Total: 416
Deep networks should be robust to rare events if they are to be successfully deployed in high-stakes real-world applications. Here we study the capability of deep networks to recognize objects in unusual poses. We create a synthetic dataset of images of objects in unusual orientations, and evaluate the robustness of a collection of 38 recent and competitive deep networks for image classification. We show that classifying these images is still a challenge for all networks tested, with an average accuracy drop of 29.5% compared to when the objects are presented upright. This brittleness is largely unaffected by various design choices, such as training losses, architectures, dataset modalities, and data-augmentation schemes. However, networks trained on very large datasets substantially outperform others, with the best network tested—Noisy Student trained on JFT-300M—showing a relatively small accuracy drop of only 14.5% on unusual poses. Nevertheless, a visual inspection of the failures of Noisy Student reveals a remaining gap in robustness with humans. Furthermore, combining multiple object transformations—3D-rotations and scaling—further degrades the performance of all networks. Our results provide another measurement of the robustness of deep networks to consider when using them in the real world. Code and datasets are available at https://github.com/amro-kamal/ObjectPose.
Improperly constructed datasets can result in inaccurate inferences. For instance, models trained on biased datasets perform poorly in terms of generalization (i.e., dataset bias). Recent debiasing techniques have successfully achieved generalization performance by underestimating easy-to-learn samples (i.e., bias-aligned samples) and highlighting difficult-to-learn samples (i.e., bias-conflicting samples). However, these techniques may fail owing to noisy labels, because the trained model recognizes noisy labels as difficult-to-learn and thus highlights them. In this study, we find that earlier approaches that used the provided labels to quantify difficulty could be affected by the small proportion of noisy labels. Furthermore, we find that running denoising algorithms before debiasing is ineffective because denoising algorithms reduce the impact of difficult-to-learn samples, including valuable bias-conflicting samples. Therefore, we propose an approach called denoising after entropy-based debiasing, i.e., DENEB, which has three main stages. (1) The prejudice model is trained by emphasizing (bias-aligned, clean) samples, which are selected using a Gaussian Mixture Model. (2) Using the per-sample entropy from the output of the prejudice model, the sampling probability of each sample that is proportional to the entropy is computed. (3) The final model is trained using existing denoising algorithms with the mini-batches constructed by following the computed sampling probability. Compared to existing debiasing and denoising algorithms, our method achieves better debiasing performance on multiple benchmarks.
Saliency methods provide post-hoc model interpretation by attributing input features to the model outputs. Current methods mainly achieve this using a single input sample, thereby failing to answer input-independent inquiries about the model. We also show that input-specific saliency mapping is intrinsically susceptible to misleading feature attribution. Current attempts to use `general' input features for model interpretation assume access to a dataset containing those features, which biases the interpretation. Addressing the gap, we introduce a new perspective of input-agnostic saliency mapping that computationally estimates the high-level features attributed by the model to its outputs. These features are geometrically correlated, and are computed by accumulating model's gradient information with respect to an unrestricted data distribution. To compute these features, we nudge independent data points over the model loss surface towards the local minima associated by a human-understandable concept, e.g., class label for classifiers. With a systematic projection, scaling and refinement process, this information is transformed into an interpretable visualization without compromising its model-fidelity. The visualization serves as a stand-alone qualitative interpretation. With an extensive evaluation, we not only demonstrate successful visualizations for a variety of concepts for large-scale models, but also showcase an interesting utility of this new form of saliency mapping by identifying backdoor signatures in compromised classifiers.
Self-supervised monocular depth estimation has been widely studied recently. Most of the work has focused on improving performance on benchmark datasets, such as KITTI, but has offered a few experiments on generalization performance. In this paper, we investigate the backbone networks (e.g., CNNs, Transformers, and CNN-Transformer hybrid models) toward the generalization of monocular depth estimation. We first evaluate state-of-the-art models on diverse public datasets, which have never been seen during the network training. Next, we investigate the effects of texture-biased and shape-biased representations using the various texture-shifted datasets that we generated. We observe that Transformers exhibit a strong shape bias and CNNs do a strong texture-bias. We also find that shape-biased models show better generalization performance for monocular depth estimation compared to texture-biased models. Based on these observations, we newly design a CNN-Transformer hybrid network with a multi-level adaptive feature fusion module, called MonoFormer. The design intuition behind MonoFormer is to increase shape bias by employing Transformers while compensating for the weak locality bias of Transformers by adaptively fusing multi-level representations. Extensive experiments show that the proposed method achieves state-of-the-art performance with various public datasets. Our method also shows the best generalization ability among the competitive methods.
Contrastive loss has significantly improved performance in supervised classification tasks by using a multi-viewed framework that leverages augmentation and label information. The augmentation enables contrast with another view of a single image but enlarges training time and memory usage. To exploit the strength of multi-views while avoiding the high computation cost, we introduce a multi-exit architecture that outputs multiple features of a single image in a single-viewed framework. To this end, we propose Self-Contrastive (SelfCon) learning, which self-contrasts within multiple outputs from the different levels of a single network. The multi-exit architecture efficiently replaces multi-augmented images and leverages various information from different layers of a network. We demonstrate that SelfCon learning improves the classification performance of the encoder network, and empirically analyze its advantages in terms of the single-view and the sub-network. Furthermore, we provide theoretical evidence of the performance increase based on the mutual information bound. For ImageNet classification on ResNet-50, SelfCon improves accuracy by +0.6% with 59% memory and 48% time of Supervised Contrastive learning, and a simple ensemble of multi-exit outputs boosts performance up to +1.5%. Our code is available at https://github.com/raymin0223/self-contrastive-learning.
We present a novel hierarchical modeling method for layout representation learning, the core of design documents (e.g., user interface, poster, template). Existing works on layout representation often ignore element hierarchies, which is an important facet of layouts, and mainly rely on the spatial bounding boxes for feature extraction. This paper proposes a Spatial-Structural Hierarchical Auto-Encoder (SSH-AE) that learns hierarchical representation by treating a hierarchically annotated layout as a tree format. On the one side, we model SSH-AE from both spatial (semantic views) and structural (organization and relationships) perspectives, which are two complementary aspects to represent a layout. On the other side, the semantic/geometric properties are associated at multiple resolutions/granularities, naturally handling complex layouts. Our learned representations are used for effective layout search from both spatial and structural similarity perspectives. We also newly involve the tree-edit distance (TED) as an evaluation metric to construct a comprehensive evaluation protocol for layout similarity assessment, which benefits a systematic and customized layout search. We further present a new dataset of POSTER layouts which we believe will be useful for future layout research. We show that our proposed SSH-AE outperforms the existing methods achieving state-of-the-art performance on two benchmark datasets. Code is available at github.com/yueb17/SSH-AE.
This paper for the first time explores audio-visual event localization in an unsupervised manner. Previous methods tackle this problem in a supervised setting and require segment-level or video-level event category ground-truth to train the model. However, building large-scale multi-modality datasets with category annotations is human-intensive and thus not scalable to real-world applications. To this end, we propose cross-modal label contrastive learning to exploit multi-modal information among unlabeled audio and visual streams as self-supervision signals. At the feature representation level, multi-modal representations are collaboratively learned from audio and visual components by using self-supervised representation learning. At the label level, we propose a novel self-supervised pretext task i.e. label contrasting to self-annotate videos with pseudo-labels for localization model training. Note that irrelevant background would hinder the acquisition of high-quality pseudo-labels and thus lead to an inferior localization model. To address this issue, we then propose an expectation-maximization algorithm that optimizes the pseudo-label acquisition and localization model in a coarse-to-fine manner. Extensive experiments demonstrate that our unsupervised approach performs reasonably well compared to the state-of-the-art supervised methods.
Robotic agents performing domestic chores by natural language directives are required to master the complex job of navigating environment and interacting with objects in the environments. The tasks given to the agents are often composite thus are challenging as completing them require to reason about multiple subtasks, e.g., bring a cup of coffee. To address the challenge, we propose to divide and conquer it by breaking the task into multiple subgoals and attend to them individually for better navigation and interaction. We call it Multi-level Compositional Reasoning Agent (MCR-Agent). Specifically, we learn a three-level action policy. At the highest level, we infer a sequence of human-interpretable subgoals to be executed based on language instructions by a high-level policy composition controller. At the middle level, we discriminatively control the agent’s navigation by a master policy by alternating between a navigation policy and various independent interaction policies. Finally, at the lowest level, we infer manipulation actions with the corresponding object masks using the appropriate interaction policy. Our approach not only generates human interpretable subgoals but also achieves 2.03% absolute gain to comparable state of the arts in the efficiency metric (PLWSR in unseen set) without using rule-based planning or a semantic spatial memory. The code is available at https://github.com/yonseivnl/mcr-agent.
For image local forgery detection, the existing methods require a large amount of labeled data for training, and most of them cannot detect multiple types of forgery simultaneously. In this paper, we firstly analyzed the JPEG compression traces which are mainly caused by different JPEG compression chains, and designed a trace extractor to learn such traces. Then, we utilized the trace extractor as the backbone and trained self-supervised to strengthen the discrimination ability of learned traces. With its benefits, regions with different JPEG compression chains can easily be distinguished within a forged image. Furthermore, our method does not rely on a large amount of training data, and even does not require any forged images for training. Experiments show that the proposed method can detect image local forgery on different datasets without re-training, and keep stable performance over various types of image local forgery.
A core process in human cognition is analogical mapping: the ability to identify a similar relational structure between different situations. We introduce a novel task, Visual Analogies of Situation Recognition, adapting the classical word-analogy task into the visual domain. Given a triplet of images, the task is to select an image candidate B' that completes the analogy (A to A' is like B to what?). Unlike previous work on visual analogy that focused on simple image transformations, we tackle complex analogies requiring understanding of scenes. We leverage situation recognition annotations and the CLIP model to generate a large set of 500k candidate analogies. Crowdsourced annotations for a sample of the data indicate that humans agree with the dataset label ~80% of the time (chance level 25%). Furthermore, we use human annotations to create a gold-standard dataset of 3,820 validated analogies. Our experiments demonstrate that state-of-the-art models do well when distractors are chosen randomly (~86%), but struggle with carefully chosen distractors (~53%, compared to 90% human accuracy). We hope our dataset will encourage the development of new analogy-making models. Website: https://vasr-dataset.github.io/
Designing a point cloud upsampler, which aims to generate a clean and dense point cloud given a sparse point representation, is a fundamental and challenging problem in computer vision. A line of attempts achieves this goal by establishing a point-to-point mapping function via deep neural networks. However, these approaches are prone to produce outlier points due to the lack of explicit surface-level constraints. To solve this problem, we introduce a novel surface regularizer into the upsampler network by forcing the neural network to learn the underlying parametric surface represented by bicubic functions and rotation functions, where the new generated points are then constrained on the underlying surface. These designs are integrated into two different networks for two tasks that take advantages of upsampling layers -- point cloud upsampling and point cloud completion for evaluation. The state-of-the-art experimental results on both tasks demonstrate the effectiveness of the proposed method. The implementation code will be available at https://github.com/corecai163/PSCU.
Cross-domain crowd counting has shown progressively improved performance. However, most methods fail to explicitly consider the transferability of different features between source and target domains. In this paper, we propose an innovative explicit Invariant Feature induced Cross-domain Knowledge Transformation framework to address the inconsistent domain-invariant features of different domains. The main idea is to explicitly extract domain-invariant features from both source and target domains, which builds a bridge to transfer more rich knowledge between two domains. The framework consists of three parts, global feature decoupling (GFD), relation exploration and alignment (REA), and graph-guided knowledge enhancement (GKE). In the GFD module, domain-invariant features are efficiently decoupled from domain-specific ones in two domains, which allows the model to distinguish crowds features from backgrounds in the complex scenes. In the REA module both inter-domain relation graph (Inter-RG) and intra-domain relation graph (Intra-RG) are built. Specifically, Inter-RG aggregates multi-scale domain-invariant features between two domains and further aligns local-level invariant features. Intra-RG preserves taskrelated specific information to assist the domain alignment. Furthermore, GKE strategy models the confidence of pseudolabels to further enhance the adaptability of the target domain. Various experiments show our method achieves state-of-theart performance on the standard benchmarks. Code is available at https://github.com/caiyiqing/IF-CKT.
Image harmonization aims to produce visually harmonious composite images by adjusting the foreground appearance to be compatible with the background. When the composite image has photographic foreground and painterly background, the task is called painterly image harmonization. There are only few works on this task, which are either time-consuming or weak in generating well-harmonized results. In this work, we propose a novel painterly harmonization network consisting of a dual-domain generator and a dual-domain discriminator, which harmonizes the composite image in both spatial domain and frequency domain. The dual-domain generator performs harmonization by using AdaIN modules in the spatial domain and our proposed ResFFT modules in the frequency domain. The dual-domain discriminator attempts to distinguish the inharmonious patches based on the spatial feature and frequency feature of each patch, which can enhance the ability of generator in an adversarial manner. Extensive experiments on the benchmark dataset show the effectiveness of our method. Our code and model are available at https://github.com/bcmi/PHDNet-Painterly-Image-Harmonization.
Automatic medical report generation is an essential task in applying artificial intelligence to the medical domain, which can lighten the workloads of doctors and promote clinical automation. The state-of-the-art approaches employ Transformer-based encoder-decoder architectures to generate reports for medical images. However, they do not fully explore the relationships between multi-modal medical data, and generate inaccurate and inconsistent reports. To address these issues, this paper proposes a Multi-modal Memory Transformer Network (MMTN) to cope with multi-modal medical data for generating image-report consistent medical reports. On the one hand, MMTN reduces the occurrence of image-report inconsistencies by designing a unique encoder to associate and memorize the relationship between medical images and medical terminologies. On the other hand, MMTN utilizes the cross-modal complementarity of the medical vision and language for the word prediction, which further enhances the accuracy of generating medical reports. Extensive experiments on three real datasets show that MMTN achieves significant effectiveness over state-of-the-art approaches on both automatic metrics and human evaluation.
Unpaired 3D object completion aims to predict a complete 3D shape from an incomplete input without knowing the correspondence between the complete and incomplete shapes. In this paper, we propose the novel KTNet to solve this task from the new perspective of knowledge transfer. KTNet elaborates a teacher-assistant-student network to establish multiple knowledge transfer processes. Specifically, the teacher network takes complete shape as input and learns the knowledge of complete shape. The student network takes the incomplete one as input and restores the corresponding complete shape. And the assistant modules not only help to transfer the knowledge of complete shape from the teacher to the student, but also judge the learning effect of the student network. As a result, KTNet makes use of a more comprehensive understanding to establish the geometric correspondence between complete and incomplete shapes in a perspective of knowledge transfer, which enables more detailed geometric inference for generating high-quality complete shapes. We conduct comprehensive experiments on several datasets, and the results show that our method outperforms previous methods of unpaired point cloud completion by a large margin. Code is available at https://github.com/a4152684/KT-Net.
Recent research on Generalized Zero-Shot Learning (GZSL) has focused primarily on generation-based methods. However, current literature has overlooked the fundamental principles of these methods and has made limited progress in a complex manner. In this paper, we aim to deconstruct the generator-classifier framework and provide guidance for its improvement and extension. We begin by breaking down the generator-learned unseen class distribution into class-level and instance-level distributions. Through our analysis of the role of these two types of distributions in solving the GZSL problem, we generalize the focus of the generation-based approach, emphasizing the importance of (i) attribute generalization in generator learning and (ii) independent classifier learning with partially biased data. We present a simple method based on this analysis that outperforms SotAs on four public GZSL datasets, demonstrating the validity of our deconstruction. Furthermore, our proposed method remains effective even without a generative model, representing a step towards simplifying the generator-classifier structure. Our code is available at https://github.com/cdb342/DGZ.
In this work, we tackle the challenging task of jointly tracking hand object poses and reconstructing their shapes from depth point cloud sequences in the wild, given the initial poses at frame 0. We for the first time propose a point cloud-based hand joint tracking network, HandTrackNet, to estimate the inter-frame hand joint motion. Our HandTrackNet proposes a novel hand pose canonicalization module to ease the tracking task, yielding accurate and robust hand joint tracking. Our pipeline then reconstructs the full hand via converting the predicted hand joints into a MANO hand. For object tracking, we devise a simple yet effective module that estimates the object SDF from the first frame and performs optimization-based tracking. Finally, a joint optimization step is adopted to perform joint hand and object reasoning, which alleviates the occlusion-induced ambiguity and further refines the hand pose. During training, the whole pipeline only sees purely synthetic data, which are synthesized with sufficient variations and by depth simulation for the ease of generalization. The whole pipeline is pertinent to the generalization gaps and thus directly transferable to real in-the-wild data. We evaluate our method on two real hand object interaction datasets, e.g. HO3D and DexYCB, without any fine-tuning. Our experiments demonstrate that the proposed method significantly outperforms the previous state-of-the-art depth-based hand and object pose estimation and tracking methods, running at a frame rate of 9 FPS. We have released our code on https://github.com/PKU-EPIC/HOTrack.
Amodal instance segmentation aims to infer the amodal mask, including both the visible part and occluded part of each object instance. Predicting the occluded parts is challenging. Existing methods often produce incomplete amodal boxes and amodal masks, probably due to lacking visual evidences to expand the boxes and masks. To this end, we propose a prior-guided expansion framework, which builds on a two-stage segmentation model (i.e., Mask R-CNN) and performs box-level (resp., pixel-level) expansion for amodal box (resp., mask) prediction, by retrieving regression (resp., flow) transformations from a memory bank of expansion prior. We conduct extensive experiments on KINS, D2SA, and COCOA cls datasets, which show the effectiveness of our method.
Data-driven medium-range weather forecasting has attracted much attention in recent years. However, the forecasting accuracy at high resolution is unsatisfactory currently. Pursuing high-resolution and high-quality weather forecasting, we develop a data-driven model SwinRDM which integrates an improved version of SwinRNN with a diffusion model. SwinRDM performs predictions at 0.25-degree resolution and achieves superior forecasting accuracy to IFS (Integrated Forecast System), the state-of-the-art operational NWP model, on representative atmospheric variables including 500 hPa geopotential (Z500), 850 hPa temperature (T850), 2-m temperature (T2M), and total precipitation (TP), at lead times of up to 5 days. We propose to leverage a two-step strategy to achieve high-resolution predictions at 0.25-degree considering the trade-off between computation memory and forecasting accuracy. Recurrent predictions for future atmospheric fields are firstly performed at 1.40625-degree resolution, and then a diffusion-based super-resolution model is leveraged to recover the high spatial resolution and finer-scale atmospheric details. SwinRDM pushes forward the performance and potential of data-driven models for a large margin towards operational applications.
Most existing light field (LF) disparity estimation algorithms focus on handling occlusion, texture-less or other areas that harm LF structure to improve accuracy, while ignoring other potential modeling ideas. In this paper, we propose a novel idea called Bad Pixel (BadPix) correction for method modeling, then implement a general post-refinement network for LF disparity estimation: Bad-pixel Correction Network (BpCNet). Given an initial disparity map generated by a specific algorithm, we assume that all BadPixs on it are in a small range. Then BpCNet is modeled as a fine-grained search strategy, and a more accurate result can be obtained by evaluating the consistency of LF images in this limited range. Due to the assumption and the consistency between input and output, BpCNet can perform as a general post-refinement network, and can work on almost all existing algorithms iteratively. We demonstrate the feasibility of our theory through extensive experiments, and achieve remarkable performance on the HCI 4D Light Field Benchmark.
Reconstructing a High Dynamic Range (HDR) image from several Low Dynamic Range (LDR) images with different exposures is a challenging task, especially in the presence of camera and object motion. Though existing models using convolutional neural networks (CNNs) have made great progress, challenges still exist, e.g., ghosting artifacts. Transformers, originating from the field of natural language processing, have shown success in computer vision tasks, due to their ability to address a large receptive field even within a single layer. In this paper, we propose a transformer model for HDR imaging. Our pipeline includes three steps: alignment, fusion, and reconstruction. The key component is the HDR transformer module. Through experiments and ablation studies, we demonstrate that our model outperforms the state-of-the-art by large margins on several popular public datasets.
Spiking camera, a novel retina-inspired vision sensor, has shown its great potential for capturing high-speed dynamic scenes with a sampling rate of 40,000 Hz. The spiking camera abandons the concept of exposure window, with each of its photosensitive units continuously capturing photons and firing spikes asynchronously. However, the special sampling mechanism prevents the frame-based algorithm from being used to spiking camera. It remains to be a challenge to reconstruct dynamic scenes and perform common computer vision tasks for spiking camera. In this paper, we propose a self-supervised joint learning framework for optical flow estimation and reconstruction of spiking camera. The framework reconstructs clean frame-based spiking representations in a self-supervised manner, and then uses them to train the optical flow networks. We also propose an optical flow based inverse rendering process to achieve self-supervision by minimizing the difference with respect to the original spiking temporal aggregation image. The experimental results demonstrate that our method bridges the gap between synthetic and real-world scenes and achieves desired results in real-world scenarios. To the best of our knowledge, this is the first attempt to jointly reconstruct dynamic scenes and estimate optical flow for spiking camera from a self-supervised learning perspective.
Neural Radiance Fields (NeRF) can implicitly represent 3D-consistent RGB images and geometric by optimizing an underlying continuous volumetric scene function using a sparse set of input views, which has greatly benefited view synthesis tasks. However, NeRF fails to estimate correct geometry when given fewer views, resulting in failure to synthesize novel views. Existing works rely on introducing depth images or adding depth estimation networks to resolve the problem of poor synthetic view in NeRF with fewer views. However, due to the lack of spatial consistency of the single-depth image and the poor performance of depth estimation with fewer views, the existing methods still have challenges in addressing this problem. So this paper proposes Bidirectional Optical Flow NeRF(BOF-NeRF), which addresses this problem by mining optical flow information between 2D images. Our key insight is that utilizing 2D optical flow images to design a loss can effectively guide NeRF to learn the correct geometry and synthesize the right novel view. We also propose a view-enhanced fusion method based on geometry and color consistency to solve the problem of novel view details loss in NeRF. We conduct extensive experiments on the NeRF-LLFF and DTU MVS benchmarks for novel view synthesis tasks with fewer images in different complex real scenes. We further demonstrate the robustness of BOF-NeRF under different baseline distances on the Middlebury dataset. In all cases, BOF-NeRF outperforms current state-of-the-art baselines for novel view synthesis and scene geometry estimation.
Neural scene representation and rendering methods have shown promise in learning the implicit form of scene structure without supervision. However, the implicit representation learned in most existing methods is non-expandable and cannot be inferred online for novel scenes, which makes the learned representation difficult to be applied across different reinforcement learning (RL) tasks. In this work, we introduce Scene Memory Network (SMN) to achieve online spatial memory construction and expansion for view rendering in novel scenes. SMN models the camera projection and back-projection as spatially aware memory control processes, where the memory values store the information of the partial 3D area, and the memory keys indicate the position of that area. The memory controller can learn the geometry property from observations without the camera's intrinsic parameters and depth supervision. We further apply the memory constructed by SMN to exploration and navigation tasks. The experimental results reveal the generalization ability of our proposed SMN in large-scale scene synthesis and its potential to improve the performance of spatial RL tasks.
Since rain streaks exhibit diverse geometric appearances and irregular overlapped phenomena, these complex characteristics challenge the design of an effective single image deraining model. To this end, rich local-global information representations are increasingly indispensable for better satisfying rain removal. In this paper, we propose a lightweight Hybrid CNN-Transformer Feature Fusion Network (dubbed as HCT-FFN) in a stage-by-stage progressive manner, which can harmonize these two architectures to help image restoration by leveraging their individual learning strengths. Specifically, we stack a sequence of the degradation-aware mixture of experts (DaMoE) modules in the CNN-based stage, where appropriate local experts adaptively enable the model to emphasize spatially-varying rain distribution features. As for the Transformer-based stage, a background-aware vision Transformer (BaViT) module is employed to complement spatially-long feature dependencies of images, so as to achieve global texture recovery while preserving the required structure. Considering the indeterminate knowledge discrepancy among CNN features and Transformer features, we introduce an interactive fusion branch at adjacent stages to further facilitate the reconstruction of high-quality deraining results. Extensive evaluations show the effectiveness and extensibility of our developed HCT-FFN. The source code is available at https://github.com/cschenxiang/HCT-FFN.