CVPR 2022 Papers

In spite of the dominant performances of deep neural networks, recent works have shown that they are poorly calibrated, resulting in over-confident predictions. Miscalibration can be exacerbated by overfitting due to the minimization of the cross-entropy during training, as it promotes the predicted softmax probabilities to match the one-hot label assignments. This yields a pre-softmax activation of the correct class that is significantly larger than the remaining activations. Recent evidence from the literature suggests that loss functions that embed implicit or explicit maximization of the entropy of predictions yield state-of-the-art calibration performances. We provide a unifying constrained-optimization perspective of current state-of-the-art calibration losses. Specifically, these losses could be viewed as approximations of a linear penalty (or a Lagrangian term) imposing equality constraints on logit distances. This points to an important limitation of such underlying equality constraints, whose ensuing gradients constantly push towards a non-informative solution, which might prevent from reaching the best compromise between the discriminative performance and calibration of the model during gradient-based optimization. Following our observations, we propose a simple and flexible generalization based on inequality constraints, which imposes a controllable margin on logit distances. Comprehensive experiments on a variety of image classification, semantic segmentation and NLP benchmarks demonstrate that our method sets novel state-of-the-art results on these tasks in terms of network calibration, without affecting the discriminative performance. The code is available at https://github.com/by-liu/MbLS.

Quantitative evaluation has increased dramatically among recent video inpainting work, but the video and mask content used to gauge performance has received relatively little attention. Although attributes such as camera and background scene motion inherently change the difficulty of the task and affect methods differently, existing evaluation schemes fail to control for them, thereby providing minimal insight into inpainting failure modes. To address this gap, we propose the Diagnostic Evaluation of Video Inpainting on Landscapes (DEVIL) benchmark, which consists of two contributions: (i) a novel dataset of videos and masks labeled according to several key inpainting failure modes, and (ii) an evaluation scheme that samples slices of the dataset characterized by a fixed content attribute, and scores performance on each slice according to reconstruction, realism, and temporal consistency quality. By revealing systematic changes in performance induced by particular characteristics of the input content, our challenging benchmark enables more insightful analysis into video inpainting methods and serves as an invaluable diagnostic tool for the field. Our code and data are available at github.com/MichiganCOG/devil.

Generalization to novel domains is a fundamental challenge for computer vision. Near-perfect accuracy on benchmarks is common, but these models do not work as expected when deployed outside of the training distribution. To build computer vision systems that truly solve real-world problems at global scale, we need benchmarks that fully capture real-world complexity, including geographic domain shift, long-tailed distributions, and data noise. We propose urban forest monitoring as an ideal testbed for studying and improving upon these computer vision challenges, while simultaneously working towards filling a crucial environmental and societal need. Urban forests provide significant benefits to urban societies (e.g., cleaner air and water, carbon sequestration, and energy savings among others). However, planning and maintaining these forests is expensive. One particularly costly aspect of urban forest management is monitoring the existing trees in a city: e.g., tracking tree locations, species, and health. Monitoring efforts are currently based on tree censuses built by human experts, costing cities millions of dollars per census and thus collected infrequently. Previous investigations into automating urban forest monitoring focused on small datasets from single cities, covering only common categories. To address these shortcomings, we introduce a new large-scale dataset that joins public tree censuses from 23 cities with a large collection of street level and aerial imagery. Our Auto Arborist dataset contains over 2.5M trees and 344 genera and is >2 orders of magnitude larger than the closest dataset in the literature. We introduce baseline results on our dataset across modalities as well as metrics for the detailed analysis of generalization with respect to geographic distribution shifts, vital for such a system to be deployed at-scale.

Evaluation practices for image super-resolution (SR) use a single-value metric, the PSNR or SSIM, to determine model performance. This provides little insight into the source of errors and model behavior. Therefore, it is beneficial to move beyond the conventional approach and reconceptualize evaluation with interpretability as our main priority. We focus on a thorough error analysis from a variety of perspectives. Our key contribution is to leverage a texture classifier, which enables us to assign patches with semantic labels, to identify the source of SR errors both globally and locally. We then use this to determine (a) the semantic alignment of SR datasets, (b) how SR models perform on each label, (c) to what extent high-resolution (HR) and SR patches semantically correspond, and more. Through these different angles, we are able to highlight potential pitfalls and blindspots. Our overall investigation highlights numerous unexpected insights. We hope this work serves as an initial step for debugging blackbox SR networks.

Natural language-based communication with mobile devices and home appliances is becoming increasingly popular and has the potential to become natural for communicating with mobile robots in the future. Towards this goal, we investigate cross-modal text-to-point-cloud localization that will allow us to specify, for example, a vehicle pick-up or goods delivery location. In particular, we propose Text2Pos, a cross-modal localization module that learns to align textual descriptions with localization cues in a coarse- to-fine manner. Given a point cloud of the environment, Text2Pos locates a position that is specified via a natural language-based description of the immediate surroundings. To train Text2Pos and study its performance, we construct KITTI360Pose, the first dataset for this task based on the recently introduced KITTI360 dataset. Our experiments show that we can localize 65% of textual queries within 15m distance to query locations for top-10 retrieved locations. This is a starting point that we hope will spark future developments towards language-based navigation.

Object-guided text-to-image synthesis aims to generate images from natural language descriptions built by two-step frameworks, i.e., the model generates the layout and then synthesizes images from the layout and captions. However, such frameworks have two issues: 1) complex structure, since generating language-related layout is not a trivial task; 2) error propagation, because the inappropriate layout will mislead the image synthesis and is hard to be revised. In this paper, we propose an object-guided joint-decoding module to simultaneously generate the image and the corresponding layout. Specially, we present the joint-decoding transformer to model the joint probability on images tokens and the corresponding layouts tokens, where layout tokens provide additional observed data to model the complex scene better. Then, we describe a novel Layout-VQGAN for layout encoding and decoding to provide more information about the complex scene. After that, we present the detail-enhanced module to enrich the language-related details based on two facts: 1) visual details could be omitted in the compression of VQGANs; 2) the joint-decoding transformer would not have sufficient generating capacity. The experiments show that our approach is competitive with previous object-centered models and can generate diverse and high-quality objects under the given layouts.

Text-to-image synthesis (T2I) aims to generate photo-realistic images which are semantically consistent with the text descriptions. Existing methods are usually built upon conditional generative adversarial networks (GANs) and initialize an image from noise with sentence embedding, and then refine the features with fine-grained word embedding iteratively. A close inspection of their generated images reveals a major limitation: even though the generated image holistically matches the description, individual image regions or parts of somethings are often not recognizable or consistent with words in the sentence, e.g. "a white crown". To address this problem, we propose a novel framework Semantic-Spatial Aware GAN for synthesizing images from input text. Concretely, we introduce a simple and effective Semantic-Spatial Aware block, which (1) learns semantic-adaptive transformation conditioned on text to effectively fuse text features and image features, and (2) learns a semantic mask in a weakly-supervised way that depends on the current text-image fusion process in order to guide the transformation spatially. Experiments on the challenging COCO and CUB bird datasets demonstrate the advantage of our method over the recent state-of-the-art approaches, regarding both visual fidelity and alignment with input text description. Code available at https://github.com/wtliao/text2image.

In this paper, we present TExt Spotting TRansformers (TESTR), a generic end-to-end text spotting framework using Transformers for text detection and recognition in the wild. TESTR builds upon a single encoder and dual decoders for the joint text-box control point regression and character recognition. Other than most existing literature, our method is free from Region-of-Interest operations and heuristics-driven post-processing procedures; TESTR is particularly effective when dealing with curved text-boxes where special cares are needed for the adaptation of the traditional bounding-box representations. We show our canonical representation of control points suitable for text instances in both Bezier curve and polygon annotations. In addition, we design a bounding-box guided polygon detection (box-to-polygon) process. Experiments on curved and arbitrarily shaped datasets demonstrate state-of-the-art performances of the proposed TESTR algorithm.

Multi-modal video similarity evaluation is important for video recommendation systems such as video de-duplication, relevance matching, ranking, and diversity control. However, there still lacks a benchmark dataset that can support supervised training and accurate evaluation. In this paper, we propose the Tencent-MVSE dataset, which is the first benchmark dataset for the multi-modal video similarity evaluation task. The Tencent-MVSE dataset contains video pairs similarity annotations, and diverse metadata including Chinese title, automatic speech recognition (ASR) text, as well as human-annotated categories/tags. We provide a simple baseline with a multi-modal Transformer architecture to perform supervised multi-modal video similarity evaluation. We also explore pre-training strategies to make use of the unpaired data. The whole dataset as well as our baseline will be released to promote the development of the multi-modal video similarity evaluation. The dataset has been released in https://tencent-mvse.github.io/.

We propose a novel approach to generate temporally coherent UV coordinates for loose clothing. Our method is not constrained by human body outlines and can capture loose garments and hair. We implemented a differentiable pipeline to learn UV mapping between a sequence of RGB inputs and textures via UV coordinates. Instead of treating the UV coordinates of each frame separately, our data generation approach connects all UV coordinates via feature matching for temporal stability. Subsequently, a generative model is trained to balance the spatial quality and temporal stability. It is driven by supervised and unsupervised losses in both UV and image spaces. Our experiments show that the trained models output high-quality UV coordinates and generalize to new poses. Once a sequence of UV coordinates has been inferred by our model, it can be used to flexibly synthesize new looks and modified visual styles. Compared to existing methods, our approach reduces the computational workload to animate new outfits by several orders of magnitude.

Image-to-video person re-identification aims to retrieve the same pedestrian as the image-based query from a video-based gallery set. Existing methods treat it as a cross-modality retrieval task and learn the common latent embeddings from image and video modalities, which are both less effective and efficient due to large modality gap and redundant feature learning by utilizing all video frames. In this work, we first regard this task as point-to-set matching problem identical to human decision process, and propose a novel Temporal Complementarity-Guided Reinforcement Learning (TCRL) approach for image-to-video person re-identification. TCRL employs deep reinforcement learning to make sequential judgments on dynamically selecting suitable amount of frames from gallery videos, and accumulate adequate temporal complementary information among these frames by the guidance of the query image, towards balancing efficiency and accuracy. Specifically, TCRL formulates point-to-set matching procedure as Markov decision process, where a sequential judgement agent measures the uncertainty between the query image and all historical frames at each time step, and verifies that sufficient complementary clues are accumulated for judgment (same or different) or one more frames are requested to assist judgment. Moreover, TCRL maintains a sequential feature extraction module with a complementary residual detector to dynamically suppress redundant salient regions and thoroughly mine diverse complementary clues among these selected frames for enhancing frame-level representation. Extensive experiments demonstrate the superiority of our method.

We present a method that can recognize new objects and estimate their 3D pose in RGB images even under partial occlusions. Our method requires neither a training phase on these objects nor real images depicting them, only their CAD models. It relies on a small set of training objects to learn local object representations, which allow us to locally match the input image to a set of "templates", rendered images of the CAD models for the new objects. In contrast with the state-of-the-art methods, the new objects on which our method is applied can be very different from the training objects. As a result, we are the first to show generalization without retraining on the LINEMOD and Occlusion-LINEMOD datasets. Our analysis of the failure modes of previous template-based approaches further confirms the benefits of local features for template matching. We outperform the state-of-the-art template matching methods on the LINEMOD, Occlusion-LINEMOD and T-LESS datasets. Our source code and data are publicly available at https://github.com/nv-nguyen/template-pose

Pre-training models on Imagenet or other massive datasets of real images has led to major advances in computer vision, albeit accompanied with shortcomings related to curation cost, privacy, usage rights, and ethical issues. In this paper, for the first time, we study the transferability of pre-trained models based on synthetic data generated by graphics simulators to downstream tasks from very different domains. In using such synthetic data for pre-training, we find that downstream performance on different tasks are favored by different configurations of simulation parameters (e.g. lighting, object pose, backgrounds, etc.), and that there is no one-size-fits-all solution. It is thus better to tailor synthetic pre-training data to a specific downstream task, for best performance. We introduce Task2Sim, a unified model mapping downstream task representations to optimal simulation parameters to generate synthetic pre-training data for them. Task2Sim learns this mapping by training to find the set of best parameters on a set of "seen" tasks. Once trained, it can then be used to predict best simulation parameters for novel "unseen" tasks in one shot, without requiring additional training. Given a budget in number of images per class, our extensive experiments with 20 diverse downstream tasks show Task2Sim's task-adaptive pre-training data results in significantly better downstream performance than non-adaptively choosing simulation parameters on both seen and unseen tasks. It is even competitive with pre-training on real images from Imagenet.

Detectors trained with massive labeled data often exhibit dramatic performance degradation in some particular scenarios with data distribution gap. To alleviate this problem of domain shift, conventional wisdom typically concentrates solely on reducing the discrepancy between the source and target domains via attached domain classifiers, yet ignoring the difficulty of such transferable features in coping with both classification and localization subtasks in object detection. To address this issue, in this paper, we propose Task-specific Inconsistency Alignment (TIA), by developing a new alignment mechanism in separate task spaces, improving the performance of the detector on both subtasks. Specifically, we add a set of auxiliary predictors for both classification and localization branches, and exploit their behavioral inconsistencies as finer-grained domain-specific measures. Then, we devise task-specific losses to align such cross-domain disagreement of both subtasks. By optimizing them individually, we are able to well approximate the category- and boundary-wise discrepancies in each task space, and therefore narrow them in a decoupled manner. TIA demonstrates superior results on various scenarios to the previous state-of-the-art methods. It is also observed that both the classification and localization capabilities of the detector are sufficiently strengthened, further demonstrating the effectiveness of our TIA method. Code and trained models are publicly available at https://github.com/MCG-NJU/TIA.

We study the problem of few-shot open-set recognition (FSOR), which learns a recognition system capable of both fast adaptation to new classes with limited labeled examples and rejection of unknown negative samples. Traditional large-scale open-set methods have been shown ineffective for FSOR problem due to data limitation. Current FSOR methods typically calibrate few-shot closed-set classifiers to be sensitive to negative samples so that they can be rejected via thresholding. However, threshold tuning is a challenging process as different FSOR tasks may require different rejection powers. In this paper, we instead propose task-adaptive negative class envision for FSOR to integrate threshold tuning into the learning process. Specifically, we augment the few-shot closed-set classifier with additional negative prototypes generated from few-shot examples. By incorporating few-shot class correlations in the negative generation process, we are able to learn dynamic rejection boundaries for FSOR tasks. Besides, we extend our method to generalized few-shot open-set recognition (GFSOR), which requires classification on both many-shot and few-shot classes as well as rejection of negative samples. Extensive experiments on public benchmarks validate our methods on both problems.

Reference-based super-resolution(RefSR) has achieved impressive progress on the recovery of high-frequency details thanks to an additional reference high-resolution(HR) image input. Although the superiority compared with Single-Image Super-Resolution(SISR), existing RefSR methods easily result in the reference-underuse issue and the reference-misuse as shown in Fig.1. In this work, we deeply investigate the cause of the two issues and further propose a novel framework to mitigate them. Our studies find that the issues are mostly due to the improper coupled framework design of current methods. Those methods conduct the super-resolution task of the input low-resolution(LR) image and the texture transfer task from the reference image together in one module, easily introducing the interference between LR and reference features. Inspired by this finding, we propose a novel framework, which decouples the two tasks of RefSR, eliminating the interference between the LR image and the reference image. The super-resolution task upsamples the LR image leveraging only the LR image itself. The texture transfer task extracts and transfers abundant textures from the reference image to the coarsely upsampled result of the super-resolution task. Extensive experiments demonstrate clear improvements in both quantitative and qualitative evaluations over state-of-the-art methods.

Adapting pre-trained models with broad capabilities has become standard practice for learning a wide range of downstream tasks. The typical approach of fine-tuning different models for each task is performant, but incurs a substantial memory cost. To efficiently learn multiple downstream tasks we introduce Task Adaptive Parameter Sharing (TAPS), a simple method for tuning a base model to a new task by adaptively modifying a small, task-specific subset of layers. This enables multi-task learning while minimizing the resources used and avoids catastrophic forgetting and competition between tasks. TAPS solves a joint optimization problem which determines both the layers that are shared with the base model and the value of the task-specific weights. Further, a sparsity penalty on the number of active layers promotes weight sharing with the base model. Compared to other methods, TAPS retains a high accuracy on the target tasks while still introducing only a small number of task-specific parameters. Moreover, TAPS is agnostic to the particular architecture used and requires only minor changes to the training scheme. We evaluate our method on a suite of fine-tuning tasks and architectures (ResNet, DenseNet, ViT) and show that it achieves state-of-the-art performance while being simple to implement.

Real-world data often exhibits long tail distributions with heavy class imbalance, where the majority classes can dominate the training process and alter the decision boundaries of the minority classes. Recently, researchers have investigated the potential of supervised contrastive learning for long-tailed recognition, and demonstrated that it provides a strong performance gain. In this paper, we show that while supervised contrastive learning can help improve performance, past baselines suffer from poor uniformity brought in by imbalanced data distribution. This poor uniformity manifests in samples from the minority class having poor separability in the feature space. To address this problem, we propose targeted supervised contrastive learning (TSC), which improves the uniformity of the feature distribution on the hypersphere. TSC first generates a set of targets uniformly distributed on a hypersphere. It then makes the features of different classes converge to these distinct and uniformly distributed targets during training. This forces all classes, including minority classes, to maintain a uniform distribution in the feature space, improves class boundaries, and provides better generalization even in the presence of long-tail data. Experiments on multiple datasets show that TSC achieves state-of-the-art performance on long-tailed recognition tasks.