distortions and other artifacts. This is because such models solely depend on visual information for text recognition, thus lacking semantic reasoning capabilities. In this paper, we argue that semantic information offers a complementary role in addition to visual only. More specifically, we additionally utilize semantic information by proposing a multi-stage multi-scale attentional decoder that performs joint visual-semantic reasoning. Our novelty lies in the intuition that for text recognition, prediction should be refined in a stage-wise manner. Therefore our key contribution is in designing a stage-wise unrolling attentional decoder where non-differentiability, invoked by discretely predicted character labels, needs to be bypassed for end-to-end training. While the first stage predicts using visual features, subsequent stages refine on-top of it using joint visual-semantic information. Additionally, we introduce multi-scale 2D attention along with dense and residual connections between different stages to deal with varying scales of character sizes, for better performance and faster convergence during training. Experimental results show our approach to outperform existing SOTA methods by a considerable margin.

Curvilinear structure segmentation (CSS) is under semantic segmentation, whose applications include crack detection, aerial road extraction, and biomedical image segmentation. In general, geometric topology and pixel-wise features are two critical aspects of CSS. However, most semantic segmentation methods only focus on enhancing feature representations while existing CSS techniques emphasize preserving topology alone. In this paper, we present a Joint Topology-preserving and Feature-refinement Network (JTFN) that jointly models global topology and refined features based on an iterative feedback learning strategy. Specifically, we explore the structure of objects to help preserve corresponding topologies of predicted masks, thus design a reciprocative two-stream module for CSS and boundary detection. In addition, we introduce such topology-aware predictions as feedback guidance that refines attentive features by supplementing and enhancing saliencies. To the best of our knowledge, this is the first work that jointly addresses topology preserving and feature refinement for CSS. We evaluate JTFN on four datasets of diverse applications: Crack500, CrackTree200, Roads, and DRIVE. Results show that JTFN performs best in comparison with alternative methods. Code is available.

This paper studies the problem of novel category discovery on single- and multi-modal data with labels from different but relevant categories. We present a generic, end-to-end framework to jointly learn a reliable representation and assign clusters to unlabelled data. To avoid over-fitting the learnt embedding to labelled data, we take inspiration from self-supervised representation learning by noise-contrastive estimation and extend it to jointly handle labelled and unlabelled data. In particular, we propose using category discrimination on labelled data and cross-modal discrimination on multi-modal data to augment instance discrimination used in conventional contrastive learning approaches. We further employ Winner-Take-All (WTA) hashing algorithm on the shared representation space to generate pairwise pseudo labels for unlabelled data to better predict cluster assignments. We thoroughly evaluate our framework on large-scale multi-modal video benchmarks Kinetics-400 and VGG-Sound, and image benchmarks CIFAR10, CIFAR100 and ImageNet, obtaining state-of-the-art results.